docs(claude): record plan 002 phase 3 in the project status

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.claude/agents/Explore.md

@@ -0,0 +1,8 @@
+---
+name: Explore
+description: Fast, read-only codebase search
+model: sonnet
+effort: low
+tools: Read, Grep, Glob, Bash, WebFetch, WebSearch
+maxTurns: 20
+---

.env.example

@@ -9,12 +9,19 @@
 
 # Application
 APP_ENV=local
-APP_MAX_UPLOAD_SIZE_MB=25
 APP_READINESS_CHECK_TIMEOUT_SECONDS=2.0
+# Set by CI/CD at build/deploy time; never computed at runtime.
+APP_SERVICE_VERSION=dev
 
 # Logging
 LOG_LEVEL=INFO
 LOG_JSON_FORMAT=false
+# Optional second sink, always JSON regardless of LOG_JSON_FORMAT. Local dev
+# only -- leave unset in production, where stdout/stderr collection is
+# preferred over an in-container log file.
+# LOG_FILE_PATH=logs/app.log
+LOG_FILE_MAX_BYTES=10485760
+LOG_FILE_BACKUP_COUNT=5
 
 # Postgres (application database, separate from Langfuse's Postgres)
 # Use 127.0.0.1 rather than localhost: some environments resolve localhost to
@@ -37,6 +44,7 @@ MINIO_BUCKET=chatbot-source-files
 INGESTION_MAX_CONCURRENCY=4
 INGESTION_THREAD_POOL_SIZE=8
 INGESTION_TIMEOUT_SECONDS=120.0
+INGESTION_MAX_UPLOAD_SIZE_MB=25
 INGESTION_MAX_CHUNKS_PER_FILE=5000
 INGESTION_EMBED_BATCH_SIZE=128
 INGESTION_EMBED_CONCURRENCY=4
@@ -44,6 +52,48 @@ INGESTION_EMBED_CONCURRENCY=4
 # Qdrant
 QDRANT_URL=http://127.0.0.1:6343
 QDRANT_API_KEY=
+QDRANT_COLLECTION=chunks
+QDRANT_UPSERT_BATCH_SIZE=128
+QDRANT_UPSERT_CONCURRENCY=4
+
+# Dense embedders (ADR-0001). Both speak an OpenAI-compatible /embeddings
+# endpoint, so one adapter serves both. Models and endpoints are the ones the
+# `emet` evaluation lab benchmarked as winners on the Farsi corpus.
+#
+# dense_nomic runs behind Ollama's OpenAI-compat shim, which accepts any
+# non-empty API key. KEEP_ALIVE holds the model resident: a cold load of
+# nomic-embed-text-v2-moe takes >150s, well past INGESTION_TIMEOUT_SECONDS,
+# so an idle-then-upload would otherwise 504.
+EMBEDDING_NOMIC_BASE_URL=http://192.168.10.10:11435/v1
+EMBEDDING_NOMIC_MODEL=nomic-embed-text-v2-moe
+EMBEDDING_NOMIC_API_KEY=sk-not-set
+EMBEDDING_NOMIC_KEEP_ALIVE=30m
+EMBEDDING_NOMIC_TIMEOUT_SECONDS=30.0
+# Empty = emet parity. The model card specifies `search_document: ` (ADR-0004),
+# but the benchmark ran without it and the prefix shifts the vector a lot
+# (cosine 0.57 on identical text) — so if you set this, the query side must
+# send `search_query: ` to match, or retrieval gets worse rather than better.
+EMBEDDING_NOMIC_DOCUMENT_PREFIX=
+
+# Leave DIMENSIONS empty for text-embedding-3-large's native 3072, which is
+# what was benchmarked. Setting it truncates via Matryoshka and is a
+# re-embedding migration, not a config tweak.
+EMBEDDING_OPENAI_BASE_URL=https://api.openai.com/v1
+EMBEDDING_OPENAI_MODEL=text-embedding-3-large
+EMBEDDING_OPENAI_API_KEY=
+EMBEDDING_OPENAI_DIMENSIONS=
+EMBEDDING_OPENAI_DOCUMENT_PREFIX=
+EMBEDDING_OPENAI_TIMEOUT_SECONDS=30.0
+
+# Sparse BM25 (ADR-0001, ADR-0005): the benchmarked `bm25-fa-norm-stop`.
+# k/b saturation is applied client-side; IDF comes from Qdrant's
+# modifier="idf" on the sparse vector field. AVG_LEN is the average document
+# length in analyzer tokens — emet's placeholder, worth recalibrating from
+# real corpus statistics.
+EMBEDDING_SPARSE_ANALYZER=fa_norm_stop
+EMBEDDING_SPARSE_K=1.2
+EMBEDDING_SPARSE_B=0.75
+EMBEDDING_SPARSE_AVG_LEN=256.0
 
 # Parsing and chunking (ADR-0018).
 # max_chunk_tokens is nomic-embed-text-v2-moe's sequence length; text past it

CLAUDE.md

@@ -7,11 +7,127 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 This repo is ADR-driven and early in implementation. Working today: the FastAPI
 app factory and lifespan wiring (`src/bootstrap/`), `/healthz` and `/readyz`,
 structlog config, Postgres/MinIO/Qdrant clients (`src/infrastructure/`), five
-SQLAlchemy models with one Alembic migration, and document parsing plus
-fixed-size chunking (`src/application/ingestion/`). Not built yet: API-key auth,
-`POST /v1/files` (so `/v1` currently exposes no routes), repositories, embedding
-adapters, Qdrant collection bootstrap, and `src/agent/`. That maps to plan 001
-Phase 1 done, Phase 2 partly done, and the parsing half of Phase 5.
+SQLAlchemy models with one Alembic migration, document parsing plus fixed-size
+chunking (`src/application/ingestion/`), API-key auth, `POST`/`GET /v1/files`
+with durable two-phase job creation (`src/application/files/`), and bounded
+inline embedding: `dense_nomic`/`dense_openai` adapters over an
+OpenAI-compatible HTTP client and a `bm25-fa-norm-stop` sparse adapter
+(`src/infrastructure/embedding/`), wired into the upload path behind
+`INGESTION_MAX_CONCURRENCY` (`503`), `INGESTION_TIMEOUT_SECONDS` (`504`), and
+the chunk-count ceiling (`413`). The embedding configuration is **ported from
+the `emet` evaluation lab** (`~/code/talie/emet`), which benchmarked these
+models and analyzers on the real Farsi corpus — the analyzer and BM25 weights
+are verified token-for-token against it, so treat them as a measured artifact
+and re-benchmark rather than tune them in place (ADR-0005). Also working: the
+`chunks` collection bootstrap (`src/infrastructure/qdrant/collection.py`, run as
+a deployment step via `uv run python -m src.cli.qdrant_bootstrap` — never at
+startup) and tenant-scoped point upserts (`src/application/points/` behind the
+`PointStorage` port), so an upload is searchable by the time `201` returns.
+Also working: `tenant_domains` plus `/v1/domains` (`src/application/domains/`),
+a strict per-tenant allowlist — `POST /v1/files` rejects an unregistered or
+disabled `domain` with `400` before anything is written, and domain management
+sits behind its own `domains:read`/`domains:write` scopes, never `files:write`.
+Also working: the operator runbook (`docs/runbook.md`), tenant/API-key/domain
+provisioning (`uv run python -m src.cli.provision_tenant` — the third deployment
+step, since nothing over HTTP can create the first tenant), a Testcontainers
+e2e suite in the default pytest run (`tests/e2e/test_ingestion_slice.py`:
+duplicate upload, retry after failure, tenant isolation, capacity, timeout,
+parse and Qdrant failure), and the one Compose-based test — `scripts/smoke.sh`
+driving `tests/e2e/test_compose_smoke.py` against a real uvicorn process, which
+skips itself unless `SMOKE_BASE_URL` is set. That maps to plan 001 Phases 1-6
+done.
+
+Plan 002 (`/v1/points` CRUD and keyword search) is **Phases 1-3 done**. Phase 1
+landed the `PointRepository` port (`src/application/ports/point_repository.py`)
+with its `Point` read model (`src/application/points/point.py`), the Qdrant adapter
+(`src/infrastructure/qdrant/point_repository.py`), request/response schemas
+(`src/api/schemas/points.py`), and lifespan wiring. This port is **separate from
+`PointStorage`**, which stays exactly the two bulk operations ingestion
+performs — reads, single-point edits, and keyword search have a different caller
+and a different tenant-filter obligation, so do not accrete them onto the
+ingestion port. `tenant_id` is a required keyword argument on every
+`PointRepository` method by design; keep it that way, because it is what turns a
+forgotten tenant filter into a type error. The `chunks` collection also gained
+full-text `content`, `is_active`, and `chunk_index` payload indexes, so a
+deployed environment needs `qdrant_bootstrap` re-run (indexes are additive — no
+rebuild, no re-embedding).
+
+Two adapter mechanics there are load-bearing and easy to "simplify" into bugs:
+reads go through `scroll` with a `HasIdCondition` rather than `retrieve` (which
+takes no filter, and would move the tenant check to *after* Qdrant answered),
+and ordered listing paginates by `order_id` value rather than offset (Qdrant
+returns no page offset under `order_by`, and an offset cursor skips or repeats
+rows under a concurrent insert).
+
+Phase 2 added the **read routes**: `GET /v1/points/{point_id}`,
+`GET /v1/points?file_id=...`, `GET /v1/points/count`, `GET /v1/points/search`,
+and `GET /v1/files/{file_id}/points`, over `src/application/points/queries.py`
+(`src/api/routers/points.py`). All are gated on `points:read`, which — with
+`points:write` — is now in `DEFAULT_SCOPES`; `GET /v1/files/{file_id}/points`
+uses `points:read` rather than `files:write`, so the scope follows the data
+rather than the URL prefix. `PointNotFoundError` maps to `404` in
+`src/api/errors.py`, never `403`. Three route-level rules are load-bearing:
+`/count` and `/search` are declared **before** `/{point_id}` (FastAPI matches in
+declaration order, so reordering them makes `/v1/points/count` a `422`),
+`file_id` is **required** on the listing (the cursor is an `order_id` value and
+`order_id` is unique only within one file), and `search_points` folds the query
+with `normalize_persian_text` before matching, because ingestion letter-folds
+content and an unfolded Arabic-keyboard query would return an empty result set
+silently rather than erroring (ADR-0002).
+
+Phase 3 added **soft delete**: `DELETE /v1/points/{point_id}` and
+`DELETE /v1/files/{file_id}`, over `src/application/points/deletion.py` (with
+the pure relinking primitive in `src/application/points/relinking.py`) and
+`src/application/files/deletion.py`. Both are gated on `points:write` — the
+file route included, since the data it destroys is points. Nothing is ever
+removed from Qdrant.
+
+Four rules there are load-bearing, and three of them look like complications
+until the concurrency is taken seriously:
+
+- `patches_for_removal` computes **what is still missing between the state just
+  read and the desired end state**, not "the patches a delete implies". That is
+  what makes a normal delete, a second delete of an already-inactive point (a
+  no-op success, never `404`), and recovery from a half-applied batch one code
+  path. Rewriting it as a straight-line "deactivate, patch prev, patch next"
+  breaks all three.
+- Qdrant has no multi-point transaction and reports success for a filtered
+  `set_payload` that matched nothing, so a batch whose second operation loses a
+  version race applies its first anyway. `soft_delete_point` therefore re-plans
+  and re-applies up to three times, verifying by read-back, and only then raises
+  `PointVersionConflictError` (`409`). A single-shot delete would be able to
+  leave a stale pointer, which ADR-0002 calls a defect.
+- A soft-deleted point **keeps its own** `previous_chunk_id`/`next_chunk_id`;
+  only the surviving neighbours are rewritten. Those pointers are unreachable
+  rather than stale, they are the only record of where the point sat, and the
+  retry re-plans from them. The whole-file sweep follows from the same rule:
+  every point leaves at once, so no survivor can dangle and no pointer is
+  touched at all.
+- `DELETE /v1/files/{file_id}` marks the `source_files` row `soft_deleted`
+  **after** the point sweep, in its own short transaction (no session is held
+  across the Qdrant work). Order matters: a half-finished sweep leaves the row
+  `active` and a retried `DELETE` finishes it, and retiring the row is what
+  makes a later re-upload of the same bytes re-ingest instead of matching
+  `find_active_by_content_hash` and returning a file whose points are gone.
+
+Audit rows are still Phase 4/6 work; Phase 3 emits log events only
+(`points.soft_deleted`, `files.soft_deleted`, `points.relink.neighbour_missing`,
+and the two `*.conflict` warnings). The completion and conflict events carry
+ADR-0011's `duration_ms` plus `rounds`, and the pair is what makes them
+diagnostic: relinking itself is O(1) (that is what the adjacency pointers buy),
+so a single-point delete costs a fixed ~5 Qdrant round trips and a `rounds`
+above 1 means contention, not a slow store. The whole-file sweep is the one
+whose cost scales — two round trips per 100-point page.
+
+Also worth knowing before touching the points tests: `tests/support/point_contract.py`
+holds **one** scenario suite run against both `FakePointRepository` (unit) and
+`QdrantPointRepository` (integration), so new repository behaviour belongs there
+rather than in one of the two runners — that is what keeps the fake from drifting
+more permissive than the real store.
+
+Not built yet: plan 002 Phases 4-6 — create/replace/patch, reorder and batch,
+the `api_request_logs`/`point_audit_events` tables, and the runbook section on
+inspecting and repairing a file's pointer chain — and `src/agent/`.
 
 Architecture decisions live in `docs/adr/` (18 ADRs plus the 0000 template;
 0001–0004 are `Accepted` — 0004 amended by 0018; 0014 is `Superseded by 0017`;
@@ -113,6 +229,31 @@ MinIO/Qdrant/SQLAlchemy client-construction code. Use ports only for
 external side effects/persistence — not around pure local functions.
 (ADR-0015)
 
+### Prefer deep modules over shallow ones
+
+When a package exposes several small pure functions that a caller must
+compose correctly every time (right dispatch, right order, right
+thread/async offload), give it one entry point that owns that composition,
+and keep the small functions internal — exported only where their own unit
+tests need them. A shallow interface (one whose surface is nearly as complex
+as its implementation) pushes a correctness obligation onto every call site;
+a deep one absorbs it once. Apply the deletion test when unsure: if deleting
+the wrapper would concentrate the composition logic back into every caller
+rather than just relocate it, the wrapper is worth having.
+
+Worked example: `src/application/ingestion/` exposes `parse_and_chunk_document`
+as its only caller-facing entry point. It dispatches on source type and owns
+the `anyio.to_thread.run_sync` + `CapacityLimiter` offload ADR-0017 requires;
+`parse_docx`/`parse_csv`/`parse_xlsx`/`chunk_document` stay in the package,
+exported mainly for their own tests, not for outside callers to reach for
+directly. `src/application/points/` follows the same shape: `index_chunks` is
+the only caller-facing entry point, owning payload construction, batching,
+the `upsert_concurrency` semaphore, and the ordering rule that the soft-delete
+sweep runs only after every upsert succeeds; `build_chunk_payload` stays
+internal. Follow this pattern in `application/` as new packages are added
+there — `retrieval/`, `threads/` — rather than exposing their internals as the
+primary surface.
+
 ### Resource lifetime rules (ADR-0012)
 
 - Application-lifetime objects (SQLAlchemy engine/sessionmaker, Qdrant client,
@@ -190,6 +331,11 @@ content_sha256)` idempotency, no terminal job returning to `running`.
 
 ### Postgres conventions (ADR-0009)
 
+`domain` is never free-form: it must match an `active` `tenant_domains` row for
+the authenticated tenant (ADR-0009). Domain sets are per-tenant and vary in
+size. The key itself is immutable — it is denormalized into every Qdrant point
+payload and into `source_files`, so renaming it is a migration, not an edit.
+
 UUID primary keys (app-generated), `timestamptz` for all timestamps,
 `Numeric(18, 8)` for money (never floats), `JSONB` for flexible metadata but
 typed/indexed columns for query-critical fields, string status columns with
@@ -206,7 +352,24 @@ Postgres remains system of record for tenants, API keys, audit, jobs,
 `graph_runs`, `llm_calls`/`llm_pricing`. Correlate the two via `request_id`,
 `tenant_id`, `thread_id`, `run_id`. Use `structlog` with stable event names
 and structured fields (`logger.info("graph.run.completed", ...)`), not
-interpolated prose; JSON logs by default in production.
+interpolated prose; JSON logs by default in production, plus an optional
+local-only JSON file sink independent of the console renderer (`LOG_FILE_PATH`).
+
+**Add logging in the same change that adds the code, not as a follow-up.**
+When you add a new service-level entry point (an `application/` function a
+route calls directly, an ingestion phase, a mutation) or a new failure branch
+inside one, add its `logger.*` event in that same diff, using ADR-0011's
+level/event-naming table. Deferring it means re-deriving the failure modes and
+field names later from code that no longer has them in working memory — as
+happened with `src/application/files/upload.py`, where four failure branches
+(`parse_failed`, `chunk_limit_exceeded`, `embedding_failed`, `index_failed`)
+shipped with no log event and had to be retrofitted.
+
+This does not mean logging every function. Pure functions, models, schemas,
+and repositories (`infrastructure/postgres/repositories/`) stay silent by
+convention — the caller that turns their result into a business-meaningful
+outcome (job succeeded, upload rejected, domain disabled) is where the event
+belongs, not the row-level function underneath it.
 
 ## Testing (ADR-0016)
 
@@ -220,9 +383,14 @@ interpolated prose; JSON logs by default in production.
 - Test naming: `test_<unit>_<scenario>_<outcome>`, Arrange–Act–Assert.
 - Layout mirrors architecture: `tests/unit/{application,agent}`,
   `tests/integration/{postgres,minio,qdrant}`, `tests/e2e/`.
-- Integration tests use **Testcontainers** (never a developer's local
-  services or Langfuse-owned storage/credentials) — this is the standard
-  automated mechanism, not Docker Compose. Isolate data per test via unique
+- Integration **and e2e** tests use **Testcontainers** (never a developer's
+  local services or Langfuse-owned storage/credentials) — this is the standard
+  automated mechanism, not Docker Compose. Compose is reserved for exactly one
+  thing: the serialized operational smoke test of the *running web process*
+  (`scripts/smoke.sh`), which is gated out of `uv run pytest`. Shared container
+  fixtures live in `tests/support/containers.py`, registered from the root
+  `tests/conftest.py` via `pytest_plugins` (a non-root conftest cannot declare
+  it). Isolate data per test via unique
   keys/queue/collection names; parallel integration execution is disabled
   until fixture isolation is proven safe.
 - Pytest never calls a live/paid model provider in routine runs — that's

README.md

@@ -2,6 +2,49 @@
 
 Architecture decisions live in [`docs/adr`](docs/adr). The first implementation
 milestone is documented in the [ingestion vertical-slice plan](docs/plans/001-ingestion-vertical-slice.md).
+Day-to-day operation — tuning the ingestion bounds, the proxy timeout
+requirement, and how to investigate or retry a failed upload — is the
+[operator runbook](docs/runbook.md).
+
+## Provisioning the datastores
+
+Both schema steps run as explicit deployment steps. The application performs no
+DDL at startup — not for Postgres (ADR-0009) and not for Qdrant (ADR-0001,
+"Collection provisioning").
+
+```bash
+docker compose up -d                          # Postgres, MinIO, Qdrant
+uv run alembic upgrade head                   # Postgres schema
+uv run python -m src.cli.qdrant_bootstrap     # the `chunks` collection
+uv run fastapi dev src/main.py
+```
+
+Nothing over HTTP can create the first tenant — every `/v1` route needs an API
+key, and a key cannot exist before its tenant. One command issues both, plus any
+domains, printing the key once (only its hash is stored):
+
+```bash
+uv run python -m src.cli.provision_tenant --slug acme --domain fire
+```
+
+Before a tenant can upload, its domains must be registered — `POST /v1/files`
+rejects an unregistered or disabled `domain` with `400`. The calling backend
+manages them over `/v1/domains` using a key with the `domains:write` scope:
+
+```bash
+curl -X POST http://localhost:8000/v1/domains \
+  -H "Authorization: Bearer $API_KEY" \
+  -H 'Content-Type: application/json' \
+  -d '{"domain": "fire", "display_name": "Fire insurance"}'
+```
+
+Both bootstrap commands are idempotent and safe to re-run. `qdrant_bootstrap` verifies an
+existing collection against the pinned schema and exits non-zero on a mismatch,
+rather than leaving a silently degraded sparse index in place.
+
+`./scripts/smoke.sh` verifies the whole path — Compose up, both deployment
+steps, provisioning, an upload through the running web process to indexed Qdrant
+points. See the [runbook](docs/runbook.md#12-verifying-a-deployment).
 
 ## Local Langfuse
 

alembic/versions/41335d162de8_create_tenant_domains.py

@@ -0,0 +1,48 @@
+"""create tenant_domains
+
+Revision ID: 41335d162de8
+Revises: bfc6c81c2542
+Create Date: 2026-08-20 17:48:29.443293
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = '41335d162de8'
+down_revision: Union[str, Sequence[str], None] = 'bfc6c81c2542'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Upgrade schema."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table('tenant_domains',
+    sa.Column('id', sa.Uuid(), nullable=False),
+    sa.Column('tenant_id', sa.Uuid(), nullable=False),
+    sa.Column('domain', sa.String(length=80), nullable=False),
+    sa.Column('display_name', sa.String(length=200), nullable=False),
+    sa.Column('status', sa.String(length=20), server_default='active', nullable=False),
+    sa.Column('metadata', postgresql.JSONB(astext_type=sa.Text()), server_default='{}', nullable=False),
+    sa.Column('created_at', sa.DateTime(timezone=True), server_default=sa.text('now()'), nullable=False),
+    sa.Column('updated_at', sa.DateTime(timezone=True), server_default=sa.text('now()'), nullable=False),
+    sa.Column('disabled_at', sa.DateTime(timezone=True), nullable=True),
+    sa.CheckConstraint("status IN ('active', 'disabled')", name='ck_tenant_domains_status'),
+    sa.ForeignKeyConstraint(['tenant_id'], ['tenants.id'], ondelete='CASCADE'),
+    sa.PrimaryKeyConstraint('id'),
+    sa.UniqueConstraint('tenant_id', 'domain', name='uq_tenant_domains_tenant_id_domain')
+    )
+    op.create_index(op.f('ix_tenant_domains_tenant_id'), 'tenant_domains', ['tenant_id'], unique=False)
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade schema."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_index(op.f('ix_tenant_domains_tenant_id'), table_name='tenant_domains')
+    op.drop_table('tenant_domains')
+    # ### end Alembic commands ###

docs/adr/0001-ingestion-pipeline-and-collection-schema.md

@@ -49,7 +49,7 @@ One collection, e.g. `chunks`, shared by all tenants and domains.
 | Name | Type | Purpose | Notes |
 |---|---|---|---|
 | `dense_nomic` | dense vector | primary semantic similarity (multilingual, incl. Persian) | `nomic-embed-text-v2-moe`, 768-dim ([0004](0004-docx-csv-chunking-strategy.md)) |
-| `dense_openai` | dense vector | second semantic signal | OpenAI large embedding model (e.g. `text-embedding-3-large`), dimension per OpenAI's `dimensions` param (TBD — full 3072 vs. a truncated size) |
+| `dense_openai` | dense vector | second semantic signal | `text-embedding-3-large` at its **native 3072 dimensions** — the `dimensions` param is deliberately left unset (see below) |
 | `sparse` | sparse vector | lexical/keyword-sensitive retrieval | `bm25-fa-norm-stop` — Qdrant FastEmbed's BM25 sparse encoder configured for Persian (stopword removal + normalization), not a separately trained model |
 | `late_interaction` | multivector | reserved for late-interaction rerank ([0003](0003-agent-hybrid-retrieval.md)) | `jina-colbert-v2` ([0005](0005-reranking-model-and-sparse-analyzer-selection.md)), `comparator: max_sim`, `hnsw_config: m=0` (rerank-only, never independently ANN-searched), stored **on disk** |
 
@@ -61,6 +61,35 @@ dense/sparse query latency. Two dense vectors are provisioned deliberately —
 `dense_nomic` and `dense_openai` are two independent semantic signals, both
 prefetched and fused at query time (ADR-0003), not a primary/fallback pair.
 
+#### Dense model endpoints and dimensions (resolved by the `emet` benchmark)
+
+Both dense models are reached over the **same OpenAI-compatible
+`/embeddings` API**, so one adapter
+(`src/infrastructure/embedding/openai_compatible.py`) serves both named
+vectors with different configuration:
+
+| Named vector | Model | Endpoint | Dimensions |
+|---|---|---|---|
+| `dense_nomic` | `nomic-embed-text-v2-moe` | self-hosted Ollama OpenAI-compat shim | **768** (verified against the live endpoint) |
+| `dense_openai` | `text-embedding-3-large` | OpenAI hosted API | **3072** (native; `dimensions` unset) |
+
+`dense_openai`'s dimension was previously listed as an open dependency. It is
+now pinned to the native 3072, because that is the configuration the `emet`
+lab benchmarked — it never passed a `dimensions` argument. Setting it later
+would truncate via Matryoshka and is a **re-embedding migration, not a config
+tweak**, exactly as the negative consequence below warns.
+
+Two operational notes about the self-hosted embedder, both learned by
+measurement rather than assumption:
+
+- **Cold load exceeds 150s**, far beyond `INGESTION_TIMEOUT_SECONDS`, so an
+  idle-then-upload would return `504`. Mitigated on both ends: Ollama's
+  `keep_alive` keeps the model resident, and the FastAPI lifespan warms each
+  dense embedder at startup (fail-soft — a down embedder must not block boot).
+- **Once warm it is fast**: ~0.30s for one input and ~0.34s for a batch of 16.
+  Batching is therefore nearly free, which is what keeps ADR-0017's inline
+  ingestion viable.
+
 ### Multitenancy / indexing config
 
 - HNSW: `m: 0` (disable the global index) + `payload_m: 16`, per Qdrant's
@@ -79,6 +108,38 @@ prefetched and fused at query time (ADR-0003), not a primary/fallback pair.
 - Payload index on `previous_chunk_id` / `next_chunk_id`: keyword index,
   used for O(1) adjacency retrieval (see below).
 
+### Collection provisioning
+
+The collection is created by an explicit **deployment step**, not by application
+startup and not lazily on first write:
+
+    uv run python -m src.cli.qdrant_bootstrap
+
+Creating a collection is DDL, and this project already keeps DDL out of the boot
+and request paths: [0009](0009-postgres-sqlalchemy-alembic-schema.md) requires
+Alembic for Postgres schema and forbids `create_all()` at startup, and
+[0012](0012-application-resource-lifetime-and-dependency-ownership.md) makes
+LangGraph's `.setup()` a deployment step for the same reason. Neither ADR named
+Qdrant explicitly; this section closes that gap rather than letting the placement
+be decided by whichever code happened to need it first.
+
+Doing it in the FastAPI lifespan was rejected: it couples process boot to Qdrant
+being reachable (which is `/readyz`'s job, not boot's), races across replicas,
+and turns a misconfigured collection into a silent skip. Doing it lazily on first
+upsert was rejected for putting DDL on a user request and hiding the
+misconfiguration until traffic arrives.
+
+`ensure_chunks_collection` is idempotent and **verifying**: against an existing
+collection it compares the dense dimensions and the sparse `modifier` to the
+pinned values and fails loudly on divergence. That check is the point of making
+the step explicit — both properties degrade silently in production if wrong (a
+missing `modifier="idf"` produces no error, just unweighted lexical retrieval).
+
+Payload indexes are (re)created on every run, since unlike vector configuration
+they can be added to a live collection. The full-text index on `content` is
+therefore deferred to the keyword-search work in
+[0002](0002-chunk-crud-and-search-api.md), not created here.
+
 ### Payload schema
 
 This schema is now decided for the fields below. Additional document-context
@@ -95,7 +156,7 @@ involves format-specific tradeoffs not yet made.
 | `chunk_id` | keyword | stable identifier for a single chunk |
 | `content_type` | keyword | classification of the chunk's content; exact value set (e.g. `paragraph`, `table_row`, `heading`) to be finalized alongside the chunking-strategy ADR |
 | `source_filename` | keyword | original uploaded filename |
-| `source_type` | keyword (`docx` \| `csv`) | which parser produced this chunk |
+| `source_type` | keyword (`docx` \| `xlsx` \| `csv`) | which parser produced this chunk — `xlsx` added by [0018](0018-docx-and-spreadsheet-parsing-with-fixed-size-chunking.md) |
 | `order_id` | float (see below) | chunk's *display* position within the file; mutable so the backend can reorder/insert chunks |
 | `chunk_index` | integer | chunk's *original ingestion* ordinal — immutable, used to derive the deterministic point ID below (kept separate from `order_id` precisely because `order_id` can change) |
 | `previous_chunk_id` | keyword, nullable | `chunk_id` of the preceding chunk in display order (`null` for the first chunk in a file) — O(1) adjacency pointer for context-window expansion in ADR-0003 |
@@ -107,7 +168,7 @@ involves format-specific tradeoffs not yet made.
 | `updated_at` | datetime | last modification timestamp |
 | `created_by` | keyword | user/service that created the chunk |
 | `updated_by` | keyword | user/service that last modified the chunk |
-| `version` | integer | optimistic-concurrency counter, used in ADR-0002 |
+| `version` | integer | optimistic-concurrency counter, used in ADR-0002. Ingestion currently writes `1` unconditionally: the read-check-write that makes the guard meaningful costs one read per point and belongs with the `/v1/points` write paths, so plan 002 owns it. Safe while ingestion is the only writer of a file's points; it would clobber a concurrent manual edit's counter once `/v1/points` ships. |
 | `content_hash` | keyword | hash of the chunk's raw text; lets re-ingestion detect unchanged content and skip re-embedding it |
 | `embedding_model_version` | keyword | identifies which embedding model(s) produced this chunk's vectors; needed to know which chunks require re-embedding after a future model swap |
 
@@ -212,9 +273,14 @@ them — see ADR-0002 for how reorder/insert/delete operations keep
   ingestion time and both are queried at retrieval time — roughly double
   the dense embedding cost/latency of a single-dense-vector design, plus an
   external network dependency on OpenAI's API in the ingestion path.
-- `dense_openai`'s exact output dimension is still an open dependency that
-  should be pinned before ingestion is implemented — changing it later is a
-  re-embedding migration, not a config tweak.
+- ~~`dense_openai`'s exact output dimension is still an open dependency~~ —
+  **resolved**: pinned to the native 3072 (see "Dense model endpoints and
+  dimensions" above). The warning still stands for any future change:
+  re-dimensioning is a re-embedding migration, not a config tweak.
+- The `sparse` vector must be created with `modifier="idf"`. The client
+  computes only BM25's term-frequency saturation; without that modifier
+  Qdrant applies no IDF at all and lexical retrieval silently degrades
+  (ADR-0005).
 - `jina-colbert-v2` ([0005](0005-reranking-model-and-sparse-analyzer-selection.md))
   adds a hard GPU dependency to ingestion (not just query time, since the
   document-side multivector is computed here) and its commercial license is

docs/adr/0002-chunk-crud-and-search-api.md

@@ -75,6 +75,51 @@ retrieval used by the AI agent in ADR-0003; the two "search" concepts serve
 different callers (a human/admin managing chunks vs. an agent retrieving
 context) and should not be conflated in the API or in future discussion.
 
+Two properties follow from the index being a *filter*: results carry no
+relevance score, and their order is unspecified. The API therefore returns
+neither a score field nor a ranked list, and callers must not read the array
+order as relevance. A caller that wants ranking wants ADR-0003's path.
+
+#### The query is normalized the way ingested content was
+
+`normalize_persian_text` (ADR-0018) folds Arabic letterforms to their Persian
+equivalents — U+064A to U+06CC, U+0643 to U+06A9 — on every text block before
+chunking, so stored `content` is uniformly Persian-formed. A query string is
+not chunk content and never passes through that path, so a term typed on an
+Arabic keyboard reaches the index as a different codepoint sequence than the
+document it should match.
+
+The service therefore applies the same folding to the query before matching.
+Without it the endpoint fails in the worst available way: an exact-looking
+query returns an empty result set, with no error, no warning, and nothing in
+the logs to distinguish "no such term" from "the term is spelled with the
+other yeh". Note this is a *query-side* transformation only — it changes what
+is compared, never what is stored.
+
+This does not extend to stemming or synonyms. Qdrant's full-text index offers
+neither, and adding a Farsi analyzer here would duplicate the benchmarked BM25
+sparse pipeline (ADR-0005) in a code path that is not benchmarked against
+anything.
+
+#### Listing is scoped to one file, and paginates by `order_id`
+
+`GET /points?file_id=...` requires `file_id` rather than treating it as one
+optional filter among several, and its pagination cursor is an `order_id`
+value rather than an offset. Both follow from `order_id` being per-file:
+
+- A cursor is only meaningful against a totally ordered key. `order_id` orders
+  points within one file and says nothing across files, so an unscoped listing
+  has no stable sort to paginate along.
+- An offset cursor is wrong even within one file. Insert, reorder, and delete
+  all shift positions, so a page-two request issued after a concurrent insert
+  ahead of the cursor would repeat a row already returned — silently. Ranging
+  on `order_id > cursor` is unaffected: the reader has passed that value, and a
+  point inserted behind it was already served.
+
+The second point depends on `order_id` being unique within a file, which the
+gap-exhaustion rule below preserves by rejecting a reorder whose computed gap
+would collapse onto a neighbour value.
+
 ### Delete is soft by default
 
 `DELETE /points/{point_id}` and `DELETE /points?file_id=...` set
@@ -105,6 +150,41 @@ Qdrant's `update_filter`, giving an optimistic-concurrency-style guard
 against races between a concurrent ingestion re-run (ADR-0001) and a manual
 edit through this API.
 
+### Re-ingestion versus manual edits
+
+A file can be re-uploaded after someone has hand-edited one of its points
+through this API. **The newly ingested file wins.** Ingestion is authoritative
+for the content of the file it ingested; a manual edit is a correction that
+survives only until the source document is replaced.
+
+Concretely:
+
+- A point that still exists in the new version (same `file_id` +
+  `chunk_index`, hence the same deterministic point ID) is **overwritten in
+  place**. Ingestion performs a read-check-write so `version` is incremented
+  from whatever the manual edit left it at, rather than reset to `1`.
+- A point from the previous ingestion that is **absent** from the new version
+  is flagged `is_active: false` with `deleted_at` set. It is never removed
+  from Qdrant — the soft-delete rule above applies to re-ingestion exactly as
+  it applies to `DELETE`.
+- A manually created point (`POST /points`) is assigned a `chunk_index` past
+  the ingested range, so the same sweep deactivates it on the next upload of
+  its file. This is the intended consequence of "the new file wins", not an
+  accident of the sweep's bounds.
+
+Because the point ID is derived from the immutable `chunk_index`, an
+overwritten point cannot hold both the manual edit and the new file's content.
+The clobbered content is therefore recorded in `point_audit_events`
+(ADR-0009) as a `reingest_overwrite` operation carrying `before_version`, so
+the edit is recoverable from the audit trail even though it is no longer a
+live point.
+
+Rejected alternative: preserving manual edits by having ingestion skip points
+with `version > 1`. It breaks the guarantee that a successful upload leaves
+Qdrant matching the uploaded document, and it needs a second, separate rule
+for edited points that no longer exist in the new version — two divergent
+notions of authority over one file.
+
 ### Re-embedding on content edit
 
 `PUT /points/{point_id}` can change `content`, which leaves the stored

docs/adr/0004-docx-csv-chunking-strategy.md

@@ -62,6 +62,23 @@ from its model card: 768-dim output, Matryoshka-truncatable down to 256;
 every embedded string — `search_document: ` at ingestion time, `search_query: `
 on the agent's query side (ADR-0003).
 
+> **Amendment — the task prefix is currently not applied.** The `emet`
+> benchmark that selected this model ran *without* any prefix: its Ollama
+> deployment's template is a bare `{{ .Prompt }}` passthrough that injects
+> nothing, which was verified directly against the running endpoint. The
+> prefix is not cosmetic — embedding the same Persian text with and without
+> `search_document: ` yields a cosine of only **0.5741** — so applying it at
+> ingest while the query side omits `search_query: ` would make retrieval
+> *worse* than using neither.
+>
+> Implementation therefore defaults `EMBEDDING_NOMIC_DOCUMENT_PREFIX` to
+> empty, matching the measured configuration, and exposes it as config so the
+> prefixed variant is a one-line experiment rather than a code change. The
+> model card remains the reason to expect prefixing to help; what is missing
+> is evidence on *this* corpus. Turning it on is a paired change — ingest and
+> query must move together — and should be settled by an emet run that
+> measures the pair, not by an unmeasured edit here.
+
 ## Decision
 
 ### Parsing order: structural extraction before chunking

docs/adr/0005-reranking-model-and-sparse-analyzer-selection.md

@@ -2,9 +2,10 @@
 
 ## Status
 
-Proposed — the fusion/rerank *shape* and reranker model are decided; the
-final BM25 analyzer and the commercial license status of the reranker are
-still open per the follow-up items below.
+Proposed — the fusion/rerank *shape*, the reranker model, and (as of the
+`emet` benchmark, see "Benchmark outcome" below) the **BM25 analyzer** are
+decided. The commercial license status of the reranker remains open per the
+follow-up items below.
 
 ## Context
 
@@ -100,6 +101,50 @@ entirely to the analyzer stage, not the ranking formula:
 consistent with Farsi's high density of function words (ezafe particles,
 prepositions, common verbs) adding TF/IDF noise if left in.
 
+### 3a. Benchmark outcome: `bm25-fa-norm-stop` confirmed, and where the BM25 math runs
+
+The `emet` evaluation lab (`~/code/talie/emet`) ran the four-variant
+comparison above against the real Farsi corpus and confirmed
+**`bm25-fa-norm-stop`** as the winner. It is the only sparse variant promoted
+into emet's hybrid matrix (`emet/hybrid.yaml`). This closes follow-up item 4
+below.
+
+The winning analyzer is a specific, reproducible artifact, ported into
+`src/infrastructure/embedding/analyzers.py` and verified token-for-token
+against emet's implementation. Its details are load-bearing:
+
+- Unicode **NFC** (not NFKC), then ZWNJ → space, then Persian/Arabic-Indic
+  digits → ASCII, then `ي→ی ك→ک ة→ه ؤ→و إ→ا أ→ا`.
+- Tokenizer `[^\W_]+`, which **keeps digits**. This matters for an insurance
+  corpus: policy numbers, dates, and amounts are exactly the terms lexical
+  retrieval should match, and the digit folding above means a query in ASCII
+  digits matches a document authored in Persian ones.
+- A 51-entry stopword set (40 Persian/Arabic + 11 English, the corpus being
+  mixed-script). Deliberately not a full `hazm` list.
+- No stemming — `fa_norm_stem` was the losing arm.
+
+**The BM25 formula is split across two systems, deliberately.** The client
+applies term-frequency saturation, including the `k`/`b` document-length
+normalization; **IDF is supplied by Qdrant** via `modifier="idf"` on the
+sparse vector field, computed from collection-wide statistics rather than
+from a fixed client-side corpus.
+
+That split is a correctness trap worth stating plainly: a `chunks` collection
+created *without* `modifier="idf"` will score these vectors as saturated term
+frequencies with no IDF weighting at all — no error, no warning, just
+materially worse lexical retrieval. The collection bootstrap must set it.
+
+Document and query encoding are asymmetric in exactly one term: documents
+carry the `b` length normalization, queries do not (standard BM25 practice).
+Both sides must therefore encode through the same implementation, which is
+why the sparse port carries a `query` flag rather than leaving retrieval to
+grow a second, silently divergent encoder.
+
+Term → sparse-index mapping is `blake2b(token, digest_size=8) % (2**31 - 1)`,
+a pure hash with no vocabulary table, so it needs no shared state and stays
+identical across processes and between ingest and query time. Changing the
+hash orphans every stored sparse vector: that is a re-ingestion, not a deploy.
+
 ### 4. BM25 parameters: keep `k=1.2`, `b=0.75`; tune analyzer, not formula
 
 These are standard, well-validated defaults (Trotman, Puurula & Burgess,
@@ -159,8 +204,21 @@ comparison and `b` sweep in the follow-ups below.
 3. Run an ablation: single dense model + sparse + rerank vs. the current
    dual-dense-model + sparse + rerank setup, on real Farsi queries, to
    justify (or drop) the second dense vector (`dense_openai`).
-4. Compare `bm25-fa-norm-stop` vs. `bm25-fa-norm-stem` in isolation to
-   determine whether gains come from stopword removal, stemming, or both.
+4. ~~Compare `bm25-fa-norm-stop` vs. `bm25-fa-norm-stem` in isolation~~ —
+   **done**, see "Benchmark outcome" above. `fa_norm_stop` won; stemming was
+   not adopted.
 5. Sweep BM25 `b` (e.g. 0.5–0.9) for the winning analyzer, since document
    length varies significantly across the corpus (short chat messages vs.
    long articles) and `0.75` is a generic default, not corpus-tuned.
+6. **Recalibrate `avg_len`.** The client-side `b` term needs an average
+   document length in *analyzer tokens*. The ported value (256.0) is emet's
+   own placeholder, and emet measured it over short Q&A records rather than
+   this service's ~400-token chunks, so it is very likely miscalibrated here.
+   Exposed as `EMBEDDING_SPARSE_AVG_LEN` so it can be corrected from real
+   corpus statistics without a code change.
+7. **Re-benchmark the analyzer with diacritic stripping.** `fa_norm_stop`
+   does not remove harakat or tatweel, so `ســلام` and `سلام` are distinct
+   terms. `src/application/ingestion/normalization.py` already strips both
+   for chunk *content*; extending that to the analyzer is plausibly an
+   improvement but would deviate from the measured configuration, so it
+   belongs in an emet run rather than an unmeasured edit.

docs/adr/0009-postgres-sqlalchemy-alembic-schema.md

@@ -98,13 +98,16 @@ One row per customer/tenant.
 | `slug` | Stable short name, unique, human-readable. |
 | `name` | Display name. |
 | `status` | `active` \| `suspended` \| `deleted`. Suspended tenants authenticate to a clear error but cannot run work. |
-| `settings` | JSONB for tenant-level feature flags/limits (max upload size, enabled file types, allowed domains, etc.). |
+| `settings` | JSONB for tenant-level feature flags/limits (max upload size, enabled file types, etc.). Allowed domains were previously listed here as well; they live in `tenant_domains` instead, per this ADR's own rule that query-critical fields get typed columns — `domain` is validated on every upload and filtered on every query. |
 | `created_at`, `updated_at`, `deleted_at` | Audit/soft-delete timestamps. |
 
 #### `tenant_domains`
 
-Optional but recommended. Validates the `domain` values used throughout Qdrant
-payloads (`car`, `fire`, etc.) per tenant.
+**Required.** (Previously "optional but recommended"; implemented and made
+mandatory alongside `/v1/domains`.) Validates the `domain` values used
+throughout Qdrant payloads (`car`, `fire`, etc.) per tenant. Domain sets are
+per-tenant and differ in size — one tenant may run 14 insurance lines and
+another 6 — so this is data, not an enum.
 
 | Column | Notes |
 |---|---|
@@ -116,7 +119,37 @@ payloads (`car`, `fire`, etc.) per tenant.
 | `metadata` | JSONB for domain-specific ingestion/retrieval settings. |
 
 This prevents arbitrary caller-supplied domains from silently creating new
-partitions in Qdrant.
+partitions in Qdrant. The failure it guards against is quiet: a typo such as
+`fier` for `fire` produces no error anywhere — the file is stored, parsed,
+embedded, and indexed into a partition retrieval never queries, so it is
+invisible rather than failed.
+
+##### Enforcement and management
+
+- **Strict allowlist.** `POST /v1/files` rejects a domain with no `active` row
+  for the tenant (`400`, error code `unknown_domain`). There is no auto-create
+  on first use: that would record the typo rather than prevent it. The check
+  runs inside the upload's first transaction, before any MinIO object, job row,
+  or Qdrant point is written.
+- **Managed over the API, not by an operator.** `/v1/domains` (list, create,
+  update, disable, enable) is the surface the calling backend uses. Domains are
+  created by an explicit, scoped call rather than as a side effect of an upload
+  — that distinction, not who makes the call, is what "strict" means here.
+- **Its own scope.** `domains:read`/`domains:write`, deliberately separate from
+  `files:write`. Folding domain creation into the upload scope would let an
+  upload key create partitions again, which is the exact hole this closes.
+  `api_keys.scopes` is already a free JSONB list, so this needs no schema change.
+- **`tenant_id` stays derived from the API key.** One key per tenant; nothing
+  request-suppliable. A platform key acting across tenants would need a real
+  actor model and is not adopted.
+- **`domain` is immutable; `display_name` is not.** The key is denormalized into
+  every Qdrant point payload and into `source_files`, so renaming it means
+  rewriting all of them — a migration, not a `PATCH`. The update schema
+  therefore has no `domain` field.
+- **Disable is not delete.** `status='disabled'` blocks new uploads and hides
+  the domain from listings, leaving already-indexed points intact and
+  retrievable. Actual removal needs the retention/erasure workflow this ADR and
+  plan 001 defer.
 
 #### `api_keys`
 

docs/adr/0011-python-structured-logging-with-structlog.md

@@ -70,17 +70,30 @@ logger.info(
 Do not build log messages by interpolating operational metadata into prose.
 Prefer fields over long strings because fields are queryable.
 
-### Emit JSON logs by default in production
+### Emit JSON logs by default in production; console and file are independent sinks locally
 
 Production logs are JSON on stdout so process managers, container runtimes, and
-log collectors can ingest them directly. Local development may use a colored
-console renderer controlled by configuration.
+log collectors can ingest them directly. This does not change.
 
-File logging is optional and mainly for local development. If enabled, it must
-use explicit rotation settings such as `maxBytes` and `backupCount`. Do not rely
-on a default `RotatingFileHandler` with no rotation parameters. In containerized
-production, stdout/stderr collection is preferred over writing `logs/app.log`
-inside the application container.
+Locally, stdout and an optional file are two **independent, simultaneous**
+handlers on the same logger, not a single renderer chosen by a flag — the same
+structlog event fans out to both:
+
+- **Console handler**: always on, `structlog.dev.ConsoleRenderer(colors=True)`.
+  This is what a developer reads while the process runs, so it stays
+  human-readable regardless of whether file logging is also enabled.
+- **File handler**: off by default, enabled by setting `LOG_FILE_PATH`. Always
+  renders JSON (`structlog.processors.JSONRenderer()`), independent of the
+  console handler's renderer, so a saved log is machine-parseable even though
+  the terminal output next to it is not. Must use explicit rotation
+  (`RotatingFileHandler` with `maxBytes`/`backupCount` — never an unrotated
+  handler).
+
+In containerized production, stdout/stderr collection remains preferred over
+writing `logs/app.log` inside the application container, so `LOG_FILE_PATH` is
+expected to be unset there; the file handler exists for local development,
+where reading a colored terminal *and* keeping a JSON trail to grep/parse later
+are both useful at once.
 
 ### Configure stdlib and structlog together
 
@@ -188,6 +201,36 @@ Notes:
 - `structlog.contextvars.merge_contextvars` ensures request-bound fields appear
   on both structlog and stdlib logs processed through the formatter.
 
+### Bind process-level environment context once at startup
+
+Deployment identity — which build is running, in which environment, on which
+instance — answers a different question than request correlation: "is this
+issue specific to one deployment / one region / one instance?" rather than "is
+this issue specific to one request?" It does not vary per request, so it must
+not go through `structlog.contextvars`, which `RequestIdMiddleware` clears on
+every request; a value bound there before the first request would be wiped the
+moment that middleware runs.
+
+Instead, add a static structlog **processor** — a plain closure over values read
+once at `configure_logging()` time — so it runs on every event regardless of
+request context:
+
+```python
+def _bind_environment(settings: AppLimitSettings):
+    def processor(logger, method_name, event_dict):
+        event_dict["env"] = settings.env
+        event_dict["service_version"] = settings.service_version
+        return event_dict
+
+    return processor
+```
+
+`service_version` should be the deployed commit SHA or release tag (e.g. from a
+`GIT_SHA`/`APP_VERSION` build-time env var — not computed at runtime by
+shelling out to `git`). This makes "is this only happening on the new
+deployment?" answerable directly from logs, without cross-referencing a
+separate deployment record.
+
 ### Bind request context with contextvars
 
 At FastAPI ingress, clear stale context, bind request identifiers, and return the

docs/adr/0017-synchronous-ingestion-in-the-request-path.md

@@ -159,8 +159,25 @@ retry, and phase 2 has no transaction protecting it:
   return the existing file/job rather than re-ingesting (plan 001).
 - `tenant_id` comes from `AuthContext`, never from the request body.
 - A terminal job is never transitioned back to `running`.
-- Qdrant points from a failed attempt do not replace the previous successful
-  index; replacement happens only after a successful attempt.
+- A failed attempt never *removes* content from a working index. The
+  soft-delete sweep that retires a shortened file's leftover points runs only
+  after every upsert in the attempt has succeeded.
+
+  This is deliberately weaker than "replacement happens only after a successful
+  attempt", which an earlier revision of this ADR claimed. That guarantee is not
+  achievable alongside ADR-0001's deterministic point ids: those ids are exactly
+  what makes a retry idempotent, and they also mean a re-ingestion overwrites
+  points **in place**, so a crash partway through leaves a prefix updated and the
+  remainder still on the old content. Buying literal atomicity would mean
+  generation-suffixed ids and an activation flip, which contradicts ADR-0001 and
+  ADR-0002's stable point ids. Staging the new points as `is_active=false` and
+  flipping them on success is strictly worse — the in-place overwrite would
+  deactivate the previously live points, silently emptying a working index if the
+  attempt were interrupted.
+
+  What holds instead: the index is never emptied, never partially deleted, and a
+  retry converges — deterministic ids rewrite every point and the sweep re-runs,
+  reaching the exact correct state.
 
 ### Failures are HTTP failures
 

docs/plans/001-ingestion-vertical-slice.md

@@ -125,9 +125,12 @@ them:
 - Use `(tenant_id, domain, content_sha256)` to recognize identical uploads.
 - An identical active upload should return the existing source-file/job reference
   rather than create a duplicate ingestion.
-- A changed upload creates a new ingestion job. Existing active Qdrant points are
-  replaced only after the new job completes successfully, so a failed re-ingestion
-  does not remove a working index.
+- A changed upload creates a new ingestion job. A failed re-ingestion never
+  removes a working index: the soft-delete sweep for a shortened file runs only
+  after every upsert has succeeded. Because ADR-0001's point ids are
+  deterministic, upserts overwrite in place, so an interrupted attempt can leave
+  a prefix updated — it cannot empty or partially delete the index, and a retry
+  converges. See ADR-0017, "Re-running an ingestion stays safe".
 - Preserve the original filename in Postgres metadata. MinIO object keys remain
   internal ID-based paths.
 
@@ -247,6 +250,16 @@ code and a terminal job row.
 
 ### Phase 5: Ingestion execution and Qdrant Chunk/Point CRUD
 
+> **Carried forward from Phase 4 — the `chunks` collection must create the
+> `sparse` vector with `modifier="idf"`.** The BM25 adapter computes only
+> term-frequency saturation client-side; IDF comes from Qdrant's
+> collection-wide statistics. Omit the modifier and there is no error and no
+> warning — sparse scoring silently loses its IDF term and lexical retrieval
+> degrades. See ADR-0005, "Benchmark outcome".
+>
+> Collection creation must also use the pinned dimensions from ADR-0001:
+> `dense_nomic` 768, `dense_openai` 3072.
+
 1. Implement the ingestion service called by the route, using the
    application-lifetime database, MinIO, Qdrant, model, and logging clients.
 2. Validate the persisted records before fetching the MinIO object.
@@ -262,6 +275,11 @@ code and a terminal job row.
    tenant-filtered upserts, terminal state persistence, retrying an upload, and
    parser/Qdrant failure handling.
 
+The `chunks` collection itself is provisioned by a deployment step —
+`uv run python -m src.cli.qdrant_bootstrap` — not by FastAPI startup, for the
+same reason ADR-0009 keeps Alembic out of startup and ADR-0012 makes LangGraph's
+`.setup()` a deployment step. See ADR-0001, "Collection provisioning".
+
 **Exit criteria:** a successful upload returns `201` with a terminal status, and
 its points are retrievable only under the owning tenant's Qdrant filter. A forced
 failure mid-ingestion produces a `failed` job and the right HTTP status, and
@@ -271,17 +289,28 @@ retrying the upload produces a correct final state without duplicate chunks.
 
 1. Add an operator runbook covering local startup, migrations, MinIO bucket
    setup, the run command, ingestion-bound tuning, the proxy/client timeout
-   requirement, and how to retry a failed ingestion.
+   requirement, and how to retry a failed ingestion. — `docs/runbook.md`.
 2. Add a serialized Compose-based operational smoke test covering upload through
    indexed points against the running web process. Testcontainers remains the
-   standard pytest mechanism for individual adapter integration tests.
+   standard pytest mechanism for individual adapter integration tests. —
+   `scripts/smoke.sh` driving `tests/e2e/test_compose_smoke.py`, which skips
+   itself unless `SMOKE_BASE_URL` is set so `uv run pytest` never invokes
+   Compose.
 3. Add end-to-end tests for duplicate upload, retrying a failed upload, tenant
-   isolation, capacity/timeout rejection, and failed parser/Qdrant behavior.
+   isolation, capacity/timeout rejection, and failed parser/Qdrant behavior. —
+   `tests/e2e/test_ingestion_slice.py`, on Testcontainers, in the default suite.
 4. Add health/readiness checks that distinguish process health from dependency
-   readiness.
+   readiness. — `/healthz` and `/readyz`; `/readyz` additionally requires the
+   `chunks` collection to exist, since a reachable but unbootstrapped Qdrant
+   would `502` on the first upload.
 5. Update the README with local-start instructions and links to ADRs, this plan,
    and the operations runbook.
 
+Provisioning a tenant and its first API key turned out to be a prerequisite for
+1 and 2 rather than a separate milestone: nothing over HTTP can create the first
+tenant, so `src/cli/provision_tenant.py` was added alongside the other two
+deployment-step commands.
+
 **Exit criteria:** a new developer can start the stack, apply migrations, upload a
 a document, observe the job through completion, and understand how to investigate or
 retry a failure.

docs/plans/002-point-crud-and-keyword-search.md

@@ -15,8 +15,9 @@ are; this document defines order, scope, and verification criteria.
 
 ## Prerequisite
 
-Plan 001 must be complete through **Phase 5** before Phase 3 of this plan
-starts. Specifically this plan depends on: the `chunks` collection and its
+Plan 001 is complete through Phase 6, so this prerequisite is satisfied. It
+required plan 001 through **Phase 5** before Phase 3 of this plan starts.
+Specifically this plan depends on: the `chunks` collection and its
 payload indexes actually existing, API-key authentication and `AuthContext`
 tenant derivation, the application-lifetime Qdrant client from the FastAPI
 lifespan, and the request-lifetime `AsyncSession` wiring. Phases 1–2 below
@@ -71,7 +72,9 @@ project owner accepts them, and update the ADR rather than diverging silently.
   (bulk soft delete of a file's points), from ADR-0008.
 - Soft delete as the default for every delete path, with neighbor relinking.
 - Optimistic concurrency on every mutating path via the `version` payload field.
-- Audit rows in Postgres for mutating operations.
+- Audit rows in Postgres for mutating operations: both ADR-0009 tables,
+  `api_request_logs` (one row per API call, written from the request middleware)
+  and `point_audit_events` with the real `api_request_log_id` foreign key.
 - Automated tests for tenant isolation, pointer integrity, concurrency
   conflicts, and pagination.
 
@@ -113,40 +116,22 @@ project owner accepts them, and update the ADR rather than diverging silently.
 9. Routers contain no Qdrant SDK calls and no filter construction. The Qdrant
    client is injected from the lifespan (ADR-0012).
 
-## Decisions needed before the affected phase
+## Decisions resolved before implementation
 
-### Re-embedding on content edit (blocks Phase 4)
+An earlier revision of this plan listed three open decisions here. All are now
+settled, and one further question this plan deferred to a Phase 6 test has been
+settled too. They are recorded in the ADRs — these lines are a pointer, not a
+second source of truth.
 
-`PUT /v1/points/{point_id}` can change `content`. The stored vectors then no
-longer match the text. Three options, in order of preference:
+| Question | Resolution | Recorded in |
+|---|---|---|
+| Re-embedding on content edit | Re-embed inline, reusing ingestion's ports and bounds and its `502`/`504` codes. The re-embed happens *before* the version-guarded write, so a stale edit still `409`s rather than re-embedding for nothing. | ADR-0002, "Re-embedding on content edit" |
+| Fractional-key exhaustion | No renormalize endpoint in this slice. Log `points.order_id.gap_low` under a safety threshold; reject with `409` and a distinct error code if the gap would collapse onto a neighbor value. Recovery is a runbook operation. | ADR-0002, "`order_id` gap exhaustion" |
+| Batch semantics | All-or-nothing, capped at 100 operations. Every operation's `version` precondition is validated before any is applied; one failure rejects the whole request and nothing reaches Qdrant. | ADR-0002, "`POST /points/batch` semantics" |
+| Re-ingestion versus manual edits | The newly uploaded file wins. Surviving points are overwritten in place with an incremented `version`; points absent from the new version are flagged inactive, never removed; manually created points sit past the ingested `chunk_index` range and are swept by the same rule. Clobbered content is recorded in `point_audit_events` as `reingest_overwrite`. | ADR-0002, "Re-ingestion versus manual edits" |
 
-1. **Re-embed inline** on content change, reusing plan 001's embedding ports and
-   bounds. Consistent, but puts embedder latency and `502`/`504` failure modes
-   on an admin edit path.
-2. **Require caller-supplied vectors** when content changes, and reject the edit
-   otherwise. Simple and honest, but pushes model knowledge to the client.
-3. **Mark the point stale** (a payload flag) and re-embed later. Needs
-   background work, which ADR-0017 currently rules out.
-
-Default to (1) for parity with ingestion, with the same batch/semaphore bounds
-and the same status codes. Record whichever is chosen in ADR-0002 before
-implementing Phase 4 — this is a real behavioral contract, not an
-implementation detail.
-
-### Fractional-key exhaustion
-
-ADR-0001 notes float keys eventually need renormalization. Decide now whether
-this slice ships a renormalize path (an internal operation rewriting a file's
-`order_id` values to `1000, 2000, 3000, ...`) or explicitly defers it with a
-logged warning when the gap between neighbors falls under a threshold. Deferring
-is acceptable; silently producing unrepresentable gaps is not.
-
-### Batch semantics
-
-`POST /v1/points/batch` must define, in the API schema and the tests: whether
-operations are all-or-nothing, what happens when operation 3 of 5 fails a
-version check, and the maximum operation count per request. Decide before
-Phase 5; do not let the answer be "whatever Qdrant happened to do."
+Phase 6's cross-slice end-to-end test therefore *verifies* the re-ingestion rule
+rather than forcing the decision.
 
 ## Build order
 
@@ -251,8 +236,8 @@ of mutations.
    ordering behavior, isolated per test by unique collection or tenant keys.
 2. An end-to-end test crossing plan 001 and this slice: ingest a CSV, list its
    points, reorder one, soft-delete another, re-upload the same file, and assert
-   the manual edits interact with re-ingestion exactly as ADR-0001/0002 specify.
-   If that interaction is not yet decided, this test is what forces the decision.
+   the manual edits interact with re-ingestion exactly as ADR-0002's
+   "Re-ingestion versus manual edits" specifies.
 3. Structured logging at the mutation boundary with stable event names
    (`points.updated`, `points.reordered`, `points.soft_deleted`) carrying
    `request_id`, `tenant_id`, `file_id`, and the resulting version.

docs/runbook.md

@@ -0,0 +1,303 @@
+# Operator runbook
+
+How to start this service, configure its ingestion bounds, and investigate or
+retry a failed upload. Architecture rationale lives in [`docs/adr/`](adr/); the
+implementation milestone is
+[plan 001](plans/001-ingestion-vertical-slice.md). This document covers
+operating what those describe.
+
+The service is a **single process with no background work**. `POST /v1/files`
+parses, chunks, embeds, and indexes inline and returns a terminal result
+(ADR-0017). There is no queue, no worker, and no automatic retry — the caller
+owns the retry decision, which makes the request's duration a deployment
+constraint. That fact drives most of this document.
+
+## 1. Prerequisites and local startup
+
+Docker, and [`uv`](https://docs.astral.sh/uv/) with Python 3.13.
+
+```bash
+cp .env.example .env      # non-secret local defaults; .env is gitignored
+uv sync
+docker compose up -d --wait
+```
+
+`docker-compose.yml` runs Postgres (`127.0.0.1:5433`), MinIO
+(`127.0.0.1:9100`, console `9101`), and Qdrant (`127.0.0.1:6343`). It is the
+local development stack and says so in its header — it is not a production
+deployment.
+
+## 2. Deployment steps
+
+Two schema steps run **before** the application, never at startup: FastAPI
+performs no DDL, for Postgres (ADR-0009) or for Qdrant (ADR-0001, "Collection
+provisioning"). Both commands and their reasoning are in the README's
+[Provisioning the datastores](../README.md#provisioning-the-datastores)
+section:
+
+```bash
+uv run alembic upgrade head                   # Postgres schema
+uv run python -m src.cli.qdrant_bootstrap     # the `chunks` collection
+```
+
+Both are idempotent. `qdrant_bootstrap` verifies an existing collection against
+the pinned schema and **exits non-zero on a mismatch** rather than leaving a
+silently degraded sparse index in place — the `sparse` vector's
+`modifier="idf"` and the pinned dense dimensions (768 / 3072) fail silently if
+wrong, which is why they are checked rather than assumed.
+
+Run both again after every deploy that ships a migration or a collection-schema
+change.
+
+## 3. MinIO bucket
+
+Under Compose the bucket already exists: the `app-minio` service's entrypoint
+runs `mkdir -p /data/${MINIO_BUCKET:-chatbot-source-files}` before starting the
+server, so first boot creates it. Nothing else needs to be done locally.
+
+Outside Compose, create the bucket named by `MINIO_BUCKET` before the first
+upload — the application never creates it. It must stay **private**; ADR-0013
+keeps source bytes non-public and this slice ships no download API or presigned
+URLs.
+
+## 4. Provisioning a tenant, an API key, and its domains
+
+Nothing over HTTP can bootstrap a tenant: every `/v1` route needs an API key,
+and a key cannot exist before its tenant. So the first key is issued by an
+operator command:
+
+```bash
+uv run python -m src.cli.provision_tenant \
+  --slug acme --domain fire --domain life --scopes files:write,domains:read
+```
+
+It prints `api_key=sk_...` **once**. Postgres stores only its SHA-256 hash
+(ADR-0009), so a lost key is reissued by re-running the command, never
+recovered. Structured logs carry only the non-secret `key_prefix` — a plaintext
+key must never reach a log sink (ADR-0011).
+
+Re-running with the same `--slug` reuses the tenant and any domains it already
+has, and issues an **additional** key. Both keys stay valid; this adds a key, it
+does not rotate one.
+
+Scopes are the security boundary between uploading, reading chunks, and
+managing the allowlist. Give an upload client `files:write` only. `domains:write`
+lets its holder create new domains, which is exactly what the allowlist exists to
+prevent an upload key from doing, and `points:read` lets its holder read the text
+of every chunk of every file — so an upload-only key gets neither.
+
+| Scope | Grants |
+|---|---|
+| `files:write` | Upload a document and read its ingestion status. |
+| `points:read` | Read, list, count, and keyword-search this tenant's points, including `GET /v1/files/{file_id}/points`. |
+| `points:write` | Create, edit, reorder, and soft-delete points (plan 002 Phases 3-5; no route uses it yet). |
+| `domains:read` / `domains:write` | Inspect and manage the domain allowlist. |
+| `admin` | Satisfies every scope check. |
+
+The command's `--scopes` default issues all of the above except `admin`, which
+suits a first operator key; narrow it explicitly for per-client keys.
+
+### Domains after the first one
+
+`POST /v1/files` rejects an unregistered or disabled `domain` with `400`
+(`unknown_domain`) before anything is written. Ongoing domain management is the
+`/v1/domains` API, under `domains:read` / `domains:write`:
+
+```bash
+curl -X POST http://localhost:8000/v1/domains \
+  -H "Authorization: Bearer $API_KEY" -H 'Content-Type: application/json' \
+  -d '{"domain": "fire", "display_name": "Fire insurance"}'
+```
+
+The `domain` key itself is immutable — it is denormalized into every Qdrant
+point payload and into `source_files`, so renaming it is a migration, not an
+edit (ADR-0009). Disabling a domain blocks new uploads; it does not delete
+existing points.
+
+## 5. Running the service
+
+```bash
+uv run fastapi dev src/main.py                                   # local, reload
+uv run uvicorn src.main:app --host 0.0.0.0 --port 8000           # deployed shape
+```
+
+Run more than one worker/replica only after reading §6: ingestion bounds are
+**per process**, so `INGESTION_MAX_CONCURRENCY` multiplies by the number of
+processes.
+
+## 6. Ingestion bounds and tuning
+
+Every bound is enforced server-side and maps to a status code. All are in
+`.env.example`. Ingestion is CPU- and network-bound in the request, so these are
+the numbers that decide whether the service degrades gracefully or falls over.
+
+| Setting | Bounds | On breach | Size it against |
+|---|---|---|---|
+| `INGESTION_MAX_CONCURRENCY` | Ingestions in flight **per process** | `503` + `Retry-After` | Memory per in-flight upload (whole file plus its chunks and vectors are resident) and the embedder's capacity. Rejecting is deliberate: ADR-0017 refuses rather than queues. |
+| `INGESTION_THREAD_POOL_SIZE` | Threads for blocking work (parse, chunk, hash, BM25, the sync `minio` SDK) | — (waits) | CPU cores. It exists to stop ingestion exhausting Starlette's own thread pool, so keep it below the total thread budget. |
+| `INGESTION_TIMEOUT_SECONDS` | The whole work phase | `504`, job marked `failed` | The slowest legitimate document, plus headroom. See §7 — this must stay under every read timeout in front of it. |
+| `INGESTION_MAX_UPLOAD_SIZE_MB` | Bytes accepted | `413` | Memory: the upload is read fully into the process before any work starts. |
+| `INGESTION_MAX_CHUNKS_PER_FILE` | Chunks per file, checked before embedding | `413` | Embedder cost/time per chunk × `INGESTION_TIMEOUT_SECONDS`. This is the real defence against one pathological file eating a slot. |
+| `INGESTION_EMBED_BATCH_SIZE` | Texts per embedder request | `502` on embedder failure | The provider's per-request limits. Batch before parallelizing. |
+| `INGESTION_EMBED_CONCURRENCY` | Concurrent embed batches | `502` | Provider rate limits and the self-hosted embedder's throughput. Never unbounded. |
+| `QDRANT_UPSERT_BATCH_SIZE` / `_CONCURRENCY` | Points per upsert and concurrent upserts | `502` (`index_error`) | Qdrant's ingest capacity; the batch size stays in ADR-0001's 64–256 band. |
+
+Two settings that look like tuning knobs but are not:
+
+- **`EMBEDDING_NOMIC_KEEP_ALIVE`** holds the self-hosted model resident. A cold
+  load of `nomic-embed-text-v2-moe` takes over 150 s — longer than any sane
+  `INGESTION_TIMEOUT_SECONDS` — so an idle period followed by an upload would
+  otherwise `504`. The lifespan also warms both dense embedders at startup for
+  the same reason.
+- **The BM25 analyzer and weights** (`EMBEDDING_SPARSE_*`) are a measured
+  artifact ported from the `emet` evaluation lab, verified token-for-token
+  against it (ADR-0005). Re-benchmark; do not tune them in place.
+
+## 7. The proxy and client read-timeout requirement
+
+**Every read timeout in front of this service must exceed
+`INGESTION_TIMEOUT_SECONDS`.** That includes the reverse proxy / ingress, any
+load balancer, and the calling backend's own HTTP client.
+
+If a proxy times out first, the client gets that proxy's error, the upload keeps
+running in the process, and the caller learns nothing about the outcome from the
+response. The job row still reaches a terminal status, so
+`GET /v1/files/{file_id}` remains the way to find out what happened — but the
+response contract is broken for that request. ADR-0017 names this the main cost
+of inline ingestion.
+
+A workable local ordering: client read timeout > proxy read timeout >
+`INGESTION_TIMEOUT_SECONDS`.
+
+## 8. Health and readiness
+
+| Endpoint | Question it answers | Use for |
+|---|---|---|
+| `GET /healthz` | Is the process alive? | Liveness probes / restart policy. Never depends on Postgres, MinIO, or Qdrant. |
+| `GET /readyz` | Can it actually serve? | Load-balancer admission and post-deploy gating. `200` with each dependency `true`, `503` if any is `false`. |
+
+`/readyz` checks Postgres, MinIO, and Qdrant reachability **and** that the
+`chunks` collection exists. A reachable-but-unbootstrapped Qdrant reports
+`{"qdrant": false}` on purpose: uploads to it would fail with `502`, so it is
+not ready, and this is how a skipped `qdrant_bootstrap` surfaces at deploy time
+instead of on a user's first upload.
+
+## 9. Investigating a failure
+
+Logs are structured (`structlog`, JSON in production) with stable event names —
+grep the event name, not prose (ADR-0011). Set `LOG_FILE_PATH` for a local
+JSON file sink alongside the console renderer; leave it unset in production,
+where stdout collection is preferred.
+
+**Correlate by `request_id`.** Every request has one, echoed in the
+`X-Request-Id` response header and included in every error envelope, and bound
+into every log line emitted while handling that request. A client reporting a
+failed upload should quote it. `tenant_id`, `file_id`, and `ingestion_job_id`
+are the other join keys.
+
+Events worth knowing:
+
+| Event | Level | Means |
+|---|---|---|
+| `ingestion.job.started` | info | Txn A committed; work phase beginning. Carries `tenant_id`, `ingestion_job_id`, `file_id`, `domain`, `source_type`. |
+| `ingestion.job.completed` | info | Terminal success, with `chunks_parsed`, `points_upserted`, `points_soft_deleted`. |
+| `ingestion.job.failed` | warning | Terminal failure. **`error_code` says which stage**: `storage_upload_failed`, `parse_failed`, `chunk_limit_exceeded`, `embedding_failed`, `index_failed`, `timeout`. |
+| `files.upload.duplicate` | info | Identical content already ingested; the existing file/job was returned and nothing was re-ingested. |
+| `domain.rejected` | warning | Upload refused before any row was written; `reason` is `unregistered` or `disabled`. |
+| `auth.failed` | warning | `reason` is `malformed_key`, `unknown_key`, `key_inactive`, `key_expired`, or `tenant_inactive`. Never contains key material. |
+| `auth.succeeded` | info | Carries `tenant_id`, `api_key_id`, `actor_type`. |
+| `lifespan.embedder.warm_failed` | warning | An embedder was unreachable at startup. Boot continues by design — `/readyz` and the first upload are where this bites. |
+| `qdrant.bootstrap.schema_mismatch` | error | The existing collection diverges from the pinned schema. The bootstrap exits non-zero; do not start the app against it. |
+| `api.unhandled_exception` | error | A bug: an exception with no mapping to the error envelope. Always worth a look. |
+
+A `503` (`ingestion_at_capacity`) is rejected before a job row exists, so it
+appears in the access log and metrics, not in `ingestion_jobs`.
+
+### The durable record
+
+Logs may roll; `ingestion_jobs` and `ingestion_job_events` do not. For one file:
+
+```sql
+SELECT id, status, error_code, error_message, points_created, points_soft_deleted,
+       created_at, updated_at
+FROM ingestion_jobs
+WHERE tenant_id = :tenant_id AND source_file_id = :file_id
+ORDER BY created_at DESC;
+
+SELECT stage, level, message, details, created_at
+FROM ingestion_job_events
+WHERE tenant_id = :tenant_id AND ingestion_job_id = :ingestion_job_id
+ORDER BY created_at;
+```
+
+Recent failures across a tenant:
+
+```sql
+SELECT error_code, count(*), max(created_at)
+FROM ingestion_jobs
+WHERE tenant_id = :tenant_id AND status = 'failed' AND created_at > now() - interval '1 day'
+GROUP BY error_code ORDER BY 2 DESC;
+```
+
+`GET /v1/files/{file_id}` reports the same terminal status over HTTP, scoped to
+the owning tenant — a file belonging to another tenant returns `404`, not `403`.
+
+## 10. Retrying a failed ingestion
+
+**Re-upload the same bytes.** There is no retry endpoint and no automatic retry;
+the client owns that decision (ADR-0017).
+
+What that guarantees:
+
+- Identical content with a **succeeded** job is recognized by
+  `(tenant_id, domain, content_sha256)` and returned as-is with `200` — no
+  re-ingestion, no duplicate points.
+- Identical content whose last job **failed** starts a fresh job against the
+  same `source_files` row. A terminal job is never moved back to `running`.
+- Point ids are deterministic from `file_id` + `chunk_index` (ADR-0001), so the
+  retry **overwrites in place** — it cannot duplicate chunks.
+- A failed attempt never empties or partially deletes a working index: the
+  soft-delete sweep that retires a shortened file's leftover points runs only
+  after every upsert has succeeded. An interrupted attempt can leave a prefix
+  updated; a retry converges (ADR-0017, "Re-running an ingestion stays safe").
+
+Fix the cause first — the `error_code` says where to look:
+
+| `error_code` | Usual cause |
+|---|---|
+| `parse_failed` | The file is corrupt or is not really the type its extension claims. Retrying identical bytes will fail identically. |
+| `chunk_limit_exceeded` | The file is genuinely too large for one inline ingestion. Split it, or raise `INGESTION_MAX_CHUNKS_PER_FILE` knowing what §6 says about the timeout. |
+| `embedding_failed` | The embedder is down, rate-limiting, or unauthenticated. Fix it, then retry — this one usually succeeds unchanged. |
+| `index_failed` | Qdrant is down, or the collection is missing (run `qdrant_bootstrap`). |
+| `timeout` | The work exceeded `INGESTION_TIMEOUT_SECONDS`. Check whether the embedder was cold (see `lifespan.embedder.warm_failed` and `KEEP_ALIVE`) before raising the bound. |
+| `storage_upload_failed` | MinIO is unreachable or the bucket is missing (§3). |
+
+## 11. What to alert on
+
+ADR-0017's own triggers for moving ingestion back off the request path. These
+are the numbers that say the inline design has stopped fitting:
+
+- **p95 ingestion duration** approaching `INGESTION_TIMEOUT_SECONDS`.
+- **`503` and `504` rates** ceasing to be negligible.
+- **Jobs stuck in `running` past the timeout** — every handled failure writes a
+  terminal status, so a non-zero count here means the process died mid-request:
+
+  ```sql
+  SELECT count(*) FROM ingestion_jobs
+  WHERE status = 'running' AND created_at < now() - interval '5 minutes';
+  ```
+
+Also worth alerting: any `qdrant.bootstrap.schema_mismatch`, a sustained
+`/readyz` `503`, and any `api.unhandled_exception`.
+
+## 12. Verifying a deployment
+
+```bash
+./scripts/smoke.sh
+```
+
+Brings up Compose, runs both deployment steps, provisions a throwaway tenant,
+starts the web process, and drives an upload through to indexed Qdrant points
+against the **running process** — including asserting the structured log output
+from §9. It is the only Compose-based test; everything else runs on
+Testcontainers under `uv run pytest` (ADR-0016). Run it before a release.

pyproject.toml

@@ -9,6 +9,7 @@ dependencies = [
     "anyio>=4.11.0",
     "asyncpg>=0.31.0",
     "fastapi[standard]==0.141.1",
+    "httpx>=0.28.1",
     "langgraph>=1.2.10",
     "minio>=7.2.20",
     "openpyxl>=3.1.5",
@@ -23,7 +24,6 @@ dependencies = [
 [dependency-groups]
 dev = [
     "asgi-lifespan>=2.1.0",
-    "httpx>=0.28.1",
     "pytest>=8.3.5",
     "pytest-asyncio>=0.25.3",
     "pytest-cov>=6.0.0",
@@ -37,6 +37,11 @@ dev = [
 testpaths = ["tests"]
 asyncio_mode = "strict"
 timeout = 10
+# Bound the test function only, not fixture setup. Testcontainers' container
+# startup is charged to whichever test first pulls a session-scoped container
+# fixture; on a cold Docker cache that is ~25s and would trip the 10s budget
+# for every integration test, regardless of how fast the test itself is.
+timeout_func_only = true
 markers = [
     "unit: fast tests with no external services",
     "integration: tests against a real disposable service",

scripts/smoke.sh

@@ -0,0 +1,98 @@
+#!/usr/bin/env bash
+# Serialized operational smoke test of the running web process (ADR-0016, plan
+# 001 Phase 6).
+#
+#   ./scripts/smoke.sh
+#
+# Brings up the Compose stack, runs both deployment steps for real, provisions
+# a tenant, starts uvicorn, and drives `tests/e2e/test_compose_smoke.py`
+# against it over a socket. This is the only Compose-based test: every other
+# test uses Testcontainers and an in-process ASGI transport, which is exactly
+# what makes this one worth having -- it is the only thing that exercises the
+# deployment steps, the real logging configuration, and a real HTTP server.
+#
+# Not part of `uv run pytest`: the smoke test skips itself unless SMOKE_BASE_URL
+# is set, so this script is the only way it runs. Run it before a release.
+#
+# Leaves the Compose stack running (it is the local dev stack); only the uvicorn
+# process and the temporary log file are cleaned up.
+set -euo pipefail
+
+cd "$(dirname "${BASH_SOURCE[0]}")/.."
+
+PORT="${SMOKE_PORT:-8021}"
+SLUG="smoke-$(date +%s)"
+DOMAIN="smoke"
+LOG_FILE="$(mktemp -t smoke-app-log.XXXXXX.jsonl)"
+CONSOLE_LOG="$(mktemp -t smoke-app-console.XXXXXX.log)"
+APP_PID=""
+
+cleanup() {
+  if [[ -n "${APP_PID}" ]] && kill -0 "${APP_PID}" 2>/dev/null; then
+    kill "${APP_PID}" 2>/dev/null || true
+    wait "${APP_PID}" 2>/dev/null || true
+  fi
+  rm -f "${LOG_FILE}" "${CONSOLE_LOG}"
+}
+trap cleanup EXIT
+
+if [[ ! -f .env ]]; then
+  echo "no .env found; copy .env.example first (see docs/runbook.md)" >&2
+  exit 1
+fi
+
+echo "==> starting Postgres, MinIO, Qdrant"
+docker compose up -d --wait
+
+echo "==> applying deployment steps"
+uv run alembic upgrade head
+uv run python -m src.cli.qdrant_bootstrap
+
+echo "==> provisioning tenant '${SLUG}'"
+PROVISION_OUTPUT="$(uv run python -m src.cli.provision_tenant \
+  --slug "${SLUG}" --domain "${DOMAIN}" --scopes files:write 2>/dev/null)"
+API_KEY="$(printf '%s\n' "${PROVISION_OUTPUT}" | sed -n 's/^api_key=//p')"
+if [[ -z "${API_KEY}" ]]; then
+  echo "provisioning did not return an api_key" >&2
+  exit 1
+fi
+
+echo "==> starting the web process on port ${PORT}"
+# JSON to a file sink, because the smoke test asserts the real ADR-0011 log
+# output -- the one thing no in-process test can check.
+LOG_JSON_FORMAT=true LOG_FILE_PATH="${LOG_FILE}" \
+  uv run python -m uvicorn src.main:app --host 127.0.0.1 --port "${PORT}" \
+  >"${CONSOLE_LOG}" 2>&1 &
+APP_PID=$!
+
+echo "==> waiting for /readyz"
+for _ in $(seq 1 60); do
+  if curl -fsS "http://127.0.0.1:${PORT}/readyz" >/dev/null 2>&1; then
+    break
+  fi
+  if ! kill -0 "${APP_PID}" 2>/dev/null; then
+    echo "the web process exited before becoming ready:" >&2
+    tail -20 "${CONSOLE_LOG}" >&2
+    exit 1
+  fi
+  sleep 1
+done
+
+if ! curl -fsS "http://127.0.0.1:${PORT}/readyz" >/dev/null 2>&1; then
+  # Most often an unbootstrapped Qdrant or an unreachable embedder host; the
+  # runbook's health/readiness section covers reading this.
+  echo "the web process never became ready:" >&2
+  tail -20 "${CONSOLE_LOG}" >&2
+  exit 1
+fi
+
+echo "==> running the smoke test"
+SMOKE_BASE_URL="http://127.0.0.1:${PORT}" \
+SMOKE_API_KEY="${API_KEY}" \
+SMOKE_DOMAIN="${DOMAIN}" \
+SMOKE_LOG_PATH="${LOG_FILE}" \
+SMOKE_QDRANT_URL="${QDRANT_URL:-http://127.0.0.1:6343}" \
+SMOKE_QDRANT_COLLECTION="${QDRANT_COLLECTION:-chunks}" \
+  uv run python -m pytest tests/e2e/test_compose_smoke.py -q
+
+echo "==> smoke test passed"

src/api/dependencies/auth.py

@@ -0,0 +1,38 @@
+"""Auth dependencies (ADR-0008): resolve `AuthContext` from a bearer token,
+then gate routes on scope.
+"""
+
+from collections.abc import Awaitable, Callable
+from typing import Annotated
+
+from fastapi import Depends
+from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.application.auth.context import AuthContext
+from src.application.auth.errors import InvalidApiKeyError, MissingScopeError
+from src.application.auth.service import resolve_auth_context
+from src.bootstrap.dependencies import get_sessionmaker
+
+_bearer_scheme = HTTPBearer(auto_error=False)
+
+
+async def get_auth_context(
+    credentials: Annotated[HTTPAuthorizationCredentials | None, Depends(_bearer_scheme)],
+    sessionmaker: Annotated[async_sessionmaker[AsyncSession], Depends(get_sessionmaker)],
+) -> AuthContext:
+    if credentials is None:
+        raise InvalidApiKeyError("missing Authorization header")
+    return await resolve_auth_context(sessionmaker, credentials.credentials)
+
+
+AuthContextDep = Annotated[AuthContext, Depends(get_auth_context)]
+
+
+def require_scope(scope: str) -> Callable[[AuthContext], Awaitable[AuthContext]]:
+    async def _dependency(auth: AuthContextDep) -> AuthContext:
+        if not auth.has_scope(scope):
+            raise MissingScopeError(f"missing required scope '{scope}'")
+        return auth
+
+    return _dependency

src/api/errors.py

@@ -0,0 +1,138 @@
+"""Maps application exceptions to the ADR-0008 error envelope.
+
+This is the single place that knows the exception-type -> status-code
+mapping; application/infrastructure code never imports FastAPI or raises
+`HTTPException` (ADR-0015).
+"""
+
+import structlog
+from fastapi import FastAPI, Request, status
+from fastapi.exceptions import RequestValidationError
+from fastapi.responses import JSONResponse
+from starlette.exceptions import HTTPException as StarletteHTTPException
+
+from src.application.auth.errors import (
+    InvalidApiKeyError,
+    MissingScopeError,
+    TenantInactiveError,
+)
+from src.application.domains.errors import DomainAlreadyExistsError, UnknownDomainError
+from src.application.files.errors import (
+    FileTooLargeError,
+    InvalidUploadError,
+    SourceFileNotFoundError,
+)
+from src.application.ingestion.errors import (
+    ChunkLimitExceededError,
+    DocumentParseError,
+    EmbedderError,
+    IngestionAtCapacityError,
+    IngestionTimeoutError,
+    PointIndexingError,
+    UnsupportedSourceTypeError,
+)
+from src.application.points.errors import PointVersionConflictError
+from src.application.points.point import PointNotFoundError
+
+logger = structlog.get_logger(__name__)
+
+# A fixed backoff hint, not a computed retry budget: ADR-0017 rejects a
+# request outright at capacity rather than queueing it, so there is no
+# in-process estimate of when a slot will free up to report instead.
+_CAPACITY_RETRY_AFTER_SECONDS = 1
+
+# (exception type, status code, stable error code)
+_MAPPING: tuple[tuple[type[Exception], int, str], ...] = (
+    (InvalidApiKeyError, status.HTTP_401_UNAUTHORIZED, "invalid_api_key"),
+    (TenantInactiveError, status.HTTP_401_UNAUTHORIZED, "tenant_not_found"),
+    (MissingScopeError, status.HTTP_403_FORBIDDEN, "missing_scope"),
+    (InvalidUploadError, status.HTTP_400_BAD_REQUEST, "validation_error"),
+    # 404, never 403: a cross-tenant point id must be indistinguishable from a
+    # nonexistent one, or the API becomes an existence oracle (ADR-0016).
+    (PointNotFoundError, status.HTTP_404_NOT_FOUND, "not_found"),
+    (SourceFileNotFoundError, status.HTTP_404_NOT_FOUND, "not_found"),
+    # Not "the version guard fired once" — that is retried. This is the service
+    # giving up after repeated re-plans, i.e. a genuinely contended point.
+    (PointVersionConflictError, status.HTTP_409_CONFLICT, "conflict"),
+    (UnknownDomainError, status.HTTP_400_BAD_REQUEST, "unknown_domain"),
+    (DomainAlreadyExistsError, status.HTTP_409_CONFLICT, "conflict"),
+    (DocumentParseError, status.HTTP_400_BAD_REQUEST, "validation_error"),
+    (UnsupportedSourceTypeError, status.HTTP_415_UNSUPPORTED_MEDIA_TYPE, "unsupported_media_type"),
+    (FileTooLargeError, status.HTTP_413_CONTENT_TOO_LARGE, "payload_too_large"),
+    (ChunkLimitExceededError, status.HTTP_413_CONTENT_TOO_LARGE, "payload_too_large"),
+    (EmbedderError, status.HTTP_502_BAD_GATEWAY, "embedder_error"),
+    (PointIndexingError, status.HTTP_502_BAD_GATEWAY, "index_error"),
+    (IngestionTimeoutError, status.HTTP_504_GATEWAY_TIMEOUT, "ingestion_timeout"),
+)
+
+
+def _request_id(request: Request) -> str | None:
+    return getattr(request.state, "request_id", None)
+
+
+def _envelope(
+    code: str, message: str, request_id: str | None, details: dict[str, object] | None = None
+) -> dict[str, object]:
+    return {
+        "error": {
+            "code": code,
+            "message": message,
+            "details": details or {},
+            "request_id": request_id,
+        }
+    }
+
+
+def register_exception_handlers(app: FastAPI) -> None:
+    for exc_type, status_code, error_code in _MAPPING:
+
+        def _handler(
+            request: Request,
+            exc: Exception,
+            status_code: int = status_code,
+            error_code: str = error_code,
+        ) -> JSONResponse:
+            return JSONResponse(
+                status_code=status_code,
+                content=_envelope(error_code, str(exc), _request_id(request)),
+            )
+
+        app.add_exception_handler(exc_type, _handler)
+
+    @app.exception_handler(IngestionAtCapacityError)
+    def _capacity_handler(request: Request, exc: IngestionAtCapacityError) -> JSONResponse:
+        return JSONResponse(
+            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            content=_envelope("ingestion_at_capacity", str(exc), _request_id(request)),
+            headers={"Retry-After": str(_CAPACITY_RETRY_AFTER_SECONDS)},
+        )
+
+    @app.exception_handler(RequestValidationError)
+    def _validation_handler(request: Request, exc: RequestValidationError) -> JSONResponse:
+        return JSONResponse(
+            status_code=status.HTTP_422_UNPROCESSABLE_CONTENT,
+            content=_envelope(
+                "validation_error",
+                "request validation failed",
+                _request_id(request),
+                details={"errors": exc.errors()},
+            ),
+        )
+
+    @app.exception_handler(StarletteHTTPException)
+    def _http_exception_handler(request: Request, exc: StarletteHTTPException) -> JSONResponse:
+        code = "not_found" if exc.status_code == status.HTTP_404_NOT_FOUND else "http_error"
+        return JSONResponse(
+            status_code=exc.status_code,
+            content=_envelope(code, str(exc.detail), _request_id(request)),
+        )
+
+    @app.exception_handler(Exception)
+    def _unhandled_exception_handler(request: Request, exc: Exception) -> JSONResponse:
+        logger.exception("api.unhandled_exception", path=request.url.path)
+        return JSONResponse(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            content=_envelope(
+                "internal_error", "an unexpected error occurred", _request_id(request)
+            ),
+        )

src/api/middleware.py

@@ -0,0 +1,38 @@
+"""Per-request correlation id (ADR-0008, ADR-0011).
+
+Every request gets a `request_id`: reused from an incoming `X-Request-Id` if
+the caller supplied one, otherwise generated. It is bound into structlog's
+contextvars so every log line emitted while handling the request carries it,
+stored on `request.state` for exception handlers, and echoed back in the
+response header.
+"""
+
+import uuid
+from collections.abc import Awaitable, Callable
+from typing import override
+
+import structlog
+from starlette.middleware.base import BaseHTTPMiddleware
+from starlette.requests import Request
+from starlette.responses import Response
+
+_HEADER = "X-Request-Id"
+
+
+class RequestIdMiddleware(BaseHTTPMiddleware):
+    @override
+    async def dispatch(
+        self, request: Request, call_next: Callable[[Request], Awaitable[Response]]
+    ) -> Response:
+        request_id = request.headers.get(_HEADER) or str(uuid.uuid4())
+        request.state.request_id = request_id
+
+        structlog.contextvars.clear_contextvars()
+        structlog.contextvars.bind_contextvars(request_id=request_id)
+        try:
+            response = await call_next(request)
+        finally:
+            structlog.contextvars.clear_contextvars()
+
+        response.headers[_HEADER] = request_id
+        return response

src/api/router.py

@@ -1,3 +1,10 @@
 from fastapi import APIRouter
 
+from src.api.routers.domains import router as domains_router
+from src.api.routers.files import router as files_router
+from src.api.routers.points import router as points_router
+
 router = APIRouter()
+router.include_router(domains_router)
+router.include_router(files_router)
+router.include_router(points_router)

src/api/routers/domains.py

@@ -0,0 +1,110 @@
+"""`/v1/domains` (ADR-0008, ADR-0009).
+
+The management surface for a tenant's domain allowlist, used by the calling
+backend rather than by an operator with a psql prompt.
+
+Gated on `domains:read`/`domains:write`, deliberately **not** on `files:write`:
+if an upload key could create domains, the allowlist would no longer prevent a
+typo'd `domain` from creating a Qdrant partition, which is its only purpose.
+"""
+
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, status
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.api.dependencies.auth import require_scope
+from src.api.schemas.domains import (
+    CreateDomainRequest,
+    DomainListResponse,
+    DomainResponse,
+    UpdateDomainRequest,
+)
+from src.application.auth.context import AuthContext
+from src.application.domains import (
+    create_domain,
+    list_domains,
+    set_domain_status,
+    update_domain,
+)
+from src.bootstrap.dependencies import get_sessionmaker
+
+router = APIRouter(prefix="/domains", tags=["domains"])
+
+_RequireDomainsRead = Annotated[AuthContext, Depends(require_scope("domains:read"))]
+_RequireDomainsWrite = Annotated[AuthContext, Depends(require_scope("domains:write"))]
+_SessionmakerDep = Annotated[async_sessionmaker[AsyncSession], Depends(get_sessionmaker)]
+
+
+@router.get("")
+async def list_tenant_domains(
+    auth: _RequireDomainsRead,
+    sessionmaker: _SessionmakerDep,
+    include_disabled: bool = False,
+) -> DomainListResponse:
+    results = await list_domains(
+        sessionmaker, tenant_id=auth.tenant_id, include_disabled=include_disabled
+    )
+    return DomainListResponse(domains=[DomainResponse.from_result(item) for item in results])
+
+
+@router.post("", status_code=status.HTTP_201_CREATED)
+async def create_tenant_domain(
+    request: CreateDomainRequest,
+    auth: _RequireDomainsWrite,
+    sessionmaker: _SessionmakerDep,
+) -> DomainResponse:
+    result = await create_domain(
+        sessionmaker,
+        tenant_id=auth.tenant_id,
+        domain=request.domain,
+        display_name=request.display_name,
+        metadata=request.metadata,
+    )
+    return DomainResponse.from_result(result)
+
+
+@router.patch("/{domain}")
+async def update_tenant_domain(
+    domain: str,
+    request: UpdateDomainRequest,
+    auth: _RequireDomainsWrite,
+    sessionmaker: _SessionmakerDep,
+) -> DomainResponse:
+    result = await update_domain(
+        sessionmaker,
+        tenant_id=auth.tenant_id,
+        domain=domain,
+        display_name=request.display_name,
+    )
+    return DomainResponse.from_result(result)
+
+
+@router.delete("/{domain}")
+async def disable_tenant_domain(
+    domain: str,
+    auth: _RequireDomainsWrite,
+    sessionmaker: _SessionmakerDep,
+) -> DomainResponse:
+    """Disable, not delete.
+
+    Blocks new uploads and drops the domain from pickers while leaving the
+    points already indexed under it intact and retrievable. Actually removing
+    them needs the tenant-erasure workflow plan 001 defers.
+    """
+    result = await set_domain_status(
+        sessionmaker, tenant_id=auth.tenant_id, domain=domain, status="disabled"
+    )
+    return DomainResponse.from_result(result)
+
+
+@router.post("/{domain}/enable")
+async def enable_tenant_domain(
+    domain: str,
+    auth: _RequireDomainsWrite,
+    sessionmaker: _SessionmakerDep,
+) -> DomainResponse:
+    result = await set_domain_status(
+        sessionmaker, tenant_id=auth.tenant_id, domain=domain, status="active"
+    )
+    return DomainResponse.from_result(result)

src/api/routers/files.py

@@ -0,0 +1,157 @@
+"""`POST /v1/files`, `GET /v1/files/{file_id}`, `DELETE /v1/files/{file_id}` (ADR-0008).
+
+Routes adapt HTTP to `application/files` calls; they do not parse, hash,
+touch MinIO/Qdrant, or otherwise carry ingestion business logic (ADR-0015).
+"""
+
+import uuid
+from collections.abc import Sequence
+from typing import Annotated
+
+from anyio import CapacityLimiter, Semaphore
+from fastapi import APIRouter, Depends, Form, HTTPException, Response, UploadFile, status
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.api.dependencies.auth import require_scope
+from src.api.schemas.files import FileDeleteResponse, FileStatusResponse, FileUploadResponse
+from src.api.schemas.points import DEFAULT_PAGE_SIZE, LimitQuery, PointListResponse
+from src.application.auth.context import AuthContext
+from src.application.files.deletion import delete_source_file
+from src.application.files.status import get_file_status
+from src.application.files.upload import upload_source_file
+from src.application.points.queries import list_file_points
+from src.application.ports.embedding import DenseEmbedder, SparseEmbedder
+from src.application.ports.object_storage import ObjectStorage
+from src.application.ports.point_repository import PointRepository
+from src.application.ports.point_storage import PointStorage
+from src.bootstrap.dependencies import (
+    get_dense_embedders,
+    get_ingestion_concurrency_limiter,
+    get_ingestion_limiter,
+    get_object_storage,
+    get_point_repository,
+    get_point_storage,
+    get_sessionmaker,
+    get_settings,
+    get_sparse_embedder,
+)
+from src.config import Settings
+
+router = APIRouter(prefix="/files", tags=["files"])
+
+_RequireFilesWrite = Annotated[AuthContext, Depends(require_scope("files:write"))]
+_RequirePointsRead = Annotated[AuthContext, Depends(require_scope("points:read"))]
+_RequirePointsWrite = Annotated[AuthContext, Depends(require_scope("points:write"))]
+_SessionmakerDep = Annotated[async_sessionmaker[AsyncSession], Depends(get_sessionmaker)]
+_ObjectStorageDep = Annotated[ObjectStorage, Depends(get_object_storage)]
+_PointStorageDep = Annotated[PointStorage, Depends(get_point_storage)]
+_PointRepositoryDep = Annotated[PointRepository, Depends(get_point_repository)]
+_SettingsDep = Annotated[Settings, Depends(get_settings)]
+_IngestionLimiterDep = Annotated[CapacityLimiter, Depends(get_ingestion_limiter)]
+_ConcurrencyLimiterDep = Annotated[Semaphore, Depends(get_ingestion_concurrency_limiter)]
+_DenseEmbeddersDep = Annotated[Sequence[DenseEmbedder], Depends(get_dense_embedders)]
+_SparseEmbedderDep = Annotated[SparseEmbedder, Depends(get_sparse_embedder)]
+
+
+@router.post("", status_code=status.HTTP_201_CREATED)
+async def upload_file(
+    response: Response,
+    file: UploadFile,
+    domain: Annotated[str, Form()],
+    auth: _RequireFilesWrite,
+    sessionmaker: _SessionmakerDep,
+    storage: _ObjectStorageDep,
+    point_storage: _PointStorageDep,
+    settings: _SettingsDep,
+    limiter: _IngestionLimiterDep,
+    concurrency_limiter: _ConcurrencyLimiterDep,
+    dense_embedders: _DenseEmbeddersDep,
+    sparse_embedder: _SparseEmbedderDep,
+) -> FileUploadResponse:
+    data = await file.read()
+    result = await upload_source_file(
+        sessionmaker=sessionmaker,
+        storage=storage,
+        point_storage=point_storage,
+        auth=auth,
+        domain=domain,
+        filename=file.filename or "",
+        data=data,
+        ingestion_settings=settings.ingestion,
+        chunking_settings=settings.chunking,
+        qdrant_settings=settings.qdrant,
+        thread_limiter=limiter,
+        concurrency_limiter=concurrency_limiter,
+        dense_embedders=dense_embedders,
+        sparse_embedder=sparse_embedder,
+    )
+    if not result.is_new_attempt:
+        response.status_code = status.HTTP_200_OK
+    return FileUploadResponse.from_result(result)
+
+
+@router.get("/{file_id}")
+async def get_file(
+    file_id: uuid.UUID,
+    auth: _RequireFilesWrite,
+    sessionmaker: _SessionmakerDep,
+) -> FileStatusResponse:
+    result = await get_file_status(sessionmaker, tenant_id=auth.tenant_id, source_file_id=file_id)
+    if result is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="file not found")
+    return FileStatusResponse.from_result(result)
+
+
+@router.delete("/{file_id}")
+async def delete_file(
+    file_id: uuid.UUID,
+    auth: _RequirePointsWrite,
+    sessionmaker: _SessionmakerDep,
+    repository: _PointRepositoryDep,
+) -> FileDeleteResponse:
+    """Soft-delete a file: every active point, then the `source_files` row.
+
+    Gated on `points:write` rather than `files:write` for the same reason as the
+    listing above — the data this destroys is points. Nothing is removed from
+    Qdrant (ADR-0002); the points are flagged inactive and the row is marked
+    `soft_deleted`, which is also what makes a later re-upload of the same bytes
+    ingest afresh instead of matching the duplicate path.
+
+    Deleting an already-deleted file is a success reporting `0` points.
+    """
+    points_soft_deleted = await delete_source_file(
+        sessionmaker,
+        repository,
+        tenant_id=auth.tenant_id,
+        source_file_id=file_id,
+        actor=f"api_key:{auth.api_key_id}",
+    )
+    return FileDeleteResponse(
+        file_id=file_id, status="soft_deleted", points_soft_deleted=points_soft_deleted
+    )
+
+
+@router.get("/{file_id}/points")
+async def list_points_for_file(
+    file_id: uuid.UUID,
+    auth: _RequirePointsRead,
+    repository: _PointRepositoryDep,
+    limit: LimitQuery = DEFAULT_PAGE_SIZE,
+    cursor: str | None = None,
+    include_inactive: bool = False,
+) -> PointListResponse:
+    """The same listing as `GET /v1/points?file_id=...`, addressed by file.
+
+    Gated on `points:read`, not `files:write`: the resource being read is the
+    file's chunks, so the scope follows the data rather than the URL prefix. An
+    upload-only key must not become a way to read every chunk of every file.
+    """
+    page = await list_file_points(
+        repository,
+        tenant_id=auth.tenant_id,
+        file_id=file_id,
+        limit=limit,
+        cursor=cursor,
+        include_inactive=include_inactive,
+    )
+    return PointListResponse.from_page(page)

src/api/routers/health.py

@@ -23,7 +23,9 @@ async def readyz(request: Request, response: Response) -> dict[str, bool]:
     postgres_ready, minio_ready, qdrant_ready = await asyncio.gather(
         ping_postgres(resources.db_engine, timeout),
         ping_minio(resources.minio_client, timeout),
-        ping_qdrant(resources.qdrant_client, timeout),
+        ping_qdrant(
+            resources.qdrant_client, timeout, collection=resources.settings.qdrant.collection
+        ),
     )
 
     result = {

src/api/routers/points.py

@@ -0,0 +1,162 @@
+"""`/v1/points` read and soft-delete paths (ADR-0002, ADR-0008).
+
+Routes adapt HTTP to `application/points` calls. They build no Qdrant filters
+and hold no CRUD semantics (ADR-0015), and they never read a tenant from the
+request — `auth.tenant_id` is the only source, which is what makes ADR-0002's
+isolation rule structural rather than a habit.
+
+**Route order is load-bearing.** `/count` and `/search` are declared before
+`/{point_id}`. FastAPI matches in declaration order, so with `/{point_id}` first
+a request for `/v1/points/count` would try to parse `"count"` as a UUID and
+fail with `422` instead of counting anything. The failure is loud but confusing,
+and it comes back the moment someone reorders these for tidiness.
+
+Gated on `points:read`, separately from `files:write`: a key that can upload
+documents should not thereby be able to read every chunk of every file, and
+plan 002's mutating paths will want `points:write` distinct again.
+"""
+
+import uuid
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, Query
+
+from src.api.dependencies.auth import require_scope
+from src.api.schemas.points import (
+    DEFAULT_PAGE_SIZE,
+    LimitQuery,
+    PointCountResponse,
+    PointListResponse,
+    PointResponse,
+    PointSearchResponse,
+)
+from src.application.auth.context import AuthContext
+from src.application.points.deletion import soft_delete_point
+from src.application.points.queries import (
+    count_points,
+    get_point,
+    list_file_points,
+    search_points,
+)
+from src.application.ports.point_repository import PointRepository
+from src.bootstrap.dependencies import get_point_repository
+
+router = APIRouter(prefix="/points", tags=["points"])
+
+_RequirePointsRead = Annotated[AuthContext, Depends(require_scope("points:read"))]
+_RequirePointsWrite = Annotated[AuthContext, Depends(require_scope("points:write"))]
+_PointRepositoryDep = Annotated[PointRepository, Depends(get_point_repository)]
+
+
+@router.get("/count")
+async def count_tenant_points(
+    auth: _RequirePointsRead,
+    repository: _PointRepositoryDep,
+    domain: str | None = None,
+    file_id: uuid.UUID | None = None,
+    include_inactive: bool = False,
+) -> PointCountResponse:
+    count = await count_points(
+        repository,
+        tenant_id=auth.tenant_id,
+        domain=domain,
+        file_id=file_id,
+        include_inactive=include_inactive,
+    )
+    return PointCountResponse(count=count)
+
+
+@router.get("/search")
+async def search_tenant_points(
+    auth: _RequirePointsRead,
+    repository: _PointRepositoryDep,
+    q: Annotated[str, Query(min_length=1)],
+    limit: LimitQuery = DEFAULT_PAGE_SIZE,
+    cursor: str | None = None,
+    domain: str | None = None,
+    file_id: uuid.UUID | None = None,
+    include_inactive: bool = False,
+) -> PointSearchResponse:
+    """Keyword search over point content — **not** semantic retrieval.
+
+    Matches Qdrant's full-text payload index on `content`, combined with the
+    structured filters below. Results are unranked: the index filters rather
+    than scores, so there is no relevance order and no score to return. Callers
+    wanting ranked answers want the agent retrieval path (plan 003), not this.
+    """
+    page = await search_points(
+        repository,
+        tenant_id=auth.tenant_id,
+        query=q,
+        limit=limit,
+        cursor=cursor,
+        domain=domain,
+        file_id=file_id,
+        include_inactive=include_inactive,
+    )
+    return PointSearchResponse.from_search(page, query=q)
+
+
+@router.get("")
+async def list_tenant_points(
+    auth: _RequirePointsRead,
+    repository: _PointRepositoryDep,
+    file_id: uuid.UUID,
+    limit: LimitQuery = DEFAULT_PAGE_SIZE,
+    cursor: str | None = None,
+    include_inactive: bool = False,
+) -> PointListResponse:
+    """A file's points in `order_id` order.
+
+    `file_id` is required rather than optional: the pagination cursor is an
+    `order_id` value, and `order_id` is only unique within one file. Listing
+    across files would silently drop or repeat rows at every page boundary.
+    """
+    page = await list_file_points(
+        repository,
+        tenant_id=auth.tenant_id,
+        file_id=file_id,
+        limit=limit,
+        cursor=cursor,
+        include_inactive=include_inactive,
+    )
+    return PointListResponse.from_page(page)
+
+
+@router.get("/{point_id}")
+async def get_tenant_point(
+    point_id: uuid.UUID,
+    auth: _RequirePointsRead,
+    repository: _PointRepositoryDep,
+    with_vectors: bool = False,
+) -> PointResponse:
+    point = await get_point(
+        repository, tenant_id=auth.tenant_id, point_id=point_id, with_vectors=with_vectors
+    )
+    return PointResponse.from_point(point)
+
+
+@router.delete("/{point_id}")
+async def delete_tenant_point(
+    point_id: uuid.UUID,
+    auth: _RequirePointsWrite,
+    repository: _PointRepositoryDep,
+) -> PointResponse:
+    """Soft-delete one point and relink its neighbours around the gap.
+
+    The point is never removed from Qdrant (ADR-0002): it is flagged
+    `is_active=false` with `deleted_at` set, and its old neighbours are pointed
+    at each other in the same batch, so context-window expansion never walks
+    into it.
+
+    Deleting an already-inactive point is a no-op success rather than a `404` —
+    the response is the point as it stands, so the resulting `version` and
+    `deleted_at` are visible either way.
+    """
+    point = await soft_delete_point(
+        repository,
+        tenant_id=auth.tenant_id,
+        point_id=point_id,
+        actor=f"api_key:{auth.api_key_id}",
+    )
+    return PointResponse.from_point(point)

src/api/schemas/domains.py

@@ -0,0 +1,63 @@
+"""Public request/response models for `/v1/domains` (ADR-0008, ADR-0009).
+
+`tenant_id` appears in none of these: it comes from the authenticated key, and
+accepting it from a body would break the isolation boundary (ADR-0002).
+"""
+
+import uuid
+from datetime import datetime
+
+from pydantic import BaseModel, Field, field_validator
+
+from src.application.domains.models import DomainResult
+
+# Lowercase alphanumerics plus - and _; the key is embedded in every Qdrant
+# payload and filtered on as a keyword, so it stays boring on purpose.
+_DOMAIN_PATTERN = r"^[a-z0-9][a-z0-9_-]*$"
+
+
+class DomainResponse(BaseModel):
+    id: uuid.UUID
+    domain: str
+    display_name: str
+    status: str
+    metadata: dict[str, object]
+    created_at: datetime
+    updated_at: datetime
+
+    @classmethod
+    def from_result(cls, result: DomainResult) -> "DomainResponse":
+        return cls(
+            id=result.id,
+            domain=result.domain,
+            display_name=result.display_name,
+            status=result.status,
+            metadata=result.metadata,
+            created_at=result.created_at,
+            updated_at=result.updated_at,
+        )
+
+
+class DomainListResponse(BaseModel):
+    domains: list[DomainResponse]
+
+
+class CreateDomainRequest(BaseModel):
+    domain: str = Field(min_length=1, max_length=80, pattern=_DOMAIN_PATTERN)
+    display_name: str = Field(min_length=1, max_length=200)
+    metadata: dict[str, object] = Field(default_factory=dict)
+
+    @field_validator("domain")
+    @classmethod
+    def _normalize(cls, value: str) -> str:
+        return value.strip()
+
+
+class UpdateDomainRequest(BaseModel):
+    """`domain` is absent by design — the key is immutable.
+
+    It is denormalized into every point payload and into `source_files`, so
+    renaming it is a migration rather than an edit (ADR-0009).
+    """
+
+    display_name: str = Field(min_length=1, max_length=200)

src/api/schemas/errors.py

@@ -0,0 +1,16 @@
+"""The ADR-0008 error envelope."""
+
+from typing import Any
+
+from pydantic import BaseModel
+
+
+class ErrorDetail(BaseModel):
+    code: str
+    message: str
+    details: dict[str, Any] = {}
+    request_id: str | None = None
+
+
+class ErrorResponse(BaseModel):
+    error: ErrorDetail

src/api/schemas/files.py

@@ -0,0 +1,64 @@
+"""Public request/response models for `/v1/files` (ADR-0008).
+
+Separate from the SQLAlchemy ORM models and the `application/files` domain
+dataclasses (ADR-0015): this is the shape callers see.
+"""
+
+import uuid
+
+from pydantic import BaseModel
+
+from src.application.files.models import UploadResult
+from src.application.files.status import FileStatusResult
+
+
+class FileUploadResponse(BaseModel):
+    file_id: uuid.UUID
+    ingestion_job_id: uuid.UUID
+    status: str
+    chunks_indexed: int
+
+    @classmethod
+    def from_result(cls, result: UploadResult) -> "FileUploadResponse":
+        return cls(
+            file_id=result.file_id,
+            ingestion_job_id=result.ingestion_job_id,
+            status=result.status,
+            chunks_indexed=result.chunks_indexed,
+        )
+
+
+class FileDeleteResponse(BaseModel):
+    """What `DELETE /v1/files/{file_id}` did.
+
+    `points_soft_deleted` is reported rather than left implicit because the
+    delete is a soft one: nothing is removed from Qdrant, and the count is the
+    only way a caller can tell "deactivated 40 points" from "the file was
+    already deleted" — both of which are successes.
+    """
+
+    file_id: uuid.UUID
+    status: str
+    points_soft_deleted: int
+
+
+class FileStatusResponse(BaseModel):
+    file_id: uuid.UUID
+    source_filename: str
+    domain: str
+    status: str
+    ingestion_job_id: uuid.UUID | None
+    ingestion_status: str | None
+    chunks_indexed: int
+
+    @classmethod
+    def from_result(cls, result: FileStatusResult) -> "FileStatusResponse":
+        return cls(
+            file_id=result.file_id,
+            source_filename=result.source_filename,
+            domain=result.domain,
+            status=result.status,
+            ingestion_job_id=result.ingestion_job_id,
+            ingestion_status=result.ingestion_status,
+            chunks_indexed=result.chunks_indexed,
+        )

src/api/schemas/points.py

@@ -0,0 +1,165 @@
+"""Public request/response models for `/v1/points` (ADR-0002, ADR-0008).
+
+The shape callers see, kept separate from `application/points`' domain models
+(ADR-0015). Two rules are encoded here rather than left to route code:
+
+- **Vectors are opt-in.** `PointResponse` omits them unless the caller asked,
+  so a listing does not ship megabytes of floats nobody reads (ADR-0008).
+- **Server-owned fields are not accepted on input.** The request models simply
+  do not declare `tenant_id`, `version`, or `chunk_index`, and forbid extra
+  keys, so a client that sends one gets `422` from Pydantic instead of having
+  it silently ignored — ADR-0002's isolation rule enforced at the boundary.
+"""
+
+import uuid
+from datetime import datetime
+from typing import Annotated
+
+from fastapi import Query
+from pydantic import BaseModel, ConfigDict, Field
+
+from src.application.points.point import Point
+from src.application.ports.point_repository import PointPage
+
+# A page ceiling the caller cannot raise. Scroll pages are materialized in
+# memory both here and in Qdrant, so an unbounded `limit` is a cheap way for one
+# request to hurt every other tenant sharing the process. Declared once because
+# two routers paginate points -- `/v1/points` and `/v1/files/{file_id}/points` --
+# and a ceiling that differs between them is a ceiling in only one of them.
+DEFAULT_PAGE_SIZE = 50
+MAX_PAGE_SIZE = 200
+
+LimitQuery = Annotated[int, Query(ge=1, le=MAX_PAGE_SIZE)]
+
+
+class PointResponse(BaseModel):
+    point_id: uuid.UUID
+    domain: str
+    file_id: uuid.UUID
+    chunk_id: uuid.UUID
+
+    content: str
+    content_type: str
+    source_filename: str
+    source_type: str
+
+    order_id: float
+    chunk_index: int
+    previous_chunk_id: uuid.UUID | None
+    next_chunk_id: uuid.UUID | None
+
+    is_active: bool
+    deleted_at: datetime | None
+    created_at: datetime
+    updated_at: datetime
+    created_by: str
+    updated_by: str
+
+    version: int
+    content_hash: str
+    embedding_model_version: str
+
+    vectors: dict[str, object] | None = None
+
+    @classmethod
+    def from_point(cls, point: Point) -> "PointResponse":
+        # `tenant_id` is present on `Point` and deliberately absent here: the
+        # caller already knows which tenant it authenticated as, and echoing it
+        # back invites clients to start sending it.
+        return cls.model_validate(point.model_dump(exclude={"tenant_id"}))
+
+
+class PointListResponse(BaseModel):
+    """A page of points plus the cursor for the next one.
+
+    Cursor-based rather than `limit`/`offset`: an offset cursor silently skips
+    or repeats rows when a concurrent insert shifts positions, which is exactly
+    the pagination defect plan 002 requires a test for.
+    """
+
+    points: list[PointResponse]
+    next_cursor: str | None = None
+
+    @classmethod
+    def from_page(cls, page: PointPage) -> "PointListResponse":
+        return cls(
+            points=[PointResponse.from_point(point) for point in page.points],
+            next_cursor=page.next_cursor,
+        )
+
+
+class PointCountResponse(BaseModel):
+    count: int
+
+
+class PointSearchResponse(PointListResponse):
+    """Results of a **keyword** match, not of semantic retrieval.
+
+    Named and documented so it cannot be mistaken for ADR-0003's hybrid
+    retrieval: these points matched a full-text filter on `content`, they are
+    not ranked by relevance, and there is no score to report. Anything that
+    wants ranked results wants the agent retrieval path in plan 003.
+    """
+
+    query: str
+
+    @classmethod
+    def from_search(cls, page: PointPage, *, query: str) -> "PointSearchResponse":
+        return cls(
+            query=query,
+            points=[PointResponse.from_point(point) for point in page.points],
+            next_cursor=page.next_cursor,
+        )
+
+
+class PointCreateRequest(BaseModel):
+    """Create one point. The server assigns identity, ordering, and provenance.
+
+    `after_point_id` positions the new point rather than a raw `order_id`: the
+    caller says where in the sequence it goes and the server computes the
+    fractional key and relinks neighbours, which a client-supplied `order_id`
+    could not do correctly (ADR-0002). `None` means "at the start of the file".
+    """
+
+    model_config = ConfigDict(extra="forbid")
+
+    file_id: uuid.UUID
+    content: str = Field(min_length=1)
+    content_type: str = "paragraph"
+    after_point_id: uuid.UUID | None = None
+
+
+class PointReplaceRequest(BaseModel):
+    """Replace a point's content under a version guard.
+
+    `version` here is the *expected* version, not a value being written — the
+    optimistic-concurrency precondition. A mismatch is `409`.
+    """
+
+    model_config = ConfigDict(extra="forbid")
+
+    content: str = Field(min_length=1)
+    content_type: str | None = None
+    version: int
+
+
+class PointPayloadPatchRequest(BaseModel):
+    """Payload-only update of caller-writable fields, under a version guard."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    payload: dict[str, object]
+    version: int
+
+
+class PointReorderRequest(BaseModel):
+    """Move a point to sit immediately after `after_point_id`.
+
+    `None` moves it to the front of the file. Expressed as a neighbour rather
+    than an `order_id` for the same reason as `PointCreateRequest`.
+    """
+
+    model_config = ConfigDict(extra="forbid")
+
+    after_point_id: uuid.UUID | None = None
+    version: int

src/application/auth/__init__.py

@@ -0,0 +1,29 @@
+"""API-key authentication and tenant resolution (ADR-0008).
+
+`resolve_auth_context` is the entry point: it takes a bearer token and
+returns a trusted `AuthContext`. Everything downstream of the FastAPI
+boundary receives `tenant_id` only through that context — never from a
+request body, query string, or object metadata.
+"""
+
+from src.application.auth.context import AuthContext
+from src.application.auth.errors import (
+    AuthError,
+    InvalidApiKeyError,
+    MissingScopeError,
+    TenantInactiveError,
+)
+from src.application.auth.keys import generate_api_key, hash_secret, verify_secret
+from src.application.auth.service import resolve_auth_context
+
+__all__ = [
+    "AuthContext",
+    "AuthError",
+    "InvalidApiKeyError",
+    "MissingScopeError",
+    "TenantInactiveError",
+    "generate_api_key",
+    "hash_secret",
+    "resolve_auth_context",
+    "verify_secret",
+]

src/application/auth/context.py

@@ -0,0 +1,16 @@
+"""The trusted request-scoped auth/tenant context (ADR-0008)."""
+
+import uuid
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class AuthContext:
+    tenant_id: uuid.UUID
+    tenant_slug: str
+    api_key_id: uuid.UUID
+    scopes: frozenset[str]
+    actor_type: str
+
+    def has_scope(self, scope: str) -> bool:
+        return scope in self.scopes or "admin" in self.scopes

src/application/auth/errors.py

@@ -0,0 +1,22 @@
+"""Auth failures (ADR-0008). No HTTP knowledge here — `src/api/errors.py` maps
+these to status codes.
+"""
+
+
+class AuthError(Exception):
+    """Base class for auth failures."""
+
+
+class InvalidApiKeyError(AuthError):
+    """The bearer token is missing, malformed, unknown, revoked, or expired.
+
+    Maps to `401`.
+    """
+
+
+class TenantInactiveError(AuthError):
+    """The key's tenant is suspended or deleted. Maps to `401`."""
+
+
+class MissingScopeError(AuthError):
+    """The key is valid but lacks a scope the route requires. Maps to `403`."""

src/application/auth/keys.py

@@ -0,0 +1,39 @@
+"""API-key generation and hashing (ADR-0008, ADR-0009).
+
+Keys are `sk_{prefix}_{secret}`. `prefix` is non-secret and indexed
+(`api_keys.key_prefix`); `secret` is 256 bits of `secrets.token_urlsafe`
+entropy, stored only as a SHA-256 hash. A random 256-bit secret does not
+benefit from a slow password-hashing KDF the way a human-chosen password
+does — the cost that defends against dictionary/brute-force guessing over a
+low-entropy input has nothing to defend here, and would only tax every
+request. Comparison is constant-time to avoid a hash-timing oracle.
+"""
+
+import hashlib
+import hmac
+import secrets
+
+_PREFIX_LENGTH = 16
+
+
+def generate_api_key() -> tuple[str, str, str]:
+    """Return `(key_prefix, secret, full_key)` for a newly issued key."""
+    key_prefix = secrets.token_hex(_PREFIX_LENGTH // 2)
+    secret = secrets.token_urlsafe(32)
+    return key_prefix, secret, f"sk_{key_prefix}_{secret}"
+
+
+def parse_api_key(full_key: str) -> tuple[str, str] | None:
+    """Return `(key_prefix, secret)`, or `None` if the token is malformed."""
+    parts = full_key.split("_", 2)
+    if len(parts) != 3 or parts[0] != "sk" or not parts[1] or not parts[2]:
+        return None
+    return parts[1], parts[2]
+
+
+def hash_secret(secret: str) -> str:
+    return hashlib.sha256(secret.encode("utf-8")).hexdigest()
+
+
+def verify_secret(secret: str, key_hash: str) -> bool:
+    return hmac.compare_digest(hash_secret(secret), key_hash)

src/application/auth/service.py

@@ -0,0 +1,86 @@
+"""Resolve a bearer token to a trusted `AuthContext` (ADR-0008).
+
+This opens and releases its own session rather than borrowing a
+request-scoped one, so auth resolution never pins a pool connection across
+the rest of the request — including the ADR-0017 ingestion work phase, which
+must run with no Postgres session held open at all.
+"""
+
+from datetime import UTC, datetime
+
+import structlog
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.application.auth.context import AuthContext
+from src.application.auth.errors import InvalidApiKeyError, TenantInactiveError
+from src.application.auth.keys import parse_api_key, verify_secret
+from src.infrastructure.postgres.repositories import api_keys as api_keys_repo
+from src.infrastructure.postgres.repositories import tenants as tenants_repo
+
+logger = structlog.get_logger(__name__)
+
+
+async def resolve_auth_context(
+    sessionmaker: async_sessionmaker[AsyncSession], bearer_token: str
+) -> AuthContext:
+    """Resolve a bearer token, logging the outcome either way (ADR-0011).
+
+    This runs on every authenticated request, so `auth.failed` is the one
+    event most likely to matter first when diagnosing a client integration
+    issue -- and the reason string alone (never logged; it can echo back
+    attacker-supplied key material) is not enough to tell a malformed token
+    apart from a revoked one without this.
+    """
+    parsed = parse_api_key(bearer_token)
+    if parsed is None:
+        logger.warning("auth.failed", reason="malformed_key")
+        raise InvalidApiKeyError("malformed API key")
+    key_prefix, secret = parsed
+
+    async with sessionmaker() as session:
+        api_key = await api_keys_repo.get_by_prefix(session, key_prefix)
+        if api_key is None or not verify_secret(secret, api_key.key_hash):
+            logger.warning("auth.failed", reason="unknown_key", key_prefix=key_prefix)
+            raise InvalidApiKeyError("unknown API key")
+        if api_key.status != "active":
+            logger.warning(
+                "auth.failed",
+                reason="key_inactive",
+                key_prefix=key_prefix,
+                api_key_id=str(api_key.id),
+                key_status=api_key.status,
+            )
+            raise InvalidApiKeyError(f"API key is {api_key.status}")
+        if api_key.expires_at is not None and api_key.expires_at <= datetime.now(UTC):
+            logger.warning(
+                "auth.failed",
+                reason="key_expired",
+                key_prefix=key_prefix,
+                api_key_id=str(api_key.id),
+            )
+            raise InvalidApiKeyError("API key has expired")
+
+        tenant = await tenants_repo.get_by_id(session, api_key.tenant_id)
+        if tenant is None or tenant.status != "active":
+            logger.warning(
+                "auth.failed",
+                reason="tenant_inactive",
+                key_prefix=key_prefix,
+                api_key_id=str(api_key.id),
+                tenant_id=str(api_key.tenant_id),
+            )
+            raise TenantInactiveError("tenant is not active")
+
+        logger.info(
+            "auth.succeeded",
+            tenant_id=str(tenant.id),
+            api_key_id=str(api_key.id),
+            actor_type=api_key.actor_type,
+        )
+        return AuthContext(
+            tenant_id=tenant.id,
+            tenant_slug=tenant.slug,
+            api_key_id=api_key.id,
+            scopes=frozenset(api_key.scopes),
+            actor_type=api_key.actor_type,
+        )

src/application/domains/__init__.py

@@ -0,0 +1,27 @@
+"""Tenant-domain management and the upload-time allowlist check (ADR-0009)."""
+
+from src.application.domains.errors import (
+    DomainAlreadyExistsError,
+    DomainsError,
+    UnknownDomainError,
+)
+from src.application.domains.models import DomainResult
+from src.application.domains.service import (
+    create_domain,
+    ensure_domain_allowed,
+    list_domains,
+    set_domain_status,
+    update_domain,
+)
+
+__all__ = [
+    "DomainAlreadyExistsError",
+    "DomainResult",
+    "DomainsError",
+    "UnknownDomainError",
+    "create_domain",
+    "ensure_domain_allowed",
+    "list_domains",
+    "set_domain_status",
+    "update_domain",
+]

src/application/domains/errors.py

@@ -0,0 +1,21 @@
+"""Domain-management failures (ADR-0009). No HTTP knowledge here —
+`src/api/errors.py` maps these to status codes.
+"""
+
+
+class DomainsError(Exception):
+    """Base class for tenant-domain failures."""
+
+
+class UnknownDomainError(DomainsError):
+    """The upload named a domain the tenant has not registered, or one that is
+    disabled. Maps to `400`.
+
+    Rejecting is the whole point: an unrecognized `domain` would otherwise
+    create a new Qdrant partition silently, and a file in a partition nothing
+    queries is invisible rather than failed (ADR-0009).
+    """
+
+
+class DomainAlreadyExistsError(DomainsError):
+    """The tenant already has a domain with this key. Maps to `409`."""

src/application/domains/models.py

@@ -0,0 +1,16 @@
+"""Transport-agnostic results for the domain-management service."""
+
+import uuid
+from dataclasses import dataclass
+from datetime import datetime
+
+
+@dataclass(frozen=True)
+class DomainResult:
+    id: uuid.UUID
+    domain: str
+    display_name: str
+    status: str
+    metadata: dict[str, object]
+    created_at: datetime
+    updated_at: datetime

src/application/domains/service.py

@@ -0,0 +1,164 @@
+"""Tenant-domain management (ADR-0009).
+
+A tenant's domain set is per-tenant and varies in size — one may run 14
+insurance lines, another 6 — so it is data, not an enum.
+
+`ensure_domain_allowed` is the reason this package exists: it is the strict
+allowlist check the upload path runs before anything is written. Everything
+else here is the management surface the calling backend uses to populate that
+allowlist, under its own `domains:write` scope so an upload key cannot create
+partitions.
+
+`tenant_id` is always a required parameter taken from `AuthContext`, never from
+a request body (ADR-0002).
+"""
+
+import uuid
+
+import structlog
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.application.domains.errors import DomainAlreadyExistsError, UnknownDomainError
+from src.application.domains.models import DomainResult
+from src.infrastructure.postgres.models.tenant_domain import TenantDomain
+from src.infrastructure.postgres.repositories import tenant_domains as repo
+
+logger = structlog.get_logger(__name__)
+
+
+async def _flush_and_refresh(session: AsyncSession, tenant_domain: TenantDomain) -> None:
+    """Materialize server-generated columns before the row leaves the session.
+
+    `updated_at` is `onupdate=func.now()`, so after an UPDATE its value lives in
+    the database, not in the instance. Reading it later would trigger a lazy
+    load outside any greenlet context (`MissingGreenlet`), so it is fetched here
+    while the session is still open.
+    """
+    await session.flush()
+    await session.refresh(tenant_domain)
+
+
+def _to_result(tenant_domain: TenantDomain) -> DomainResult:
+    return DomainResult(
+        id=tenant_domain.id,
+        domain=tenant_domain.domain,
+        display_name=tenant_domain.display_name,
+        status=tenant_domain.status,
+        metadata=tenant_domain.metadata_,
+        created_at=tenant_domain.created_at,
+        updated_at=tenant_domain.updated_at,
+    )
+
+
+async def ensure_domain_allowed(
+    session: AsyncSession, *, tenant_id: uuid.UUID, domain: str
+) -> None:
+    """Raise `UnknownDomainError` unless the tenant has this domain active.
+
+    Takes a session rather than a sessionmaker: the upload path calls this
+    inside its existing txn A, so the check costs no extra connection and
+    cannot pass and then go stale before the row is written.
+
+    Logs the rejection here rather than at the call site: this runs before any
+    `ingestion_jobs` row exists, so `upload_source_file`'s job-level
+    `ingestion.job.failed` event (ADR-0011) never fires for it -- without a log
+    here, a rejected upload would leave no operational trace at all.
+    """
+    tenant_domain = await repo.get(session, tenant_id=tenant_id, domain=domain)
+    if tenant_domain is None:
+        logger.warning(
+            "domain.rejected", tenant_id=str(tenant_id), domain=domain, reason="unregistered"
+        )
+        raise UnknownDomainError(
+            f"domain '{domain}' is not registered for this tenant; "
+            f"create it via POST /v1/domains before uploading to it"
+        )
+    if tenant_domain.status != "active":
+        logger.warning(
+            "domain.rejected", tenant_id=str(tenant_id), domain=domain, reason="disabled"
+        )
+        raise UnknownDomainError(f"domain '{domain}' is disabled for this tenant")
+
+
+async def list_domains(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    tenant_id: uuid.UUID,
+    include_disabled: bool = False,
+) -> list[DomainResult]:
+    async with sessionmaker() as session:
+        found = await repo.list_for_tenant(
+            session, tenant_id=tenant_id, include_disabled=include_disabled
+        )
+    return [_to_result(item) for item in found]
+
+
+async def create_domain(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    tenant_id: uuid.UUID,
+    domain: str,
+    display_name: str,
+    metadata: dict[str, object] | None = None,
+) -> DomainResult:
+    async with sessionmaker() as session:
+        if await repo.get(session, tenant_id=tenant_id, domain=domain) is not None:
+            raise DomainAlreadyExistsError(f"domain '{domain}' already exists for this tenant")
+        created = repo.create(
+            session,
+            tenant_id=tenant_id,
+            domain=domain,
+            display_name=display_name,
+            metadata=metadata,
+        )
+        await session.commit()
+
+    logger.info("domain.created", tenant_id=str(tenant_id), domain=domain)
+    return _to_result(created)
+
+
+async def update_domain(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    tenant_id: uuid.UUID,
+    domain: str,
+    display_name: str,
+) -> DomainResult:
+    """Only the label is mutable — see `repo.update_display_name`."""
+    async with sessionmaker() as session:
+        found = await repo.get(session, tenant_id=tenant_id, domain=domain)
+        if found is None:
+            raise UnknownDomainError(f"domain '{domain}' is not registered for this tenant")
+        repo.update_display_name(found, display_name=display_name)
+        await _flush_and_refresh(session, found)
+        await session.commit()
+        result = _to_result(found)
+
+    logger.info("domain.updated", tenant_id=str(tenant_id), domain=domain)
+    return result
+
+
+async def set_domain_status(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    tenant_id: uuid.UUID,
+    domain: str,
+    status: str,
+) -> DomainResult:
+    """Disable or re-enable a domain.
+
+    Disabling blocks new uploads and hides the domain from pickers. It does not
+    touch the points already indexed under it — removing those needs the
+    tenant-erasure workflow plan 001 defers.
+    """
+    async with sessionmaker() as session:
+        found = await repo.get(session, tenant_id=tenant_id, domain=domain)
+        if found is None:
+            raise UnknownDomainError(f"domain '{domain}' is not registered for this tenant")
+        repo.set_status(found, status=status)
+        await _flush_and_refresh(session, found)
+        await session.commit()
+        result = _to_result(found)
+
+    logger.info("domain.status_changed", tenant_id=str(tenant_id), domain=domain, status=status)
+    return result

src/application/files/__init__.py

@@ -0,0 +1,19 @@
+"""Source-file upload and status use cases (ADR-0008, ADR-0009, ADR-0017)."""
+
+from src.application.files.errors import FilesError, FileTooLargeError, InvalidUploadError
+from src.application.files.models import UploadResult, ValidatedUpload
+from src.application.files.status import FileStatusResult, get_file_status
+from src.application.files.upload import upload_source_file
+from src.application.files.validation import validate_and_hash_upload
+
+__all__ = [
+    "FileStatusResult",
+    "FileTooLargeError",
+    "FilesError",
+    "InvalidUploadError",
+    "UploadResult",
+    "ValidatedUpload",
+    "get_file_status",
+    "upload_source_file",
+    "validate_and_hash_upload",
+]

src/application/files/deletion.py

@@ -0,0 +1,81 @@
+"""`DELETE /v1/files/{file_id}` — retire a file and deactivate its points.
+
+Two stores have to agree here, and the phase boundaries are the same ones
+ingestion uses (ADR-0017): a short Postgres transaction to authorize, then the
+Qdrant work with **no session held**, then a short transaction to record the
+outcome. Holding a session across the sweep would pin a pool connection for the
+length of a multi-page delete.
+
+The order — points first, Postgres second — is deliberate. If the sweep dies
+half way, the row stays `active` and a retried `DELETE` finishes the job, since
+the sweep only ever looks at points that are still active. The reverse order
+would leave a row marked deleted while its points are still live and still
+retrievable by the agent, which is the failure that actually matters.
+"""
+
+import uuid
+from datetime import UTC, datetime
+from time import perf_counter
+
+import structlog
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.application.files.errors import SourceFileNotFoundError
+from src.application.points.deletion import soft_delete_file_points
+from src.application.ports.point_repository import PointRepository
+from src.infrastructure.postgres.repositories import source_files as source_files_repo
+
+logger = structlog.get_logger(__name__)
+
+
+async def delete_source_file(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    source_file_id: uuid.UUID,
+    actor: str,
+) -> int:
+    """Soft-delete a file: every active point, then the `source_files` row.
+
+    Returns how many points the sweep deactivated. Raises
+    `SourceFileNotFoundError` (`404`) when the file is not this tenant's — the
+    check happens before anything is written, so a probe for another tenant's
+    file id cannot deactivate a single point.
+
+    Idempotent: a second call finds no active points and a row already marked
+    `soft_deleted`, and returns `0`.
+    """
+    started = perf_counter()
+    async with sessionmaker() as session:
+        source_file = await source_files_repo.get_by_id(
+            session, tenant_id=tenant_id, source_file_id=source_file_id
+        )
+        if source_file is None:
+            raise SourceFileNotFoundError(f"file {source_file_id} not found")
+
+    points_soft_deleted = await soft_delete_file_points(
+        repository, tenant_id=tenant_id, file_id=source_file_id, actor=actor
+    )
+
+    async with sessionmaker() as session:
+        source_file = await source_files_repo.get_by_id(
+            session, tenant_id=tenant_id, source_file_id=source_file_id
+        )
+        if source_file is None:
+            raise SourceFileNotFoundError(f"file {source_file_id} not found")
+        source_files_repo.mark_soft_deleted(source_file, deleted_at=datetime.now(UTC))
+        await session.commit()
+
+    logger.info(
+        "files.soft_deleted",
+        tenant_id=str(tenant_id),
+        file_id=str(source_file_id),
+        points_soft_deleted=points_soft_deleted,
+        actor=actor,
+        # End to end, including both Postgres transactions. Comparing it with
+        # the sweep's own `duration_ms` on `points.file_soft_deleted` is what
+        # separates a slow Qdrant from a slow database.
+        duration_ms=round((perf_counter() - started) * 1000, 2),
+    )
+    return points_soft_deleted

src/application/files/errors.py

@@ -0,0 +1,25 @@
+"""Upload-validation failures (ADR-0008). No HTTP knowledge here —
+`src/api/errors.py` maps these to status codes.
+"""
+
+
+class FilesError(Exception):
+    """Base class for file-upload failures."""
+
+
+class InvalidUploadError(FilesError):
+    """Missing domain, empty file, or content that doesn't match its
+    declared extension. Maps to `400`.
+    """
+
+
+class FileTooLargeError(FilesError):
+    """The upload exceeds `INGESTION_MAX_UPLOAD_SIZE_MB`. Maps to `413`."""
+
+
+class SourceFileNotFoundError(FilesError):
+    """No such source file *within the requesting tenant*. Maps to `404`.
+
+    Same non-disclosure rule as points (ADR-0016): a cross-tenant file id and a
+    nonexistent one are indistinguishable to the caller, so this is never `403`.
+    """

src/application/files/models.py

@@ -0,0 +1,26 @@
+"""Domain models for the upload use case (ADR-0008, ADR-0009)."""
+
+import uuid
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class ValidatedUpload:
+    """The result of extension/content validation, before any I/O."""
+
+    source_type: str
+    content_type: str
+    content_sha256: str
+
+
+@dataclass(frozen=True)
+class UploadResult:
+    """What `upload_source_file` returns; the route maps this to `FileUploadResponse`."""
+
+    file_id: uuid.UUID
+    ingestion_job_id: uuid.UUID
+    status: str
+    chunks_indexed: int
+    is_new_attempt: bool
+    """`False` when an identical active upload already succeeded and no new
+    ingestion attempt was made (route returns `200`, not `201`)."""

src/application/files/status.py

@@ -0,0 +1,52 @@
+"""`GET /v1/files/{file_id}` read model (ADR-0008, ADR-0009)."""
+
+import uuid
+from dataclasses import dataclass
+
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.infrastructure.postgres.repositories import ingestion_jobs as jobs_repo
+from src.infrastructure.postgres.repositories import source_files as source_files_repo
+
+
+@dataclass(frozen=True)
+class FileStatusResult:
+    file_id: uuid.UUID
+    source_filename: str
+    domain: str
+    status: str
+    ingestion_job_id: uuid.UUID | None
+    ingestion_status: str | None
+    chunks_indexed: int
+
+
+async def get_file_status(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    tenant_id: uuid.UUID,
+    source_file_id: uuid.UUID,
+) -> FileStatusResult | None:
+    """Returns `None` when the file doesn't exist under this tenant — the
+    route maps that to `404`, never `403` (ADR-0016: cross-tenant access
+    returns 404).
+    """
+    async with sessionmaker() as session:
+        source_file = await source_files_repo.get_by_id(
+            session, tenant_id=tenant_id, source_file_id=source_file_id
+        )
+        if source_file is None:
+            return None
+
+        latest_job = await jobs_repo.get_latest_for_source_file(
+            session, tenant_id=tenant_id, source_file_id=source_file.id
+        )
+
+        return FileStatusResult(
+            file_id=source_file.id,
+            source_filename=source_file.source_filename,
+            domain=source_file.domain,
+            status=source_file.status,
+            ingestion_job_id=latest_job.id if latest_job else None,
+            ingestion_status=latest_job.status if latest_job else None,
+            chunks_indexed=latest_job.points_created if latest_job else 0,
+        )

src/application/files/storage_keys.py

@@ -0,0 +1,11 @@
+"""Object-storage key derivation (ADR-0013).
+
+Object keys are internal identifiers, never the caller-supplied filename.
+Pure and synchronous.
+"""
+
+import uuid
+
+
+def source_file_object_key(tenant_id: uuid.UUID, source_file_id: uuid.UUID) -> str:
+    return f"tenants/{tenant_id}/source-files/{source_file_id}/original"

src/application/files/upload.py

@@ -0,0 +1,366 @@
+"""`POST /v1/files` orchestration: the ADR-0017 three-phase upload.
+
+This service owns two separate short-lived sessions/transactions rather than
+one request-scoped session, because the request is two units of work
+(ADR-0012, ADR-0017):
+
+    txn A (short):  source_files [+ ingestion_jobs(status='running')], commit
+    no txn:          store bytes in MinIO, parse/chunk (threads),
+                     embed dense+sparse (bounded/batched)
+    txn B (short):  ingestion_jobs -> succeeded/failed, append event, commit
+
+No Postgres session is open during phase 2. A failure at any point between
+txn A and txn B still leaves a durable, inspectable `failed` job — never a
+job stuck in `running`. The whole request additionally holds one of
+`INGESTION_MAX_CONCURRENCY` process-wide slots (`503` when exhausted) and
+phase 2 is bounded by `INGESTION_TIMEOUT_SECONDS` (`504`) (ADR-0017, plan 001
+Phase 4).
+
+Phase 2 ends by upserting the embedded chunks as tenant-scoped Qdrant points
+(`src/application/points/`), so a successful upload is searchable by the time
+the `201` returns. The collection those points land in is provisioned by a
+deployment step, not by this path — see `src/cli/qdrant_bootstrap.py`.
+"""
+
+import uuid
+from collections.abc import Sequence
+
+import structlog
+from anyio import CapacityLimiter, Semaphore, fail_after, to_thread
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.application.auth.context import AuthContext
+from src.application.domains import ensure_domain_allowed
+from src.application.files.errors import InvalidUploadError
+from src.application.files.models import UploadResult
+from src.application.files.storage_keys import source_file_object_key
+from src.application.files.validation import validate_and_hash_upload
+from src.application.ingestion import (
+    ChunkTooLargeError,
+    DocumentParseError,
+    UnsupportedSourceTypeError,
+    parse_and_chunk_document,
+)
+from src.application.ingestion.bounds import acquire_ingestion_slot, enforce_chunk_limit
+from src.application.ingestion.embedding import embed_chunks
+from src.application.ingestion.errors import (
+    ChunkLimitExceededError,
+    EmbedderError,
+    IngestionTimeoutError,
+    PointIndexingError,
+)
+from src.application.points import index_chunks
+from src.application.ports.embedding import DenseEmbedder, SparseEmbedder
+from src.application.ports.object_storage import ObjectStorage
+from src.application.ports.point_storage import PointStorage
+from src.config import ChunkingSettings, IngestionSettings, QdrantSettings
+from src.infrastructure.postgres.repositories import ingestion_jobs as jobs_repo
+from src.infrastructure.postgres.repositories import source_files as source_files_repo
+
+logger = structlog.get_logger(__name__)
+
+
+async def _mark_job_failed(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    tenant_id: uuid.UUID,
+    ingestion_job_id: uuid.UUID,
+    error_code: str,
+    error_message: str,
+) -> None:
+    """Write the terminal `failed` job row and emit its log event together.
+
+    Every failure branch below calls this, so logging here once closes every
+    branch at once rather than duplicating a `logger.warning` at each call
+    site (CLAUDE.md, "prefer deep modules") -- previously only
+    `storage_upload_failed` and `timeout` did that ad hoc, and
+    `parse_failed`/`chunk_limit_exceeded`/`embedding_failed`/`index_failed`
+    logged nothing at all: visible in `ingestion_job_events` but invisible to
+    log-based alerting (ADR-0011).
+    """
+    async with sessionmaker() as session:
+        job = await jobs_repo.mark_terminal(
+            session,
+            tenant_id=tenant_id,
+            ingestion_job_id=ingestion_job_id,
+            status="failed",
+            error_code=error_code,
+            error_message=error_message,
+        )
+        if job is not None:
+            jobs_repo.append_event(
+                session,
+                tenant_id=tenant_id,
+                ingestion_job_id=ingestion_job_id,
+                level="error",
+                stage="received",
+                message=error_message,
+            )
+        await session.commit()
+
+    logger.warning(
+        "ingestion.job.failed",
+        tenant_id=str(tenant_id),
+        ingestion_job_id=str(ingestion_job_id),
+        error_code=error_code,
+        error_message=error_message,
+    )
+
+
+async def upload_source_file(
+    *,
+    sessionmaker: async_sessionmaker[AsyncSession],
+    storage: ObjectStorage,
+    point_storage: PointStorage,
+    auth: AuthContext,
+    domain: str,
+    filename: str,
+    data: bytes,
+    ingestion_settings: IngestionSettings,
+    chunking_settings: ChunkingSettings,
+    qdrant_settings: QdrantSettings,
+    thread_limiter: CapacityLimiter,
+    concurrency_limiter: Semaphore,
+    dense_embedders: Sequence[DenseEmbedder],
+    sparse_embedder: SparseEmbedder,
+) -> UploadResult:
+    domain = domain.strip()
+    if not domain:
+        raise InvalidUploadError("domain is required")
+
+    validated = await to_thread.run_sync(
+        lambda: validate_and_hash_upload(
+            filename=filename, data=data, max_size_bytes=ingestion_settings.max_upload_size_bytes
+        ),
+        limiter=thread_limiter,
+    )
+
+    async with acquire_ingestion_slot(concurrency_limiter):
+        async with sessionmaker() as session:
+            # Strict allowlist, checked inside txn A before anything is written
+            # (ADR-0009). An unregistered domain would otherwise create a new
+            # Qdrant partition silently, leaving the file invisible to
+            # retrieval rather than failing.
+            await ensure_domain_allowed(session, tenant_id=auth.tenant_id, domain=domain)
+
+            existing = await source_files_repo.find_active_by_content_hash(
+                session,
+                tenant_id=auth.tenant_id,
+                domain=domain,
+                content_sha256=validated.content_sha256,
+            )
+
+            if existing is not None:
+                latest_job = await jobs_repo.get_latest_for_source_file(
+                    session, tenant_id=auth.tenant_id, source_file_id=existing.id
+                )
+                if latest_job is not None and latest_job.status == "succeeded":
+                    logger.info(
+                        "files.upload.duplicate",
+                        tenant_id=str(auth.tenant_id),
+                        file_id=str(existing.id),
+                    )
+                    return UploadResult(
+                        file_id=existing.id,
+                        ingestion_job_id=latest_job.id,
+                        status=latest_job.status,
+                        chunks_indexed=latest_job.points_created,
+                        is_new_attempt=False,
+                    )
+                source_file_id = existing.id
+                object_key = existing.storage_uri or source_file_object_key(
+                    auth.tenant_id, source_file_id
+                )
+            else:
+                source_file_id = uuid.uuid4()
+                object_key = source_file_object_key(auth.tenant_id, source_file_id)
+                source_files_repo.create(
+                    session,
+                    source_file_id=source_file_id,
+                    tenant_id=auth.tenant_id,
+                    domain=domain,
+                    source_filename=filename,
+                    source_type=validated.source_type,
+                    content_sha256=validated.content_sha256,
+                    byte_size=len(data),
+                    storage_uri=object_key,
+                    created_by_api_key_id=auth.api_key_id,
+                )
+                # `ingestion_jobs.source_file_id` FKs to this row; flush so the
+                # insert below sees it, since the two mapped classes carry no
+                # ORM relationship for the unit of work to order by itself.
+                await session.flush()
+
+            job = jobs_repo.create_running(
+                session,
+                tenant_id=auth.tenant_id,
+                source_file_id=source_file_id,
+                requested_by_api_key_id=auth.api_key_id,
+                chunking_strategy=chunking_settings.strategy,
+            )
+            jobs_repo.append_event(
+                session,
+                tenant_id=auth.tenant_id,
+                ingestion_job_id=job.id,
+                level="info",
+                stage="received",
+                message="upload accepted, storing object",
+            )
+            await session.commit()
+            ingestion_job_id = job.id
+
+        logger.info(
+            "ingestion.job.started",
+            tenant_id=str(auth.tenant_id),
+            ingestion_job_id=str(ingestion_job_id),
+            file_id=str(source_file_id),
+            domain=domain,
+            source_type=validated.source_type,
+        )
+
+        # Phase 2: no Postgres session open across this work (ADR-0017),
+        # bounded end-to-end by INGESTION_TIMEOUT_SECONDS.
+        try:
+            with fail_after(ingestion_settings.timeout_seconds):
+                try:
+                    await storage.put_object(
+                        key=object_key, data=data, content_type=validated.content_type
+                    )
+                except Exception as exc:
+                    await _mark_job_failed(
+                        sessionmaker,
+                        tenant_id=auth.tenant_id,
+                        ingestion_job_id=ingestion_job_id,
+                        error_code="storage_upload_failed",
+                        error_message=f"failed to store object: {exc}",
+                    )
+                    raise
+
+                try:
+                    chunks = await parse_and_chunk_document(
+                        data,
+                        source_type=validated.source_type,
+                        file_id=source_file_id,
+                        settings=chunking_settings,
+                        limiter=thread_limiter,
+                    )
+                except (DocumentParseError, UnsupportedSourceTypeError, ChunkTooLargeError) as exc:
+                    await _mark_job_failed(
+                        sessionmaker,
+                        tenant_id=auth.tenant_id,
+                        ingestion_job_id=ingestion_job_id,
+                        error_code="parse_failed",
+                        error_message=str(exc),
+                    )
+                    raise
+
+                try:
+                    enforce_chunk_limit(chunks, max_chunks=ingestion_settings.max_chunks_per_file)
+                except ChunkLimitExceededError as exc:
+                    await _mark_job_failed(
+                        sessionmaker,
+                        tenant_id=auth.tenant_id,
+                        ingestion_job_id=ingestion_job_id,
+                        error_code="chunk_limit_exceeded",
+                        error_message=str(exc),
+                    )
+                    raise
+
+                try:
+                    embedded = await embed_chunks(
+                        chunks,
+                        dense_embedders=dense_embedders,
+                        sparse_embedder=sparse_embedder,
+                        settings=ingestion_settings,
+                        thread_limiter=thread_limiter,
+                    )
+                except EmbedderError as exc:
+                    await _mark_job_failed(
+                        sessionmaker,
+                        tenant_id=auth.tenant_id,
+                        ingestion_job_id=ingestion_job_id,
+                        error_code="embedding_failed",
+                        error_message=str(exc),
+                    )
+                    raise
+
+                try:
+                    indexed = await index_chunks(
+                        embedded,
+                        storage=point_storage,
+                        tenant_id=auth.tenant_id,
+                        domain=domain,
+                        file_id=source_file_id,
+                        source_filename=filename,
+                        source_type=validated.source_type,
+                        actor=f"api_key:{auth.api_key_id}",
+                        dense_embedders=dense_embedders,
+                        sparse_embedder=sparse_embedder,
+                        settings=qdrant_settings,
+                        thread_limiter=thread_limiter,
+                    )
+                except PointIndexingError as exc:
+                    await _mark_job_failed(
+                        sessionmaker,
+                        tenant_id=auth.tenant_id,
+                        ingestion_job_id=ingestion_job_id,
+                        error_code="index_failed",
+                        error_message=str(exc),
+                    )
+                    raise
+        except TimeoutError:
+            await _mark_job_failed(
+                sessionmaker,
+                tenant_id=auth.tenant_id,
+                ingestion_job_id=ingestion_job_id,
+                error_code="timeout",
+                error_message=f"ingestion exceeded {ingestion_settings.timeout_seconds}s",
+            )
+            raise IngestionTimeoutError(
+                f"ingestion exceeded {ingestion_settings.timeout_seconds}s"
+            ) from None
+
+        async with sessionmaker() as session:
+            await jobs_repo.mark_terminal(
+                session,
+                tenant_id=auth.tenant_id,
+                ingestion_job_id=ingestion_job_id,
+                status="succeeded",
+                # An upsert with deterministic ids cannot tell an insert from
+                # an overwrite, so every written point is reported here and
+                # `points_updated` stays 0 rather than being guessed at.
+                points_created=indexed.points_upserted,
+                points_soft_deleted=indexed.points_soft_deleted,
+            )
+            jobs_repo.append_event(
+                session,
+                tenant_id=auth.tenant_id,
+                ingestion_job_id=ingestion_job_id,
+                level="info",
+                stage="completed",
+                message="chunks parsed, embedded, and indexed",
+                details={
+                    "chunks_parsed": len(chunks),
+                    "chunks_embedded": len(embedded),
+                    "points_upserted": indexed.points_upserted,
+                    "points_soft_deleted": indexed.points_soft_deleted,
+                },
+            )
+            await session.commit()
+
+    logger.info(
+        "ingestion.job.completed",
+        tenant_id=str(auth.tenant_id),
+        ingestion_job_id=str(ingestion_job_id),
+        file_id=str(source_file_id),
+        chunks_parsed=len(chunks),
+        points_upserted=indexed.points_upserted,
+        points_soft_deleted=indexed.points_soft_deleted,
+    )
+    return UploadResult(
+        file_id=source_file_id,
+        ingestion_job_id=ingestion_job_id,
+        status="succeeded",
+        chunks_indexed=indexed.points_upserted,
+        is_new_attempt=True,
+    )

src/application/files/validation.py

@@ -0,0 +1,55 @@
+"""Upload extension/content-type/size validation (ADR-0008).
+
+Pure and synchronous: no I/O. `content_sha256` computation lives here too —
+hashing is blocking CPU work (ADR-0017), so the caller runs this whole
+function through `anyio.to_thread.run_sync` with the ingestion
+`CapacityLimiter`, the same rule applied to parsing/chunking.
+"""
+
+import hashlib
+
+from src.application.files.errors import FileTooLargeError, InvalidUploadError
+from src.application.files.models import ValidatedUpload
+from src.application.ingestion.errors import UnsupportedSourceTypeError
+
+_CONTENT_TYPES = {
+    "csv": "text/csv",
+    "xlsx": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+    "docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
+}
+_OOXML_MAGIC = b"PK\x03\x04"
+
+
+def _source_type_from_filename(filename: str) -> str:
+    suffix = filename.rsplit(".", 1)[-1].lower() if "." in filename else ""
+    if suffix == "doc":
+        raise UnsupportedSourceTypeError(
+            "legacy .doc is not ingestible until an out-of-process conversion "
+            "service exists (ADR-0018)"
+        )
+    if suffix not in _CONTENT_TYPES:
+        raise UnsupportedSourceTypeError(f"'.{suffix}' is not an ingestible file type")
+    return suffix
+
+
+def validate_and_hash_upload(*, filename: str, data: bytes, max_size_bytes: int) -> ValidatedUpload:
+    source_type = _source_type_from_filename(filename)
+
+    if not data:
+        raise InvalidUploadError("uploaded file is empty")
+    if len(data) > max_size_bytes:
+        raise FileTooLargeError(
+            f"upload is {len(data)} bytes, over the {max_size_bytes}-byte limit"
+        )
+
+    is_ooxml = data[:4] == _OOXML_MAGIC
+    if source_type in ("docx", "xlsx") and not is_ooxml:
+        raise InvalidUploadError(f"content does not match the declared .{source_type} extension")
+    if source_type == "csv" and is_ooxml:
+        raise InvalidUploadError("content does not match the declared .csv extension")
+
+    return ValidatedUpload(
+        source_type=source_type,
+        content_type=_CONTENT_TYPES[source_type],
+        content_sha256=hashlib.sha256(data).hexdigest(),
+    )

src/application/ingestion/__init__.py

@@ -1,9 +1,11 @@
 """Document parsing and fixed-size chunking (ADR-0004, ADR-0018).
 
-Everything here is pure and synchronous: no I/O, no ports, no SDK clients
-(ADR-0015 reserves ports for external side effects). Parsing and chunking are
-blocking CPU work, so callers run them through `anyio.to_thread.run_sync` with
-the ingestion `CapacityLimiter` rather than on the event loop (ADR-0017).
+`parse_and_chunk_document` is the entry point callers outside this package
+should use: it dispatches on source type and owns the
+`anyio.to_thread.run_sync` + `CapacityLimiter` offload required by ADR-0017.
+The individual parsers and `chunk_document` are pure, synchronous, and
+exported mainly for their own unit tests — calling them directly from an
+`async def` route or service is the defect ADR-0017 warns about.
 """
 
 from src.application.ingestion.chunking import chunk_document, chunk_id_for, split_by_tokens
@@ -22,6 +24,7 @@ from src.application.ingestion.models import (
     StructuralUnit,
 )
 from src.application.ingestion.normalization import normalize_persian_text
+from src.application.ingestion.pipeline import parse_and_chunk_document
 from src.application.ingestion.spreadsheet_parser import parse_csv, parse_xlsx
 from src.application.ingestion.tokenizer import count_tokens, get_encoder
 
@@ -40,6 +43,7 @@ __all__ = [
     "count_tokens",
     "get_encoder",
     "normalize_persian_text",
+    "parse_and_chunk_document",
     "parse_csv",
     "parse_docx",
     "parse_xlsx",

src/application/ingestion/bounds.py

@@ -0,0 +1,48 @@
+"""Request bounds for inline ingestion (ADR-0017).
+
+Three independent bounds, each mapping to its own status code: the chunk
+ceiling (`413`, checked before embedding starts), process-wide concurrency
+(`503` + `Retry-After`, rejected rather than queued), and the work-phase
+deadline (`504`, and the caller must still write a terminal job status).
+"""
+
+from collections.abc import AsyncIterator, Sequence
+from contextlib import asynccontextmanager
+
+from anyio import Semaphore, WouldBlock
+
+from src.application.ingestion.errors import ChunkLimitExceededError, IngestionAtCapacityError
+from src.application.ingestion.models import Chunk
+
+
+def enforce_chunk_limit(chunks: Sequence[Chunk], *, max_chunks: int) -> None:
+    """Raise `ChunkLimitExceededError` if `chunks` exceeds `max_chunks`.
+
+    Call this immediately after parsing/chunking and before any embedding
+    call — the ceiling must be discovered up front, not mid-batch.
+    """
+    if len(chunks) > max_chunks:
+        raise ChunkLimitExceededError(
+            f"document produced {len(chunks)} chunks, over the {max_chunks}-chunk limit"
+        )
+
+
+@asynccontextmanager
+async def acquire_ingestion_slot(limiter: Semaphore) -> AsyncIterator[None]:
+    """Hold one of `INGESTION_MAX_CONCURRENCY` process-wide slots for the block.
+
+    `limiter` is an `anyio.Semaphore` created once in the lifespan. Rejects
+    immediately with `IngestionAtCapacityError` when the process is already at
+    capacity, rather than queueing the request behind an unbounded wait
+    (ADR-0017) — the semaphore's own async `acquire()` would do the latter.
+    """
+    try:
+        limiter.acquire_nowait()
+    except WouldBlock:
+        raise IngestionAtCapacityError(
+            "ingestion is at capacity; retry after the configured backoff"
+        ) from None
+    try:
+        yield
+    finally:
+        limiter.release()

src/application/ingestion/embedding.py

@@ -0,0 +1,112 @@
+"""The one caller-facing entry point for embedding chunks (ADR-0001, ADR-0017).
+
+`embed_chunks` is the only version of this step callers should reach for: it
+owns batching, the `embed_concurrency` semaphore bounding in-flight dense
+batches, and the `anyio.to_thread.run_sync` + `CapacityLimiter` offload for
+the blocking BM25 pipeline. Composing these correctly at every call site is
+exactly the obligation a deep module absorbs once (see CLAUDE.md's "prefer
+deep modules").
+
+Per-provider text shaping — task prefixes, `keep_alive`, request payload —
+belongs to the adapters in `src/infrastructure/embedding/`, not here. This
+module knows only that an embedder turns texts into vectors.
+"""
+
+import asyncio
+from collections.abc import Sequence
+from functools import partial
+
+from anyio import CapacityLimiter, to_thread
+
+from src.application.ingestion.errors import EmbedderError
+from src.application.ingestion.models import Chunk, EmbeddedChunk
+from src.application.ports.embedding import DenseEmbedder, SparseEmbedder
+from src.config import IngestionSettings
+
+
+def _batches(texts: Sequence[str], size: int) -> list[Sequence[str]]:
+    return [texts[i : i + size] for i in range(0, len(texts), size)]
+
+
+async def _embed_dense_bounded(
+    embedder: DenseEmbedder,
+    batch: Sequence[str],
+    *,
+    semaphore: asyncio.Semaphore,
+) -> list[list[float]]:
+    async with semaphore:
+        try:
+            return await embedder.embed_batch(batch)
+        except Exception as exc:
+            raise EmbedderError(f"{embedder.name} embedding batch failed: {exc}") from exc
+
+
+async def _embed_dense_all(
+    embedder: DenseEmbedder,
+    texts: Sequence[str],
+    *,
+    batch_size: int,
+    semaphore: asyncio.Semaphore,
+) -> list[list[float]]:
+    batches = _batches(texts, batch_size)
+    results = await asyncio.gather(
+        *(_embed_dense_bounded(embedder, batch, semaphore=semaphore) for batch in batches)
+    )
+    return [vector for batch_result in results for vector in batch_result]
+
+
+def _embed_sparse_sync(embedder: SparseEmbedder, texts: Sequence[str]):
+    try:
+        return embedder.embed_batch(texts)
+    except Exception as exc:
+        raise EmbedderError(f"{embedder.name} embedding batch failed: {exc}") from exc
+
+
+async def embed_chunks(
+    chunks: Sequence[Chunk],
+    *,
+    dense_embedders: Sequence[DenseEmbedder],
+    sparse_embedder: SparseEmbedder,
+    settings: IngestionSettings,
+    thread_limiter: CapacityLimiter,
+) -> list[EmbeddedChunk]:
+    """Embed every chunk into all dense vectors plus the sparse vector.
+
+    Dense embedders run concurrently with each other; each one's batches are
+    concurrent among themselves too, bounded by one `embed_concurrency`
+    semaphore shared across all dense embedders (ADR-0017: the limit exists
+    for both providers' rate limits and the self-hosted server's capacity —
+    not a per-provider budget). The sparse (BM25) pass is blocking and runs
+    once, off the event loop.
+
+    Raises `EmbedderError` (502) if any embedder call fails.
+    """
+    if not chunks:
+        return []
+
+    texts = [chunk.content for chunk in chunks]
+    semaphore = asyncio.Semaphore(settings.embed_concurrency)
+
+    dense_task = asyncio.gather(
+        *(
+            _embed_dense_all(
+                embedder, texts, batch_size=settings.embed_batch_size, semaphore=semaphore
+            )
+            for embedder in dense_embedders
+        )
+    )
+    sparse_task = to_thread.run_sync(
+        partial(_embed_sparse_sync, sparse_embedder, texts), limiter=thread_limiter
+    )
+    dense_results, sparse_vectors = await asyncio.gather(dense_task, sparse_task)
+
+    dense_by_name = {
+        embedder.name: vectors
+        for embedder, vectors in zip(dense_embedders, dense_results, strict=True)
+    }
+
+    embedded: list[EmbeddedChunk] = []
+    for index, chunk in enumerate(chunks):
+        dense = {name: vectors[index] for name, vectors in dense_by_name.items()}
+        embedded.append(EmbeddedChunk(chunk=chunk, dense=dense, sparse=sparse_vectors[index]))
+    return embedded

src/application/ingestion/errors.py

@@ -39,3 +39,35 @@ class ChunkTooLargeError(IngestionError):
     against is silent — `nomic-embed-text-v2-moe` truncates over-long input
     without raising (ADR-0004).
     """
+
+
+class EmbedderError(IngestionError):
+    """A dense or sparse embedder call failed (transport error, non-2xx, or
+    a malformed response).
+
+    Maps to `502` per ADR-0017.
+    """
+
+
+class PointIndexingError(IngestionError):
+    """Upserting or soft-deleting Qdrant points failed.
+
+    Maps to `502` — like `EmbedderError`, this is an upstream dependency
+    failing, not a malformed request. Kept distinct from `EmbedderError` so the
+    job's `error_code` says which dependency broke.
+    """
+
+
+class IngestionAtCapacityError(IngestionError):
+    """`INGESTION_MAX_CONCURRENCY` in-process ingestions are already running.
+
+    Maps to `503` with `Retry-After`, not a queued wait (ADR-0017).
+    """
+
+
+class IngestionTimeoutError(IngestionError):
+    """The work phase (parse/embed/upsert) exceeded `INGESTION_TIMEOUT_SECONDS`.
+
+    Maps to `504`. The caller must still write a terminal `failed` job status
+    before this propagates (ADR-0017).
+    """

src/application/ingestion/models.py

@@ -64,3 +64,28 @@ class Chunk(BaseModel):
     next_chunk_id: uuid.UUID | None = None
     token_count: int
     character_count: int
+
+
+class SparseVector(BaseModel):
+    """A sparse (term-index -> weight) vector, Qdrant's `modifier="idf"` shape.
+
+    Kept free of the `qdrant_client` SDK (ADR-0015: ports carry no infra
+    imports) — `src/infrastructure/qdrant/` converts this to the SDK's own
+    `SparseVector` type at upsert time (Phase 5).
+    """
+
+    indices: list[int]
+    values: list[float]
+
+
+class EmbeddedChunk(BaseModel):
+    """A chunk plus every vector it will be upserted with (ADR-0001).
+
+    `dense` is keyed by named-vector name (`dense_nomic`, `dense_openai`).
+    `late_interaction` is deliberately absent — not computed at ingest
+    (ADR-0017).
+    """
+
+    chunk: Chunk
+    dense: dict[str, list[float]]
+    sparse: SparseVector

src/application/ingestion/pipeline.py

@@ -0,0 +1,65 @@
+"""The one caller-facing entry point for parsing and chunking (ADR-0017).
+
+`parse_docx`/`parse_csv`/`parse_xlsx`/`chunk_document` are blocking, pure
+functions; calling any of them directly from an `async def` route or service
+is the defect ADR-0017 names explicitly ("one large `python-docx` parse would
+stall every concurrent request"). `parse_and_chunk_document` is the only
+version of this pipeline callers should reach for: it owns source-type
+dispatch and the `anyio.to_thread.run_sync` + `CapacityLimiter` offload, so
+that obligation cannot be forgotten at a call site.
+"""
+
+import uuid
+from functools import partial
+
+from anyio import CapacityLimiter, to_thread
+
+from src.application.ingestion.chunking import chunk_document
+from src.application.ingestion.docx_parser import parse_docx
+from src.application.ingestion.errors import UnsupportedSourceTypeError
+from src.application.ingestion.models import Chunk, ParsedDocument
+from src.application.ingestion.spreadsheet_parser import parse_csv, parse_xlsx
+from src.config import ChunkingSettings
+
+_PARSERS = {"csv", "xlsx", "docx"}
+
+
+def _parse(data: bytes, source_type: str, settings: ChunkingSettings) -> ParsedDocument:
+    if source_type == "docx":
+        return parse_docx(data, settings)
+    if source_type == "xlsx":
+        return parse_xlsx(data)
+    if source_type == "csv":
+        return parse_csv(data)
+    raise UnsupportedSourceTypeError(f"'{source_type}' is not an ingestible source type")
+
+
+def _parse_and_chunk(
+    data: bytes, source_type: str, file_id: uuid.UUID, settings: ChunkingSettings
+) -> list[Chunk]:
+    parsed = _parse(data, source_type, settings)
+    return chunk_document(parsed, file_id=file_id, settings=settings)
+
+
+async def parse_and_chunk_document(
+    data: bytes,
+    *,
+    source_type: str,
+    file_id: uuid.UUID,
+    settings: ChunkingSettings,
+    limiter: CapacityLimiter,
+) -> list[Chunk]:
+    """Parse and chunk a document off the event loop, bounded by `limiter`.
+
+    Raises `UnsupportedSourceTypeError` (415), `DocumentParseError` (400), or
+    `ChunkTooLargeError` — see `src/application/ingestion/errors.py`. Callers
+    map these to status codes; this module carries no HTTP knowledge
+    (ADR-0015). The `max_chunks_per_file` ceiling (413) is enforced by the
+    caller, not here — see Phase 4 of plan 001.
+    """
+    if source_type not in _PARSERS:
+        raise UnsupportedSourceTypeError(f"'{source_type}' is not an ingestible source type")
+    return await to_thread.run_sync(
+        partial(_parse_and_chunk, data, source_type, file_id, settings),
+        limiter=limiter,
+    )

src/application/points/__init__.py

@@ -0,0 +1,18 @@
+"""Ingestion-generated Qdrant point CRUD (ADR-0001, ADR-0002).
+
+`index_chunks` is the entry point callers outside this package should use: it
+dispatches payload construction, batching, bounded-concurrency upserts, and the
+post-success soft-delete sweep. `build_chunk_payload` and the batching helpers
+stay internal, exported mainly for their own unit tests.
+
+The `/v1/points` surface lives here too, in its own modules with their own
+entry points: `queries.py` for the read paths and `deletion.py` for soft delete
+with neighbour relinking. They share this package because they share ADR-0001's
+payload schema, not because they share a caller — `index_chunks` writes a whole
+file at once, while those serve one admin edit at a time.
+"""
+
+from src.application.points.indexing import IndexingResult, index_chunks
+from src.application.points.models import ChunkPoint
+
+__all__ = ["ChunkPoint", "IndexingResult", "index_chunks"]

src/application/points/deletion.py

@@ -0,0 +1,253 @@
+"""Soft delete for `/v1/points` and for a whole file's points (ADR-0002).
+
+The caller-facing entry points are `soft_delete_point` and
+`soft_delete_file_points`. Routers call these; `patches_for_removal` and the
+planning helpers stay internal, because getting a delete right is exactly the
+composition a caller should not have to reassemble: read the point, load its
+neighbours, compute the patches still missing, send them in **one** batch,
+verify they landed, and retry against fresh versions if they did not.
+
+Why the retry exists. Qdrant has no multi-point transaction, so a batch whose
+second operation loses a version race applies its first operation anyway — and
+a filtered `set_payload` that matched nothing still reports success. That
+combination means "did my write land?" is only answerable by reading back, and
+a single-shot delete would be able to leave the deactivation applied and a
+neighbour's pointer stale. Since `patches_for_removal` plans from current state
+towards a fixed end state, simply re-planning emits precisely the patches that
+did not land, so the loop converges instead of re-doing work. Only exhausting
+the attempts raises `PointVersionConflictError` (`409`).
+
+Deleting an already-inactive point falls out of the same machinery rather than
+needing a special case: its neighbours were relinked by the first delete, so the
+plan is empty and the call is a no-op success — not a `404`, and not a second
+relink.
+"""
+
+import uuid
+from collections.abc import Mapping, Sequence
+from datetime import UTC, datetime
+from time import perf_counter
+
+import structlog
+
+from src.application.points.errors import PointVersionConflictError
+from src.application.points.point import Point, PointNotFoundError
+from src.application.points.relinking import (
+    neighbour_ids,
+    patch_for_deactivation,
+    patches_for_removal,
+)
+from src.application.ports.point_repository import PayloadPatch, PointRepository
+
+logger = structlog.get_logger(__name__)
+
+# Three plan-apply rounds, then a final verifying plan. Each round only re-emits
+# what a concurrent writer displaced, so a caller that legitimately needs more
+# than this is contending on the same points continuously and deserves the
+# `409` rather than an unbounded loop inside a request.
+_MAX_ATTEMPTS = 3
+
+# One sweep page. Matches ADR-0002's 100-operation batch cap, so a page of
+# points is always expressible as a single `points/batch` request.
+_SWEEP_BATCH_SIZE = 100
+
+# A hard ceiling on sweep rounds, so a file being concurrently re-ingested while
+# it is deleted cannot spin here for the life of the request.
+_MAX_SWEEP_ROUNDS = 1_000
+
+
+def _elapsed_ms(started: float) -> float:
+    """Wall-clock milliseconds since `started` (ADR-0011's `duration_ms`).
+
+    Worth carrying on these events even though the relinking itself is O(1):
+    what a delete actually spends is Qdrant round trips, and the whole-file
+    sweep spends a number of them proportional to the file's length. Timing the
+    operation is the only way to tell a slow store from a contended one, which
+    `rounds` on the same event then disambiguates.
+    """
+    return round((perf_counter() - started) * 1000, 2)
+
+
+async def soft_delete_point(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    point_id: uuid.UUID,
+    actor: str,
+) -> Point:
+    """Deactivate one point and relink its neighbours around the gap.
+
+    Returns the point as it now stands. Raises `PointNotFoundError` (`404`) if
+    it is not this tenant's — the same non-disclosure rule the read paths
+    follow — or `PointVersionConflictError` (`409`) if concurrent writers keep
+    displacing the plan.
+    """
+    started = perf_counter()
+    rounds = 0
+    point, patches = await _plan_removal(
+        repository, tenant_id=tenant_id, point_id=point_id, actor=actor
+    )
+
+    for _ in range(_MAX_ATTEMPTS):
+        if not patches:
+            break
+        await repository.apply_patches(tenant_id=tenant_id, patches=patches)
+        rounds += 1
+        # The next plan doubles as verification: anything that did not land is
+        # still missing from the end state and comes back as a patch.
+        point, patches = await _plan_removal(
+            repository, tenant_id=tenant_id, point_id=point_id, actor=actor
+        )
+
+    if patches:
+        logger.warning(
+            "points.soft_delete.conflict",
+            tenant_id=str(tenant_id),
+            point_id=str(point_id),
+            file_id=str(point.file_id),
+            unsettled_points=[str(patch.point_id) for patch in patches],
+            rounds=rounds,
+            duration_ms=_elapsed_ms(started),
+        )
+        raise PointVersionConflictError(
+            f"point {point_id} could not be soft-deleted under concurrent modification"
+        )
+
+    if rounds:
+        logger.info(
+            "points.soft_deleted",
+            tenant_id=str(tenant_id),
+            point_id=str(point_id),
+            file_id=str(point.file_id),
+            version=point.version,
+            actor=actor,
+            # `rounds` is 1 unless a concurrent writer forced a re-plan, so a
+            # rising value here is contention, not slow relinking.
+            rounds=rounds,
+            duration_ms=_elapsed_ms(started),
+        )
+    return point
+
+
+async def soft_delete_file_points(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    file_id: uuid.UUID,
+    actor: str,
+) -> int:
+    """Deactivate every active point of one file, in batches.
+
+    No relinking: the whole file leaves the sequence at once, so no surviving
+    active point can be left pointing at a deactivated one, and the chain is
+    preserved intact for whoever reads the deleted file later.
+
+    Returns how many points were active when the sweep reached them. Each round
+    re-lists from the start rather than paging with a cursor — deactivated
+    points drop straight out of the default listing, so the listing itself is
+    the progress check, and a round that attempts the exact same ids as the one
+    before it made no progress and raises `PointVersionConflictError`.
+    """
+    started = perf_counter()
+    swept: set[uuid.UUID] = set()
+    previous_attempt: frozenset[uuid.UUID] = frozenset()
+    rounds = 0
+
+    for _ in range(_MAX_SWEEP_ROUNDS):
+        page = await repository.list_by_file(
+            tenant_id=tenant_id, file_id=file_id, limit=_SWEEP_BATCH_SIZE
+        )
+        if not page.points:
+            if swept:
+                logger.info(
+                    "points.file_soft_deleted",
+                    tenant_id=str(tenant_id),
+                    file_id=str(file_id),
+                    points_soft_deleted=len(swept),
+                    actor=actor,
+                    # Two Qdrant round trips per round, so this is the delete
+                    # path whose cost tracks the size of the file.
+                    rounds=rounds,
+                    duration_ms=_elapsed_ms(started),
+                )
+            return len(swept)
+
+        attempt = frozenset(point.point_id for point in page.points)
+        if attempt == previous_attempt:
+            logger.warning(
+                "points.file_soft_delete.conflict",
+                tenant_id=str(tenant_id),
+                file_id=str(file_id),
+                unsettled_points=[str(point_id) for point_id in sorted(attempt, key=str)],
+                rounds=rounds,
+                duration_ms=_elapsed_ms(started),
+            )
+            raise PointVersionConflictError(
+                f"file {file_id} could not be soft-deleted under concurrent modification"
+            )
+        previous_attempt = attempt
+
+        now = datetime.now(UTC)
+        await repository.apply_patches(
+            tenant_id=tenant_id,
+            patches=[patch_for_deactivation(point, actor=actor, now=now) for point in page.points],
+        )
+        swept |= attempt
+        rounds += 1
+
+    logger.warning(
+        "points.file_soft_delete.conflict",
+        tenant_id=str(tenant_id),
+        file_id=str(file_id),
+        reason="sweep_rounds_exhausted",
+        rounds=rounds,
+        duration_ms=_elapsed_ms(started),
+    )
+    raise PointVersionConflictError(f"file {file_id} still had active points after the sweep")
+
+
+async def _plan_removal(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    point_id: uuid.UUID,
+    actor: str,
+) -> tuple[Point, tuple[PayloadPatch, ...]]:
+    point = await repository.get(tenant_id=tenant_id, point_id=point_id)
+    if point is None:
+        raise PointNotFoundError(f"point {point_id} not found")
+
+    neighbours = await _load_neighbours(repository, tenant_id=tenant_id, point=point)
+    _warn_on_missing_neighbours(point, neighbours, tenant_id=tenant_id)
+    patches = patches_for_removal(point, neighbours, actor=actor, now=datetime.now(UTC))
+    return point, patches
+
+
+async def _load_neighbours(
+    repository: PointRepository, *, tenant_id: uuid.UUID, point: Point
+) -> dict[uuid.UUID, Point]:
+    wanted: Sequence[uuid.UUID] = neighbour_ids(point)
+    if not wanted:
+        return {}
+    found = await repository.get_many(tenant_id=tenant_id, point_ids=wanted)
+    return {neighbour.point_id: neighbour for neighbour in found}
+
+
+def _warn_on_missing_neighbours(
+    point: Point, neighbours: Mapping[uuid.UUID, Point], *, tenant_id: uuid.UUID
+) -> None:
+    """A pointer naming a point that is not there means the chain is already broken.
+
+    Worth a log line rather than an exception: the delete can still complete the
+    part of the relink that does exist, and refusing would leave the caller with
+    a point it cannot remove through any endpoint.
+    """
+    missing = [pointer for pointer in neighbour_ids(point) if pointer not in neighbours]
+    if missing:
+        logger.warning(
+            "points.relink.neighbour_missing",
+            tenant_id=str(tenant_id),
+            point_id=str(point.point_id),
+            file_id=str(point.file_id),
+            missing_neighbours=[str(pointer) for pointer in missing],
+        )

src/application/points/errors.py

@@ -0,0 +1,19 @@
+"""Mutation failures for the `/v1/points` write paths (ADR-0002).
+
+No HTTP knowledge here — `src/api/errors.py` owns the status mapping. Absence
+lives on `PointNotFoundError` in `point.py`, next to the model whose read paths
+raise it; this module is for the failures only a *write* can produce.
+"""
+
+
+class PointVersionConflictError(Exception):
+    """A version-guarded write could not be landed against a moving target.
+
+    Raised when the service has re-read, recomputed, and re-applied its patches
+    the allowed number of times and the desired state still has not settled —
+    something else is writing the same points concurrently. Maps to `409`.
+
+    This is not "the guard fired once": a single stale guard is expected and is
+    retried, because Qdrant reports success for a filtered `set_payload` that
+    matched nothing. It means the retries were exhausted.
+    """

src/application/points/indexing.py

@@ -0,0 +1,203 @@
+"""The one caller-facing entry point for indexing embedded chunks (ADR-0001, ADR-0017).
+
+`index_chunks` is the only version of this step callers should reach for. It
+owns the whole composition a correct upsert needs:
+
+- building ADR-0001's payload for every chunk, with `tenant_id`/`domain` taken
+  from server-derived context;
+- offloading that (and the per-chunk content hashing) to a thread, since it is
+  blocking CPU work (ADR-0017);
+- batching at `QDRANT_UPSERT_BATCH_SIZE` inside ADR-0001's 64-256 band;
+- bounding in-flight batches with an `asyncio.Semaphore` rather than an
+  unbounded `gather` (ADR-0017);
+- running the soft-delete sweep for a shortened file **only after every batch
+  has succeeded**.
+
+That last ordering is the point, not an implementation detail — see
+`_deactivate_stale` below. `build_chunk_payload` and `_batches` stay internal;
+pushing that composition onto every call site is exactly the obligation a deep
+module absorbs once (CLAUDE.md, "prefer deep modules").
+"""
+
+import asyncio
+import uuid
+from collections.abc import Sequence
+from dataclasses import dataclass
+from datetime import UTC, datetime
+
+from anyio import CapacityLimiter, to_thread
+
+from src.application.ingestion.errors import PointIndexingError
+from src.application.ingestion.models import EmbeddedChunk
+from src.application.points.models import ChunkPoint
+from src.application.points.payload import build_chunk_payload
+from src.application.ports.embedding import DenseEmbedder, SparseEmbedder
+from src.application.ports.point_storage import PointStorage
+from src.config import QdrantSettings
+
+
+@dataclass(frozen=True)
+class IndexingResult:
+    """What one indexing pass wrote.
+
+    `points_upserted` counts points written, not points *created* — a
+    deterministic-id upsert cannot distinguish an insert from an overwrite, so
+    the ingestion job reports this as `points_created` and leaves
+    `points_updated` at zero rather than guessing.
+    """
+
+    points_upserted: int
+    points_soft_deleted: int
+
+
+def _embedding_model_version(
+    dense_embedders: Sequence[DenseEmbedder], sparse_embedder: SparseEmbedder
+) -> str:
+    """Compose the `embedding_model_version` payload value (ADR-0001).
+
+    Sorted so the string is stable regardless of the order the embedders were
+    wired in — an unstable value would make "which chunks need re-embedding?"
+    unanswerable, which is the field's only reason to exist.
+    """
+    versions = sorted(
+        [embedder.model_version for embedder in dense_embedders] + [sparse_embedder.model_version]
+    )
+    return "+".join(versions)
+
+
+def _batches(points: Sequence[ChunkPoint], size: int) -> list[Sequence[ChunkPoint]]:
+    return [points[i : i + size] for i in range(0, len(points), size)]
+
+
+def _build_points(
+    embedded: Sequence[EmbeddedChunk],
+    *,
+    tenant_id: uuid.UUID,
+    domain: str,
+    file_id: uuid.UUID,
+    source_filename: str,
+    source_type: str,
+    actor: str,
+    embedding_model_version: str,
+    indexed_at: datetime,
+) -> list[ChunkPoint]:
+    """Blocking: hashes every chunk's content. Always called through a thread."""
+    return [
+        ChunkPoint(
+            point_id=item.chunk.chunk_id,
+            dense=item.dense,
+            sparse=item.sparse,
+            payload=build_chunk_payload(
+                item.chunk,
+                tenant_id=tenant_id,
+                domain=domain,
+                file_id=file_id,
+                source_filename=source_filename,
+                source_type=source_type,
+                actor=actor,
+                embedding_model_version=embedding_model_version,
+                indexed_at=indexed_at,
+            ),
+        )
+        for item in embedded
+    ]
+
+
+async def _upsert_bounded(
+    storage: PointStorage, batch: Sequence[ChunkPoint], *, semaphore: asyncio.Semaphore
+) -> None:
+    async with semaphore:
+        try:
+            await storage.upsert_points(batch)
+        except Exception as exc:
+            raise PointIndexingError(f"upserting {len(batch)} points failed: {exc}") from exc
+
+
+async def _deactivate_stale(
+    storage: PointStorage,
+    *,
+    tenant_id: uuid.UUID,
+    file_id: uuid.UUID,
+    from_chunk_index: int,
+    actor: str,
+    deleted_at: datetime,
+) -> int:
+    """Soft-delete points left over from a longer previous version of this file.
+
+    Chunk indices are contiguous from 0, so "index >= the new chunk count" is
+    exactly the set of points the new version no longer produces.
+
+    This runs **only after every upsert has succeeded**, and that ordering is
+    what keeps a failed attempt from damaging a working index. ADR-0001's
+    deterministic point ids mean a re-ingestion overwrites in place, so literal
+    atomic replacement is not available; what *is* guaranteed is that a failed
+    attempt never removes content (it can only leave a prefix updated), and that
+    a retry converges to the correct state. See ADR-0017.
+    """
+    try:
+        return await storage.deactivate_points_from_index(
+            tenant_id=tenant_id,
+            file_id=file_id,
+            from_chunk_index=from_chunk_index,
+            deleted_at=deleted_at,
+            updated_by=actor,
+        )
+    except Exception as exc:
+        raise PointIndexingError(f"soft-deleting stale points failed: {exc}") from exc
+
+
+async def index_chunks(
+    embedded: Sequence[EmbeddedChunk],
+    *,
+    storage: PointStorage,
+    tenant_id: uuid.UUID,
+    domain: str,
+    file_id: uuid.UUID,
+    source_filename: str,
+    source_type: str,
+    actor: str,
+    dense_embedders: Sequence[DenseEmbedder],
+    sparse_embedder: SparseEmbedder,
+    settings: QdrantSettings,
+    thread_limiter: CapacityLimiter,
+) -> IndexingResult:
+    """Upsert every embedded chunk as a tenant-scoped point, then sweep leftovers.
+
+    Raises `PointIndexingError` (502) if any batch or the sweep fails.
+    """
+    if not embedded:
+        return IndexingResult(points_upserted=0, points_soft_deleted=0)
+
+    indexed_at = datetime.now(UTC)
+    points = await to_thread.run_sync(
+        lambda: _build_points(
+            embedded,
+            tenant_id=tenant_id,
+            domain=domain,
+            file_id=file_id,
+            source_filename=source_filename,
+            source_type=source_type,
+            actor=actor,
+            embedding_model_version=_embedding_model_version(dense_embedders, sparse_embedder),
+            indexed_at=indexed_at,
+        ),
+        limiter=thread_limiter,
+    )
+
+    semaphore = asyncio.Semaphore(settings.upsert_concurrency)
+    await asyncio.gather(
+        *(
+            _upsert_bounded(storage, batch, semaphore=semaphore)
+            for batch in _batches(points, settings.upsert_batch_size)
+        )
+    )
+
+    soft_deleted = await _deactivate_stale(
+        storage,
+        tenant_id=tenant_id,
+        file_id=file_id,
+        from_chunk_index=len(points),
+        actor=actor,
+        deleted_at=indexed_at,
+    )
+    return IndexingResult(points_upserted=len(points), points_soft_deleted=soft_deleted)

src/application/points/models.py

@@ -0,0 +1,29 @@
+"""Domain models for Qdrant points (ADR-0001).
+
+Deliberately free of the `qdrant_client` SDK: `src/infrastructure/qdrant/`
+converts these to `PointStruct`/`models.SparseVector` at upsert time
+(ADR-0015 — application code and ports carry no infrastructure imports).
+"""
+
+import uuid
+
+from pydantic import BaseModel, Field
+
+from src.application.ingestion.models import SparseVector
+
+
+class ChunkPoint(BaseModel):
+    """One chunk, ready to upsert: its id, its named vectors, and its payload.
+
+    `point_id` is the chunk's deterministic UUIDv5 (`chunk_id_for`), so
+    re-ingesting a file overwrites its points rather than duplicating them
+    (ADR-0001).
+
+    `dense` is keyed by named-vector name (`dense_nomic`, `dense_openai`).
+    `late_interaction` is absent — ADR-0017 does not compute it at ingest.
+    """
+
+    point_id: uuid.UUID
+    dense: dict[str, list[float]]
+    sparse: SparseVector
+    payload: dict[str, object] = Field(default_factory=dict)

src/application/points/payload.py

@@ -0,0 +1,70 @@
+"""Builds ADR-0001's point payload from a chunk plus its ingestion context.
+
+Internal to `src/application/points/` — callers use `index_chunks`, which owns
+composing this with batching and the deactivation sweep. Exported for its own
+unit tests, not as a surface to build payloads by hand.
+"""
+
+import uuid
+from datetime import datetime
+from hashlib import sha256
+
+from src.application.ingestion.models import Chunk
+
+
+def _optional_id(value: uuid.UUID | None) -> str | None:
+    return str(value) if value is not None else None
+
+
+def build_chunk_payload(
+    chunk: Chunk,
+    *,
+    tenant_id: uuid.UUID,
+    domain: str,
+    file_id: uuid.UUID,
+    source_filename: str,
+    source_type: str,
+    actor: str,
+    embedding_model_version: str,
+    indexed_at: datetime,
+) -> dict[str, object]:
+    """Return ADR-0001's payload for one chunk.
+
+    `tenant_id` and `domain` are passed in from the server-derived `AuthContext`
+    and the validated request — never from anything the client could assert as
+    authority (ADR-0002's non-negotiable isolation rule).
+
+    UUIDs are serialized as strings because the `tenant_id`/`domain`/`file_id`/
+    `previous_chunk_id`/`next_chunk_id` payload indexes are *keyword* indexes;
+    a native UUID would not match a keyword filter.
+
+    **Known gap — `version` is always written as `1`.** ADR-0002 uses this field
+    for optimistic concurrency between ingestion and manual `/v1/points` edits,
+    which needs a read-check-write (one read per point). Ingestion is
+    authoritative for its own file today, so writing `1` is safe until
+    `/v1/points` exists; plan 002 owns closing this.
+    """
+    timestamp = indexed_at.isoformat()
+    return {
+        "tenant_id": str(tenant_id),
+        "domain": domain,
+        "file_id": str(file_id),
+        "chunk_id": str(chunk.chunk_id),
+        "content": chunk.content,
+        "content_type": chunk.content_type.value,
+        "source_filename": source_filename,
+        "source_type": source_type,
+        "order_id": chunk.order_id,
+        "chunk_index": chunk.chunk_index,
+        "previous_chunk_id": _optional_id(chunk.previous_chunk_id),
+        "next_chunk_id": _optional_id(chunk.next_chunk_id),
+        "is_active": True,
+        "deleted_at": None,
+        "created_at": timestamp,
+        "updated_at": timestamp,
+        "created_by": actor,
+        "updated_by": actor,
+        "version": 1,
+        "content_hash": sha256(chunk.content.encode("utf-8")).hexdigest(),
+        "embedding_model_version": embedding_model_version,
+    }

src/application/points/point.py

@@ -0,0 +1,126 @@
+"""The caller-facing point model for `/v1/points` (ADR-0001, ADR-0002).
+
+`ChunkPoint` in `models.py` is the *write* shape ingestion upserts: an id, its
+named vectors, and an opaque payload dict. This module is the *read/edit* shape
+the `/v1/points` surface works in, where the payload's individual fields matter
+and the distinction between what a caller may write and what the server owns is
+a security boundary rather than a convention.
+
+That split is the reason this is a model and not a dict. ADR-0002's isolation
+rule ("never accepted as client-supplied input") and its optimistic-concurrency
+guard both fail open if a caller can smuggle `tenant_id` or `version` through a
+payload update, so the writable field set is enumerated in one place here and
+every write path validates against it.
+"""
+
+import uuid
+from collections.abc import Mapping
+from datetime import datetime
+from typing import Self
+
+from pydantic import BaseModel, ConfigDict, Field
+
+# Fields the server derives and a caller may never set, patch, or override.
+# `tenant_id` is authority, `version` is the concurrency guard, `chunk_index`
+# derives the point id, and the rest are provenance the server timestamps.
+SERVER_OWNED_FIELDS: frozenset[str] = frozenset(
+    {
+        "tenant_id",
+        "version",
+        "chunk_index",
+        "chunk_id",
+        "is_active",
+        "deleted_at",
+        "created_at",
+        "created_by",
+        "updated_at",
+        "updated_by",
+        "content_hash",
+        "embedding_model_version",
+    }
+)
+
+# Fields a caller may supply on create, replace, or payload patch. `order_id`
+# is writable on create but moves only through `PATCH /v1/points/{id}/order`
+# afterwards, because a bare `order_id` write would not relink neighbours.
+CALLER_WRITABLE_FIELDS: frozenset[str] = frozenset(
+    {
+        "content",
+        "content_type",
+        "domain",
+        "source_filename",
+        "source_type",
+        "order_id",
+        "previous_chunk_id",
+        "next_chunk_id",
+    }
+)
+
+
+class Point(BaseModel):
+    """One Qdrant point, read back with its ADR-0001 payload fields typed.
+
+    Vectors are deliberately absent: ADR-0008 returns them only when explicitly
+    requested, and every read path that does not ask for them should not pay to
+    deserialize them.
+    """
+
+    model_config = ConfigDict(frozen=True)
+
+    point_id: uuid.UUID
+
+    tenant_id: uuid.UUID
+    domain: str
+    file_id: uuid.UUID
+    chunk_id: uuid.UUID
+
+    content: str
+    content_type: str
+    source_filename: str
+    source_type: str
+
+    order_id: float
+    chunk_index: int
+    previous_chunk_id: uuid.UUID | None = None
+    next_chunk_id: uuid.UUID | None = None
+
+    is_active: bool = True
+    deleted_at: datetime | None = None
+
+    created_at: datetime
+    updated_at: datetime
+    created_by: str
+    updated_by: str
+
+    version: int
+    content_hash: str
+    embedding_model_version: str
+
+    # Only populated when the caller explicitly asked for vectors.
+    vectors: dict[str, object] | None = Field(default=None)
+
+    @classmethod
+    def from_payload(
+        cls,
+        point_id: uuid.UUID,
+        payload: Mapping[str, object],
+        *,
+        vectors: Mapping[str, object] | None = None,
+    ) -> Self:
+        """Build a `Point` from a raw Qdrant payload dict.
+
+        Lives here rather than in the Qdrant adapter so the payload field names
+        are declared once, next to the model that mirrors them. The adapter
+        stays responsible for talking to the SDK, not for knowing ADR-0001's
+        schema twice.
+        """
+        return cls.model_validate({**payload, "point_id": point_id, "vectors": vectors})
+
+
+class PointNotFoundError(LookupError):
+    """No such point *within the requesting tenant*.
+
+    Routes map this to `404`, never `403` — a caller must not be able to probe
+    for the existence of another tenant's point ids (ADR-0016). The error
+    deliberately carries no hint about which of the two cases occurred.
+    """

src/application/points/queries.py

@@ -0,0 +1,115 @@
+"""Read paths for `/v1/points` (ADR-0002, ADR-0008).
+
+The caller-facing entry points for point reads. Routers call these; they never
+touch the `PointRepository` directly, and never build a filter.
+
+This module is thin on purpose but not empty, and the two things it does own are
+exactly the ones a route would otherwise get wrong:
+
+- **`tenant_id` always comes from the caller's `AuthContext`.** Every function
+  takes it as a required keyword and hands it to the repository. Nothing here
+  reads a tenant from a query string or body.
+- **A keyword query is Persian-normalized before it reaches the index.**
+  Ingestion letter-folds chunk content (`normalize_persian_text`, ADR-0018), so
+  stored text contains Persian yeh/keheh. A query typed on an Arabic keyboard
+  carries U+064A/U+0643 and would match nothing at all — a silent empty result,
+  not an error. Folding the query the same way is what makes the two comparable.
+
+Reads emit no log events. The request middleware already records every call, and
+ADR-0011 reserves `INFO` for lifecycle events rather than per-read volume;
+mutations get their own events when those paths land.
+"""
+
+import uuid
+
+from src.application.ingestion.normalization import normalize_persian_text
+from src.application.points.point import Point, PointNotFoundError
+from src.application.ports.point_repository import PointPage, PointRepository
+
+
+async def get_point(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    point_id: uuid.UUID,
+    with_vectors: bool = False,
+) -> Point:
+    """One point, or `PointNotFoundError` if it is not this tenant's.
+
+    Raises rather than returning `None` so a route cannot forget the check and
+    serve `200 null`. "Absent" and "another tenant's" are the same outcome by
+    design (ADR-0016: cross-tenant access is `404`, never `403`).
+    """
+    point = await repository.get(tenant_id=tenant_id, point_id=point_id, with_vectors=with_vectors)
+    if point is None:
+        raise PointNotFoundError(f"point {point_id} not found")
+    return point
+
+
+async def list_file_points(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    file_id: uuid.UUID,
+    limit: int,
+    cursor: str | None = None,
+    include_inactive: bool = False,
+) -> PointPage:
+    """One file's points in display (`order_id`) order.
+
+    An unknown or foreign `file_id` yields an empty page rather than an error:
+    the two are indistinguishable to the caller, which is the same
+    non-disclosure property `get_point` gets from raising.
+    """
+    return await repository.list_by_file(
+        tenant_id=tenant_id,
+        file_id=file_id,
+        limit=limit,
+        cursor=cursor,
+        include_inactive=include_inactive,
+    )
+
+
+async def count_points(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    domain: str | None = None,
+    file_id: uuid.UUID | None = None,
+    include_inactive: bool = False,
+) -> int:
+    return await repository.count(
+        tenant_id=tenant_id,
+        domain=domain,
+        file_id=file_id,
+        include_inactive=include_inactive,
+    )
+
+
+async def search_points(
+    repository: PointRepository,
+    *,
+    tenant_id: uuid.UUID,
+    query: str,
+    limit: int,
+    cursor: str | None = None,
+    domain: str | None = None,
+    file_id: uuid.UUID | None = None,
+    include_inactive: bool = False,
+) -> PointPage:
+    """Keyword match on `content`, within this tenant.
+
+    **Not semantic retrieval.** Qdrant's full-text index filters rather than
+    scores, so results carry no relevance ranking and their order is
+    unspecified. Ranked retrieval is ADR-0003's hybrid path in plan 003; this
+    function must not grow a semantic mode (ADR-0002).
+    """
+    return await repository.keyword_search(
+        tenant_id=tenant_id,
+        query=normalize_persian_text(query),
+        limit=limit,
+        cursor=cursor,
+        domain=domain,
+        file_id=file_id,
+        include_inactive=include_inactive,
+    )

src/application/points/relinking.py

@@ -0,0 +1,135 @@
+"""Adjacency-pointer maintenance for a point leaving a file's sequence.
+
+ADR-0001 keeps `previous_chunk_id`/`next_chunk_id` on every point so ADR-0003's
+context-window expansion can walk a file in O(1) steps. ADR-0002 makes keeping
+them correct an obligation of every operation that changes a point's position:
+a partial relink is a defect, not a degraded-but-acceptable outcome.
+
+The function below is the primitive that obligation reduces to. It is pure, and
+it is written as **"what is still missing between the state I just read and the
+state I want"** rather than "the patches a delete implies". That framing is what
+makes the caller's retry loop correct: re-planning after a partial apply emits
+exactly the patches that did not land, and re-planning after a completed delete
+emits nothing at all. The three cases the plan calls out — a normal delete, a
+second delete of an already-inactive point, and recovery from a half-applied
+batch — are then one code path instead of three.
+
+Note what is deliberately *not* patched: the departing point's own
+`previous_chunk_id`/`next_chunk_id`. Nothing active points at it once its
+neighbours are relinked, so those pointers are unreachable rather than stale,
+and leaving them records where the point sat — which is what a later restore or
+an audit reader would need. `src/application/points/deletion.py` relies on that
+when it re-plans: the departing point's pointers are the only surviving record
+of which two neighbours have to be joined.
+"""
+
+import uuid
+from collections.abc import Mapping
+from datetime import datetime
+
+from src.application.points.point import Point
+from src.application.ports.point_repository import PayloadPatch
+
+
+def _optional_id(value: uuid.UUID | None) -> str | None:
+    return str(value) if value is not None else None
+
+
+def _provenance(point: Point, *, actor: str, now: datetime) -> dict[str, object]:
+    """The fields every mutation writes: who, when, and the next version.
+
+    Bumping `version` on a relinked *neighbour* is intentional. The neighbour's
+    payload really did change, so a concurrent editor holding the old version
+    must get a `409` rather than overwrite the pointer we just fixed.
+    """
+    return {
+        "updated_at": now.isoformat(),
+        "updated_by": actor,
+        "version": point.version + 1,
+    }
+
+
+def neighbour_ids(point: Point) -> tuple[uuid.UUID, ...]:
+    """The ids `patches_for_removal` needs loaded, skipping the nulls."""
+    return tuple(
+        pointer for pointer in (point.previous_chunk_id, point.next_chunk_id) if pointer is not None
+    )
+
+
+def patches_for_removal(
+    point: Point,
+    neighbours: Mapping[uuid.UUID, Point],
+    *,
+    actor: str,
+    now: datetime,
+) -> tuple[PayloadPatch, ...]:
+    """The patches still needed to remove `point` from its file's sequence.
+
+    Returns an empty tuple when the removal is already complete, which the
+    caller reads as both "converged" and "this was a no-op".
+
+    A neighbour absent from `neighbours` is skipped rather than patched blind:
+    its id came from the departing point's payload, so a missing one means the
+    chain was already broken, and inventing a patch for a point that is not
+    there would not fix it. The caller logs that case.
+    """
+    patches: list[PayloadPatch] = []
+
+    if point.is_active:
+        patches.append(
+            PayloadPatch(
+                point_id=point.point_id,
+                payload={
+                    "is_active": False,
+                    "deleted_at": now.isoformat(),
+                    **_provenance(point, actor=actor, now=now),
+                },
+                expected_version=point.version,
+            )
+        )
+
+    previous = neighbours.get(point.previous_chunk_id) if point.previous_chunk_id else None
+    if previous is not None and previous.next_chunk_id != point.next_chunk_id:
+        patches.append(
+            PayloadPatch(
+                point_id=previous.point_id,
+                payload={
+                    "next_chunk_id": _optional_id(point.next_chunk_id),
+                    **_provenance(previous, actor=actor, now=now),
+                },
+                expected_version=previous.version,
+            )
+        )
+
+    following = neighbours.get(point.next_chunk_id) if point.next_chunk_id else None
+    if following is not None and following.previous_chunk_id != point.previous_chunk_id:
+        patches.append(
+            PayloadPatch(
+                point_id=following.point_id,
+                payload={
+                    "previous_chunk_id": _optional_id(point.previous_chunk_id),
+                    **_provenance(following, actor=actor, now=now),
+                },
+                expected_version=following.version,
+            )
+        )
+
+    return tuple(patches)
+
+
+def patch_for_deactivation(point: Point, *, actor: str, now: datetime) -> PayloadPatch:
+    """Deactivate one point without touching any pointer.
+
+    Used by the whole-file sweep, where every point in the file leaves at once:
+    no active point survives to dangle, so there is no neighbour to relink and
+    the chain stays intact for a later reader of the deactivated file.
+    """
+    return PayloadPatch(
+        point_id=point.point_id,
+        payload={
+            "is_active": False,
+            "deleted_at": now.isoformat(),
+            **_provenance(point, actor=actor, now=now),
+        },
+        expected_version=point.version,
+    )

src/application/ports/__init__.py

@@ -0,0 +1,7 @@
+"""Narrow contracts for external side effects (ADR-0015).
+
+Ports exist for external side effects/persistence that need a swappable or
+fake-able boundary — not as a blanket wrapper around every database access.
+`object_storage.py` is one: MinIO is a real external system with its own
+failure modes, and ADR-0016 requires a hand-written fake for it in tests.
+"""

src/application/ports/embedding.py

@@ -0,0 +1,63 @@
+"""Embedding ports (ADR-0001, ADR-0017).
+
+`src/infrastructure/embedding/` holds the production adapters; tests use
+scripted fakes (ADR-0016). Application code depends on these Protocols, not
+on `httpx`/provider SDKs directly.
+"""
+
+from collections.abc import Sequence
+from typing import Protocol
+
+from src.application.ingestion.models import SparseVector
+
+
+class DenseEmbedder(Protocol):
+    """One named dense vector's embedding client (`dense_nomic`/`dense_openai`).
+
+    `embed_batch` is a single batched network call — callers own concurrency
+    bounding (ADR-0017's `embed_concurrency` semaphore), not this Protocol.
+    """
+
+    name: str
+    model_version: str
+    """Identifies the model that produced these vectors (ADR-0001).
+
+    Written into every point's `embedding_model_version` payload field, which
+    exists so a future model swap can tell which chunks need re-embedding. The
+    embedder is what knows this, so it is reported here rather than
+    reconstructed from configuration at the call site.
+    """
+
+    async def embed_batch(self, texts: Sequence[str]) -> list[list[float]]:
+        """Return one vector per input text, same order. Raises `EmbedderError`
+        (see `src/application/ingestion/errors.py`) on transport/response
+        failure.
+        """
+        ...
+
+
+class SparseEmbedder(Protocol):
+    """The `sparse` (BM25) vector's embedding client.
+
+    Blocking/CPU-bound (ADR-0017): callers offload it via
+    `anyio.to_thread.run_sync` with the ingestion `CapacityLimiter`, not call
+    it directly from an `async def`.
+    """
+
+    name: str
+    model_version: str
+    """Identifies the analyzer/parameters that produced these vectors.
+
+    Same purpose as `DenseEmbedder.model_version`; for BM25 the "model" is the
+    analyzer choice (ADR-0005), which is equally a re-embedding trigger.
+    """
+
+    def embed_batch(self, texts: Sequence[str], *, query: bool = False) -> list[SparseVector]:
+        """Return one sparse vector per input text, same order.
+
+        `query=True` selects the query-side weighting, which omits document
+        length normalization. Ingestion always passes `False`; the flag exists
+        so retrieval (ADR-0003) encodes queries through this same port rather
+        than growing a second, silently divergent implementation.
+        """
+        ...

src/application/ports/object_storage.py

@@ -0,0 +1,14 @@
+"""The object-storage port (ADR-0013).
+
+`src/infrastructure/minio/storage.py` is the production adapter; tests use a
+hand-written fake (ADR-0016). Application code depends on this Protocol, not
+on the `minio` SDK.
+"""
+
+from typing import Protocol
+
+
+class ObjectStorage(Protocol):
+    async def put_object(self, *, key: str, data: bytes, content_type: str) -> None:
+        """Store `data` privately under `key`. Overwrites an existing object."""
+        ...

src/application/ports/point_repository.py

@@ -0,0 +1,131 @@
+"""The point read/edit port for `/v1/points` (ADR-0002, ADR-0015).
+
+Separate from `PointStorage`, which stays exactly the two bulk operations
+ingestion performs. Reads, single-point edits, reordering, and keyword search
+have a different caller, a different failure vocabulary, and a different
+tenant-filter obligation, so they get their own port rather than accreting onto
+the ingestion one.
+
+`tenant_id` is a required keyword argument on **every** method. That is not
+style: ADR-0002's isolation rule has to hold on every code path that touches the
+collection, and an optional tenant filter is one forgotten argument away from a
+cross-tenant read. Making it required moves that from a review question to a
+type error.
+
+`src/infrastructure/qdrant/point_repository.py` is the production adapter;
+`tests.fakes.FakePointRepository` is the test double.
+"""
+
+import uuid
+from collections.abc import Sequence
+from typing import Protocol
+
+from pydantic import BaseModel
+
+from src.application.points.point import Point
+
+
+class PointPage(BaseModel):
+    """One page of points plus the cursor that continues it.
+
+    `next_cursor` is opaque to callers and encoded by the adapter: ordered
+    scrolls and keyword searches paginate by different Qdrant mechanisms, and
+    neither is a plain integer offset. `None` means the listing is exhausted.
+    """
+
+    points: tuple[Point, ...]
+    next_cursor: str | None = None
+
+
+class PayloadPatch(BaseModel):
+    """Set these payload fields on one point, optionally guarded by `version`.
+
+    When `expected_version` is set, the adapter attaches it to the operation's
+    filter, so a concurrent write that has already moved the version on means
+    this patch matches nothing rather than clobbering it. The guard is what
+    makes a lost update impossible; detecting that it fired is the service's
+    job (see `apply_patches`).
+    """
+
+    point_id: uuid.UUID
+    payload: dict[str, object]
+    expected_version: int | None = None
+
+
+class PointRepository(Protocol):
+    async def get(
+        self, *, tenant_id: uuid.UUID, point_id: uuid.UUID, with_vectors: bool = False
+    ) -> Point | None:
+        """One point, or `None` if it does not exist *under this tenant*.
+
+        The two cases are deliberately indistinguishable — the route maps both
+        to `404` so a caller cannot probe for another tenant's point ids.
+        """
+        ...
+
+    async def get_many(
+        self, *, tenant_id: uuid.UUID, point_ids: Sequence[uuid.UUID]
+    ) -> tuple[Point, ...]:
+        """The subset of `point_ids` that exists under this tenant.
+
+        Order is not guaranteed and missing ids are silently absent: callers are
+        neighbour-relinking and batch precondition checks, both of which match
+        on id rather than position.
+        """
+        ...
+
+    async def list_by_file(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        file_id: uuid.UUID,
+        limit: int,
+        cursor: str | None = None,
+        include_inactive: bool = False,
+    ) -> PointPage:
+        """One file's points in `order_id` order (ADR-0008's `scroll`).
+
+        Scoped to a single file because the cursor is an `order_id` value, and
+        `order_id` is only unique within a file.
+        """
+        ...
+
+    async def count(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> int: ...
+
+    async def keyword_search(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        query: str,
+        limit: int,
+        cursor: str | None = None,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> PointPage:
+        """Full-text payload match on `content`, plus structured filters.
+
+        Keyword matching, **not** semantic retrieval (ADR-0002). Qdrant's
+        full-text index is a filter, not a scorer, so results carry no relevance
+        ranking and their order is unspecified.
+        """
+        ...
+
+    async def apply_patches(self, *, tenant_id: uuid.UUID, patches: Sequence[PayloadPatch]) -> None:
+        """Apply every patch in one Qdrant `points/batch` request.
+
+        Qdrant has no multi-point transaction, so this is not atomic and does
+        not pretend to be. ADR-0002's all-or-nothing rule is implemented one
+        layer up as validate-every-precondition-then-apply; the per-patch
+        `expected_version` guard here is what makes the residual window safe,
+        turning a lost update into a no-op the service can detect rather than a
+        silent clobber.
+        """
+        ...

src/application/ports/point_storage.py

@@ -0,0 +1,45 @@
+"""The point-storage port (ADR-0001, ADR-0015).
+
+`src/infrastructure/qdrant/points.py` is the production adapter; tests use a
+hand-written fake (ADR-0016). Application code depends on this Protocol, not on
+the `qdrant_client` SDK.
+
+Deliberately narrow: exactly the two operations ingestion performs. Reads,
+single-point edits, reordering, and keyword search are plan 002's `/v1/points`
+surface and belong on a port of their own rather than accreting here.
+"""
+
+import uuid
+from collections.abc import Sequence
+from datetime import datetime
+from typing import Protocol
+
+from src.application.points.models import ChunkPoint
+
+
+class PointStorage(Protocol):
+    async def upsert_points(self, points: Sequence[ChunkPoint]) -> None:
+        """Upsert one batch of points.
+
+        Callers own batching and concurrency bounding (ADR-0017's
+        `upsert_concurrency` semaphore), not this Protocol — the same division
+        `DenseEmbedder.embed_batch` uses.
+        """
+        ...
+
+    async def deactivate_points_from_index(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        file_id: uuid.UUID,
+        from_chunk_index: int,
+        deleted_at: datetime,
+        updated_by: str,
+    ) -> int:
+        """Soft-delete this file's points at or past `from_chunk_index`.
+
+        Sets `is_active=false`/`deleted_at` rather than removing the points
+        (ADR-0002: delete is soft by default). Tenant-filtered — a `file_id`
+        alone is never sufficient authority. Returns how many points matched.
+        """
+        ...

src/application/tenants/__init__.py

@@ -0,0 +1,9 @@
+"""Operator-run tenant provisioning (ADR-0008, ADR-0009)."""
+
+from src.application.tenants.provisioning import (
+    DEFAULT_SCOPES,
+    ProvisionResult,
+    provision_tenant,
+)
+
+__all__ = ["DEFAULT_SCOPES", "ProvisionResult", "provision_tenant"]

src/application/tenants/provisioning.py

@@ -0,0 +1,129 @@
+"""Provision a tenant, its first API key, and its domains (ADR-0008, ADR-0009).
+
+Nothing in the HTTP surface can bootstrap a tenant: every `/v1` route needs a
+key, and a key can only exist once a tenant does. That chicken-and-egg is why
+this is an operator-run deployment step (`src/cli/provision_tenant.py`) rather
+than an endpoint — the same reasoning that keeps `alembic upgrade head` and
+`qdrant_bootstrap` off the request path.
+
+This is the package's only caller-facing entry point. It owns the whole
+composition — tenant reuse-or-create, key generation and hashing, domain
+registration, and the single transaction the three share — so a caller cannot
+get the order wrong or commit a key whose tenant never landed (CLAUDE.md,
+"prefer deep modules"). The plaintext key is returned exactly once and is never
+logged (ADR-0011 forbids plaintext keys in logs); only its non-secret
+`key_prefix` appears in the event.
+"""
+
+import uuid
+from dataclasses import dataclass
+
+import structlog
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.application.auth.keys import generate_api_key, hash_secret
+from src.infrastructure.postgres.repositories import api_keys as api_keys_repo
+from src.infrastructure.postgres.repositories import tenant_domains as domains_repo
+from src.infrastructure.postgres.repositories import tenants as tenants_repo
+
+logger = structlog.get_logger(__name__)
+
+DEFAULT_SCOPES = (
+    "files:write",
+    "domains:read",
+    "domains:write",
+    "points:read",
+    "points:write",
+)
+
+
+@dataclass(frozen=True)
+class ProvisionResult:
+    tenant_id: uuid.UUID
+    tenant_slug: str
+    tenant_created: bool
+    api_key_id: uuid.UUID
+    api_key_prefix: str
+    api_key: str
+    """The plaintext bearer token. Only ever returned here — never stored, never logged."""
+    domains_created: tuple[str, ...]
+    domains_existing: tuple[str, ...]
+
+
+async def provision_tenant(
+    sessionmaker: async_sessionmaker[AsyncSession],
+    *,
+    slug: str,
+    name: str | None = None,
+    key_name: str = "bootstrap",
+    scopes: tuple[str, ...] = DEFAULT_SCOPES,
+    domains: tuple[str, ...] = (),
+    actor_type: str = "backend",
+) -> ProvisionResult:
+    """Create (or reuse) the tenant, issue a key, and register `domains`.
+
+    Re-running with the same `slug` reuses the tenant and its existing domains
+    rather than failing, so an operator can add a key to a live tenant with the
+    same command they used to create it. A *new* key is issued on every run —
+    keys are write-once by construction (only the hash is stored), so there is
+    nothing to return for an existing one.
+    """
+    key_prefix, secret, full_key = generate_api_key()
+
+    async with sessionmaker() as session:
+        tenant = await tenants_repo.get_by_slug(session, slug)
+        tenant_created = tenant is None
+        if tenant is None:
+            tenant = tenants_repo.create(session, slug=slug, name=name or slug)
+            # `api_keys.tenant_id` and `tenant_domains.tenant_id` FK to this row
+            # and the mapped classes carry no ORM relationship for the unit of
+            # work to order by itself, so the insert has to land first.
+            await session.flush()
+
+        api_key = api_keys_repo.create(
+            session,
+            tenant_id=tenant.id,
+            name=key_name,
+            key_prefix=key_prefix,
+            key_hash=hash_secret(secret),
+            scopes=list(scopes),
+            actor_type=actor_type,
+            created_by="cli:provision_tenant",
+        )
+
+        created: list[str] = []
+        existing: list[str] = []
+        for domain in domains:
+            if await domains_repo.get(session, tenant_id=tenant.id, domain=domain) is not None:
+                existing.append(domain)
+                continue
+            domains_repo.create(session, tenant_id=tenant.id, domain=domain, display_name=domain)
+            created.append(domain)
+
+        await session.flush()
+        tenant_id, api_key_id, tenant_slug = tenant.id, api_key.id, tenant.slug
+        await session.commit()
+
+    if tenant_created:
+        logger.info("tenant.provisioned", tenant_id=str(tenant_id), tenant_slug=tenant_slug)
+    for domain in created:
+        logger.info("domain.created", tenant_id=str(tenant_id), domain=domain)
+    logger.info(
+        "api_key.provisioned",
+        tenant_id=str(tenant_id),
+        api_key_id=str(api_key_id),
+        key_prefix=key_prefix,
+        scopes=list(scopes),
+        actor_type=actor_type,
+    )
+
+    return ProvisionResult(
+        tenant_id=tenant_id,
+        tenant_slug=tenant_slug,
+        tenant_created=tenant_created,
+        api_key_id=api_key_id,
+        api_key_prefix=key_prefix,
+        api_key=full_key,
+        domains_created=tuple(created),
+        domains_existing=tuple(existing),
+    )

src/bootstrap/dependencies.py

@@ -1,11 +1,16 @@
-from collections.abc import AsyncIterator
+from collections.abc import AsyncIterator, Sequence
 from dataclasses import dataclass
 
+from anyio import CapacityLimiter, Semaphore
 from fastapi import Request
 from minio import Minio
 from qdrant_client import AsyncQdrantClient
 from sqlalchemy.ext.asyncio import AsyncEngine, AsyncSession, async_sessionmaker
 
+from src.application.ports.embedding import DenseEmbedder, SparseEmbedder
+from src.application.ports.object_storage import ObjectStorage
+from src.application.ports.point_repository import PointRepository
+from src.application.ports.point_storage import PointStorage
 from src.config import Settings
 
 
@@ -16,6 +21,13 @@ class AppResources:
     db_sessionmaker: async_sessionmaker[AsyncSession]
     minio_client: Minio
     qdrant_client: AsyncQdrantClient
+    object_storage: ObjectStorage
+    point_storage: PointStorage
+    point_repository: PointRepository
+    ingestion_limiter: CapacityLimiter
+    dense_embedders: Sequence[DenseEmbedder]
+    sparse_embedder: SparseEmbedder
+    ingestion_concurrency_limiter: Semaphore
 
 
 def _resources(request: Request) -> AppResources:
@@ -34,6 +46,45 @@ def get_qdrant_client(request: Request) -> AsyncQdrantClient:
     return _resources(request).qdrant_client
 
 
+def get_object_storage(request: Request) -> ObjectStorage:
+    return _resources(request).object_storage
+
+
+def get_point_storage(request: Request) -> PointStorage:
+    return _resources(request).point_storage
+
+
+def get_point_repository(request: Request) -> PointRepository:
+    return _resources(request).point_repository
+
+
+def get_ingestion_limiter(request: Request) -> CapacityLimiter:
+    return _resources(request).ingestion_limiter
+
+
+def get_dense_embedders(request: Request) -> Sequence[DenseEmbedder]:
+    return _resources(request).dense_embedders
+
+
+def get_sparse_embedder(request: Request) -> SparseEmbedder:
+    return _resources(request).sparse_embedder
+
+
+def get_ingestion_concurrency_limiter(request: Request) -> Semaphore:
+    return _resources(request).ingestion_concurrency_limiter
+
+
+def get_sessionmaker(request: Request) -> async_sessionmaker[AsyncSession]:
+    """The session *factory*, not a request-scoped session.
+
+    Application services that own more than one transaction in a single
+    request (ADR-0017's two-phase upload) need to open and close sessions
+    themselves rather than borrow one request-scoped session that would
+    otherwise stay open across the whole request.
+    """
+    return _resources(request).db_sessionmaker
+
+
 async def get_db_session(request: Request) -> AsyncIterator[AsyncSession]:
     sessionmaker = _resources(request).db_sessionmaker
     async with sessionmaker() as session:

src/bootstrap/lifespan.py

@@ -1,28 +1,66 @@
-from collections.abc import AsyncIterator, Callable
+from collections.abc import AsyncIterator, Callable, Sequence
 from contextlib import AbstractAsyncContextManager, asynccontextmanager
 
+import httpx
 import structlog
-from anyio import to_thread
+from anyio import CapacityLimiter, Semaphore, to_thread
 from fastapi import FastAPI
 
 from src.application.ingestion import get_encoder
+from src.application.ports.embedding import DenseEmbedder
 from src.bootstrap.dependencies import AppResources
 from src.config import Settings
+from src.infrastructure.embedding.bm25 import Bm25SparseEmbedder
+from src.infrastructure.embedding.openai_compatible import (
+    OpenAICompatibleEmbedder,
+    is_ollama_base_url,
+)
 from src.infrastructure.minio.client import create_client as create_minio_client
+from src.infrastructure.minio.storage import MinioObjectStorage
 from src.infrastructure.observability.logging import configure_logging
 from src.infrastructure.postgres.database import create_engine, create_sessionmaker
 from src.infrastructure.qdrant.client import create_client as create_qdrant_client
+from src.infrastructure.qdrant.point_repository import QdrantPointRepository
+from src.infrastructure.qdrant.points import QdrantPointStorage
 
 logger = structlog.get_logger(__name__)
 
 
+def _auth_headers(api_key: str | None) -> dict[str, str]:
+    """Bearer header, or none at all when no key is configured.
+
+    Sending an empty `Bearer ` is worse than sending nothing: some gateways
+    treat a malformed credential as an auth failure rather than as anonymous.
+    """
+    return {"Authorization": f"Bearer {api_key}"} if api_key else {}
+
+
+async def _warm_dense_embedders(embedders: Sequence[DenseEmbedder]) -> None:
+    """Force each dense model to load before the first upload needs it.
+
+    Same rationale as the tiktoken warm-up above, but with the opposite
+    failure policy. A self-hosted embedder that has unloaded the model takes
+    minutes to serve its first request — longer than
+    `INGESTION_TIMEOUT_SECONDS` — so paying that once at boot keeps it off a
+    user's upload. Unlike the tokenizer this is best-effort: an embedder that
+    is merely *down* must not stop the process from booting and reporting its
+    own health, and `/readyz` is where that condition belongs.
+    """
+    for embedder in embedders:
+        try:
+            await embedder.embed_batch(["warmup"])
+            logger.info("lifespan.embedder.warmed", embedder=embedder.name)
+        except Exception:
+            logger.warning("lifespan.embedder.warm_failed", embedder=embedder.name, exc_info=True)
+
+
 def create_lifespan(
     settings: Settings | None = None,
 ) -> Callable[[FastAPI], AbstractAsyncContextManager[None, bool | None]]:
     @asynccontextmanager
     async def lifespan(app: FastAPI) -> AsyncIterator[None]:
         resolved_settings = settings or Settings()
-        configure_logging(resolved_settings.logging)
+        configure_logging(resolved_settings.logging, resolved_settings.app)
 
         # tiktoken fetches its vocabulary over the network on first use, so warm
         # it here: a missing vocabulary should fail the process at boot, not the
@@ -41,14 +79,81 @@ def create_lifespan(
         logger.info("lifespan.minio.client.created")
 
         qdrant_client = create_qdrant_client(resolved_settings.qdrant)
+        # No collection DDL here: `ensure_chunks_collection` is a deployment
+        # step (`python -m src.cli.qdrant_bootstrap`), for the same reason
+        # ADR-0009 keeps Alembic out of startup and ADR-0012 makes LangGraph's
+        # `.setup()` a deployment step.
+        point_storage = QdrantPointStorage(
+            qdrant_client, collection=resolved_settings.qdrant.collection
+        )
+        point_repository = QdrantPointRepository(
+            qdrant_client, collection=resolved_settings.qdrant.collection
+        )
         logger.info("lifespan.qdrant.client.created")
 
+        nomic_settings = resolved_settings.embedding.nomic
+        nomic_http_client = httpx.AsyncClient(
+            base_url=nomic_settings.base_url,
+            timeout=nomic_settings.timeout_seconds,
+            headers=_auth_headers(nomic_settings.api_key),
+        )
+        openai_settings = resolved_settings.embedding.openai
+        openai_http_client = httpx.AsyncClient(
+            base_url=openai_settings.base_url,
+            timeout=openai_settings.timeout_seconds,
+            headers=_auth_headers(openai_settings.api_key),
+        )
+        dense_embedders = (
+            OpenAICompatibleEmbedder(
+                nomic_http_client,
+                name="dense_nomic",
+                model=nomic_settings.model,
+                document_prefix=nomic_settings.document_prefix,
+                keep_alive=(
+                    nomic_settings.keep_alive
+                    if is_ollama_base_url(nomic_settings.base_url)
+                    else None
+                ),
+            ),
+            OpenAICompatibleEmbedder(
+                openai_http_client,
+                name="dense_openai",
+                model=openai_settings.model,
+                dimensions=openai_settings.dimensions,
+                document_prefix=openai_settings.document_prefix,
+            ),
+        )
+        sparse_embedder = Bm25SparseEmbedder(resolved_settings.embedding.sparse)
+        logger.info("lifespan.embedders.created")
+
+        await _warm_dense_embedders(dense_embedders)
+
+        # Bounds how many ingestions run in this process at once (ADR-0017);
+        # a distinct resource from ingestion_limiter, which bounds threads
+        # spent on blocking work within a single ingestion.
+        ingestion_concurrency_limiter = Semaphore(resolved_settings.ingestion.max_concurrency)
+
+        # Bounds threads spent on blocking ingestion work (parsing, chunking,
+        # hashing, the sync minio SDK) so it cannot exhaust Starlette's own
+        # thread pool (ADR-0017).
+        ingestion_limiter = CapacityLimiter(resolved_settings.ingestion.thread_pool_size)
+        object_storage = MinioObjectStorage(
+            minio_client, bucket=resolved_settings.minio.bucket, limiter=ingestion_limiter
+        )
+
         app.state.resources = AppResources(
             settings=resolved_settings,
             db_engine=db_engine,
             db_sessionmaker=db_sessionmaker,
             minio_client=minio_client,
             qdrant_client=qdrant_client,
+            object_storage=object_storage,
+            point_storage=point_storage,
+            point_repository=point_repository,
+            ingestion_limiter=ingestion_limiter,
+            dense_embedders=dense_embedders,
+            sparse_embedder=sparse_embedder,
+            ingestion_concurrency_limiter=ingestion_concurrency_limiter,
         )
 
         try:
@@ -64,4 +169,14 @@ def create_lifespan(
             except Exception:
                 logger.exception("lifespan.qdrant.close.failed")
 
+            try:
+                await nomic_http_client.aclose()
+            except Exception:
+                logger.exception("lifespan.embedding.nomic_client.close.failed")
+
+            try:
+                await openai_http_client.aclose()
+            except Exception:
+                logger.exception("lifespan.embedding.openai_client.close.failed")
+
     return lifespan

src/cli/__init__.py

@@ -0,0 +1 @@
+"""Operator entry points that run as deployment steps, not at app startup."""

src/cli/provision_tenant.py

@@ -0,0 +1,94 @@
+"""Create a tenant, issue its first API key, and register its domains.
+
+    uv run python -m src.cli.provision_tenant --slug acme --domain fire
+
+A deployment step, like `alembic upgrade head` and `src.cli.qdrant_bootstrap`.
+It exists because nothing over HTTP can bootstrap a tenant: every `/v1` route
+requires an API key, and a key cannot exist before its tenant does.
+
+The plaintext key is printed to **stdout once** and never stored or logged —
+Postgres holds only its SHA-256 hash (ADR-0009), so a lost key is reissued by
+re-running this command, not recovered. Structured logs go to stderr/the log
+sink and carry only the non-secret `key_prefix` (ADR-0011).
+
+Re-running with the same `--slug` reuses the tenant and any domains it already
+has, and issues an additional key.
+"""
+
+import argparse
+import asyncio
+import sys
+
+import structlog
+
+from src.application.tenants import DEFAULT_SCOPES, provision_tenant
+from src.config import Settings
+from src.infrastructure.observability.logging import configure_logging
+from src.infrastructure.postgres.database import create_engine, create_sessionmaker
+
+logger = structlog.get_logger(__name__)
+
+
+def _parse_args(argv: list[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        prog="python -m src.cli.provision_tenant",
+        description="Create a tenant, issue an API key, and register domains.",
+    )
+    parser.add_argument("--slug", required=True, help="URL-safe tenant identifier, e.g. 'acme'")
+    parser.add_argument("--name", default=None, help="Display name (defaults to --slug)")
+    parser.add_argument("--key-name", default="bootstrap", help="Label for the issued API key")
+    parser.add_argument(
+        "--scopes",
+        default=",".join(DEFAULT_SCOPES),
+        help=f"Comma-separated scopes for the key (default: {','.join(DEFAULT_SCOPES)})",
+    )
+    parser.add_argument(
+        "--domain",
+        action="append",
+        default=[],
+        dest="domains",
+        help="Domain to register; repeatable. Uploads reject an unregistered domain.",
+    )
+    return parser.parse_args(argv)
+
+
+async def run(argv: list[str] | None = None, settings: Settings | None = None) -> int:
+    args = _parse_args(argv)
+    resolved = settings or Settings()
+    configure_logging(resolved.logging, resolved.app)
+
+    scopes = tuple(scope.strip() for scope in args.scopes.split(",") if scope.strip())
+    if not scopes:
+        logger.error("tenant.provision.failed", reason="no_scopes")
+        return 2
+
+    engine = create_engine(resolved.postgres)
+    try:
+        result = await provision_tenant(
+            create_sessionmaker(engine),
+            slug=args.slug,
+            name=args.name,
+            key_name=args.key_name,
+            scopes=scopes,
+            domains=tuple(args.domains),
+        )
+    finally:
+        await engine.dispose()
+
+    # stdout, not the logger: this is the one value the operator must copy, and
+    # it must never reach a log sink (ADR-0011).
+    print(f"tenant_id={result.tenant_id}")
+    print(f"tenant_slug={result.tenant_slug}")
+    print(f"api_key_id={result.api_key_id}")
+    print(f"domains={','.join(result.domains_created + result.domains_existing)}")
+    print(f"api_key={result.api_key}")
+    print("Store the api_key now -- only its hash is persisted and it cannot be shown again.")
+    return 0
+
+
+def main() -> None:
+    sys.exit(asyncio.run(run()))
+
+
+if __name__ == "__main__":
+    main()

src/cli/qdrant_bootstrap.py

@@ -0,0 +1,54 @@
+"""Create the `chunks` collection — the Qdrant analogue of `alembic upgrade head`.
+
+    uv run python -m src.cli.qdrant_bootstrap
+
+A deployment step, deliberately not part of the FastAPI lifespan: collection
+creation is DDL, which ADR-0009 keeps out of application startup for Postgres
+and ADR-0012 keeps out of it for LangGraph's `.setup()`. See
+`src/infrastructure/qdrant/collection.py` for the full reasoning.
+
+Idempotent and safe to re-run. Exits non-zero if an existing collection
+diverges from the pinned schema, rather than leaving a silently degraded
+sparse index behind.
+"""
+
+import asyncio
+import sys
+
+import structlog
+
+from src.config import Settings
+from src.infrastructure.observability.logging import configure_logging
+from src.infrastructure.qdrant.client import create_client
+from src.infrastructure.qdrant.collection import (
+    CollectionSchemaMismatchError,
+    ensure_chunks_collection,
+)
+
+logger = structlog.get_logger(__name__)
+
+
+async def bootstrap(settings: Settings | None = None) -> int:
+    resolved = settings or Settings()
+    configure_logging(resolved.logging, resolved.app)
+    client = create_client(resolved.qdrant)
+    try:
+        created = await ensure_chunks_collection(client, collection=resolved.qdrant.collection)
+    except CollectionSchemaMismatchError as exc:
+        logger.error("qdrant.bootstrap.schema_mismatch", error=str(exc))
+        return 1
+    finally:
+        await client.close()
+
+    logger.info(
+        "qdrant.bootstrap.completed", collection=resolved.qdrant.collection, created=created
+    )
+    return 0
+
+
+def main() -> None:
+    sys.exit(asyncio.run(bootstrap()))
+
+
+if __name__ == "__main__":
+    main()

src/config.py

@@ -1,9 +1,31 @@
+from typing import Any
+
 from pydantic import Field, model_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
+# pydantic-settings does not cascade `env_file` from a parent `BaseSettings`
+# to a nested one: each `BaseSettings` subclass only reads `.env` if its own
+# `model_config` names it. `Settings` has no fields of its own -- only
+# nested settings objects -- so every nested class below repeats
+# `env_file=".env"` as its own default, or its fields would silently read
+# only real process environment variables (fine under Docker Compose, broken
+# for local `.env`-file development) while still *appearing* to work
+# whenever a `.env.example` default happens to match the class default.
+#
+# That default alone isn't enough to let a caller point `Settings` at a
+# *different* file (as the test suite does, to parse `.env.example`), since
+# passing `_env_file=...` to `Settings(...)` only overrides `Settings`'s own
+# `model_config` -- nested defaults still construct against their own
+# hardcoded ".env". `Settings.__init__`/`EmbeddingSettings.__init__` below
+# thread an explicit `_env_file` override through to every nested
+# constructor so one override actually reaches the whole tree.
+_UNSET: Any = object()
+
 
 class PostgresSettings(BaseSettings):
-    model_config = SettingsConfigDict(env_prefix="POSTGRES_", extra="ignore")
+    model_config = SettingsConfigDict(
+        env_prefix="POSTGRES_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     host: str = "127.0.0.1"
     port: int = 5433
@@ -17,7 +39,9 @@ class PostgresSettings(BaseSettings):
 
 
 class MinioSettings(BaseSettings):
-    model_config = SettingsConfigDict(env_prefix="MINIO_", extra="ignore")
+    model_config = SettingsConfigDict(
+        env_prefix="MINIO_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     endpoint: str = "127.0.0.1:9100"
     access_key: str = "chatbot"
@@ -33,15 +57,22 @@ class IngestionSettings(BaseSettings):
     timeouts, or callers give up on work that is still succeeding.
     """
 
-    model_config = SettingsConfigDict(env_prefix="INGESTION_", extra="ignore")
+    model_config = SettingsConfigDict(
+        env_prefix="INGESTION_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     max_concurrency: int = 4
     thread_pool_size: int = 8
     timeout_seconds: float = 120.0
+    max_upload_size_mb: int = 25
     max_chunks_per_file: int = 5000
     embed_batch_size: int = 128
     embed_concurrency: int = 4
 
+    @property
+    def max_upload_size_bytes(self) -> int:
+        return self.max_upload_size_mb * 1024 * 1024
+
 
 class ChunkingSettings(BaseSettings):
     """Parsing and chunking parameters (ADR-0018).
@@ -52,7 +83,9 @@ class ChunkingSettings(BaseSettings):
     `search_document: ` task prefix and any heading text carried into a chunk.
     """
 
-    model_config = SettingsConfigDict(env_prefix="CHUNKING_", extra="ignore")
+    model_config = SettingsConfigDict(
+        env_prefix="CHUNKING_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     strategy: str = "fixed_size"
     chunk_size: int = 400
@@ -76,25 +109,151 @@ class ChunkingSettings(BaseSettings):
 
 
 class QdrantSettings(BaseSettings):
-    model_config = SettingsConfigDict(env_prefix="QDRANT_", extra="ignore")
+    """Qdrant connection and bulk-upsert bounds (ADR-0001, ADR-0017).
+
+    `collection` names the single shared collection all tenants live in
+    (ADR-0001); it is deliberately configurable so tests can point at a
+    disposable one. The vector *dimensions* are not settings -- they are model
+    facts pinned in `src/infrastructure/qdrant/collection.py`, and changing one
+    is a re-embedding migration.
+
+    `upsert_batch_size` sits inside ADR-0001's 64-256 bulk-upload band, and
+    `upsert_concurrency` bounds in-flight batches so ingestion issues parallel
+    streams rather than an unbounded `gather` (ADR-0017).
+    """
+
+    model_config = SettingsConfigDict(
+        env_prefix="QDRANT_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     url: str = "http://127.0.0.1:6343"
     api_key: str | None = None
+    collection: str = "chunks"
+    upsert_batch_size: int = 128
+    upsert_concurrency: int = 4
+
+
+class NomicEmbeddingSettings(BaseSettings):
+    """Self-hosted `nomic-embed-text-v2-moe` server (ADR-0001, 768-dim).
+
+    Talks the OpenAI-compatible `/embeddings` endpoint shape. The default
+    points at the Ollama host the `emet` benchmark used, whose OpenAI-compat
+    shim accepts any non-empty API key.
+
+    `document_prefix` is empty by design. ADR-0004 cites the model card's
+    `search_document: ` requirement, but emet's winning run used no prefix
+    (Ollama's template is a bare passthrough and injects none), and the
+    prefix is not cosmetic -- it moves the vector substantially. Turning it
+    on here obliges the query side to send `search_query: ` too (ADR-0003),
+    so it stays off until an emet run measures the pair together.
+    """
+
+    model_config = SettingsConfigDict(
+        env_prefix="EMBEDDING_NOMIC_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
+
+    base_url: str = "http://192.168.10.10:11435/v1"
+    model: str = "nomic-embed-text-v2-moe"
+    api_key: str = "sk-not-set"
+    document_prefix: str = ""
+    keep_alive: str = "30m"
+    timeout_seconds: float = 30.0
+
+
+class OpenaiEmbeddingSettings(BaseSettings):
+    """OpenAI's hosted embedding API (ADR-0001's second dense signal).
+
+    `dimensions` is unset, so `text-embedding-3-large` returns its native
+    3072 dimensions -- the configuration emet benchmarked. Setting it would
+    truncate via Matryoshka and is a re-embedding migration, not a config
+    tweak.
+    """
+
+    model_config = SettingsConfigDict(
+        env_prefix="EMBEDDING_OPENAI_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
+
+    base_url: str = "https://api.openai.com/v1"
+    model: str = "text-embedding-3-large"
+    api_key: str | None = None
+    dimensions: int | None = None
+    document_prefix: str = ""
+    timeout_seconds: float = 30.0
+
+
+class SparseEmbeddingSettings(BaseSettings):
+    """BM25 sparse-vector parameters (ADR-0001, ADR-0005).
+
+    `analyzer`, `k`, and `b` are the configuration emet benchmarked as
+    `bm25-fa-norm-stop`; changing them invalidates that result. `k`/`b`
+    saturation is applied client-side here, while IDF comes from Qdrant's
+    `modifier="idf"` sparse-vector config at query time.
+
+    `avg_len` is emet's placeholder average document length in analyzer
+    tokens, exposed as a setting so it can be recalibrated from real corpus
+    statistics without a code change.
+    """
+
+    model_config = SettingsConfigDict(
+        env_prefix="EMBEDDING_SPARSE_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
+
+    analyzer: str = "fa_norm_stop"
+    k: float = 1.2
+    b: float = 0.75
+    avg_len: float = 256.0
+
+
+class EmbeddingSettings(BaseSettings):
+    model_config = SettingsConfigDict(extra="ignore")
+
+    nomic: NomicEmbeddingSettings = Field(default_factory=NomicEmbeddingSettings)
+    openai: OpenaiEmbeddingSettings = Field(default_factory=OpenaiEmbeddingSettings)
+    sparse: SparseEmbeddingSettings = Field(default_factory=SparseEmbeddingSettings)
+
+    def __init__(self, _env_file: Any = _UNSET, **data: Any) -> None:
+        env_file = ".env" if _env_file is _UNSET else _env_file
+        data.setdefault("nomic", NomicEmbeddingSettings(_env_file=env_file))
+        data.setdefault("openai", OpenaiEmbeddingSettings(_env_file=env_file))
+        data.setdefault("sparse", SparseEmbeddingSettings(_env_file=env_file))
+        super().__init__(**data)
 
 
 class AppLimitSettings(BaseSettings):
-    model_config = SettingsConfigDict(env_prefix="APP_", extra="ignore")
+    model_config = SettingsConfigDict(
+        env_prefix="APP_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     env: str = "local"
-    max_upload_size_mb: int = 25
     readiness_check_timeout_seconds: float = 2.0
+    # The deployed commit SHA or release tag (ADR-0011, "Bind process-level
+    # environment context"). Set by CI/CD at build/deploy time -- never
+    # computed at runtime by shelling out to git, which would fail in a
+    # container image with no .git directory.
+    service_version: str = "dev"
 
 
 class LoggingSettings(BaseSettings):
-    model_config = SettingsConfigDict(env_prefix="LOG_", extra="ignore")
+    """Logging sinks (ADR-0011).
+
+    `json_format` controls stdout's renderer only. Production sets it `true`
+    so stdout is JSON for the container log collector; local development
+    leaves it `false` for a colored console renderer. `file_path`, when set,
+    is a second, independent handler that always renders JSON regardless of
+    `json_format` -- a developer can read a human console while still keeping
+    a machine-parseable file. Unset in production: stdout/stderr collection is
+    preferred there over a log file inside the container.
+    """
+
+    model_config = SettingsConfigDict(
+        env_prefix="LOG_", extra="ignore", env_file=".env", env_ignore_empty=True
+    )
 
     level: str = "INFO"
     json_format: bool = False
+    file_path: str | None = None
+    file_max_bytes: int = 10 * 1024 * 1024
+    file_backup_count: int = 5
 
 
 class Settings(BaseSettings):
@@ -105,5 +264,18 @@ class Settings(BaseSettings):
     ingestion: IngestionSettings = Field(default_factory=IngestionSettings)
     chunking: ChunkingSettings = Field(default_factory=ChunkingSettings)
     qdrant: QdrantSettings = Field(default_factory=QdrantSettings)
+    embedding: EmbeddingSettings = Field(default_factory=EmbeddingSettings)
     app: AppLimitSettings = Field(default_factory=AppLimitSettings)
     logging: LoggingSettings = Field(default_factory=LoggingSettings)
+
+    def __init__(self, _env_file: Any = _UNSET, **data: Any) -> None:
+        env_file = ".env" if _env_file is _UNSET else _env_file
+        data.setdefault("postgres", PostgresSettings(_env_file=env_file))
+        data.setdefault("minio", MinioSettings(_env_file=env_file))
+        data.setdefault("ingestion", IngestionSettings(_env_file=env_file))
+        data.setdefault("chunking", ChunkingSettings(_env_file=env_file))
+        data.setdefault("qdrant", QdrantSettings(_env_file=env_file))
+        data.setdefault("embedding", EmbeddingSettings(_env_file=env_file))
+        data.setdefault("app", AppLimitSettings(_env_file=env_file))
+        data.setdefault("logging", LoggingSettings(_env_file=env_file))
+        super().__init__(_env_file=env_file, **data)

src/infrastructure/embedding/analyzers.py

@@ -0,0 +1,150 @@
+"""The `fa_norm_stop` BM25 analyzer (ADR-0001, ADR-0005).
+
+Ported from the `emet` evaluation lab
+(`src/chatbot_gh/adapters/sparse/analyzers.py`), which benchmarked four Farsi
+analyzer variants on the real corpus and found `fa_norm_stop` the best
+performer. This is a *measured* artifact: changing the normalization,
+tokenization, or stopword list invalidates that result, so improvements belong
+in a new emet benchmark run rather than in an edit here.
+
+Deliberately independent of `src/application/ingestion/normalization.py`.
+Those solve different problems: `normalize_persian_text` shapes chunk content
+that gets cited back to the reader, so ADR-0018 has it preserve digits and
+punctuation as authored. This module shapes index terms nobody ever sees, so
+it folds digits and diacritics freely. Sharing one function between them would
+let a display-motivated tweak silently perturb the benchmarked sparse index.
+"""
+
+import re
+import unicodedata
+
+# Several Arabic letterforms are visually indistinguishable from Latin ones in
+# a monospace editor (alef from "l", heh from "o"), and literals render
+# right-to-left, visually reordering the source line. `normalization.py` writes
+# them as codepoints for that reason; this module follows the same convention.
+_ZWNJ = 0x200C
+_ARABIC_YEH = 0x064A
+_ARABIC_KAF = 0x0643
+_TEH_MARBUTA = 0x0629
+_HAMZA_ON_WAW = 0x0624
+_ALEF_HAMZA_BELOW = 0x0625
+_ALEF_HAMZA_ABOVE = 0x0623
+
+_PERSIAN_YEH = 0x06CC
+_PERSIAN_KEHEH = 0x06A9
+_HEH = 0x0647
+_WAW = 0x0648
+_ALEF = 0x0627
+_SPACE = 0x0020
+
+# Persian (U+06F0-U+06F9) and Arabic-Indic (U+0660-U+0669) digits both fold to
+# ASCII, so the same number matches however it was authored.
+_EASTERN_DIGITS = str.maketrans("۰۱۲۳۴۵۶۷۸۹٠١٢٣٤٥٦٧٨٩", "01234567890123456789")
+
+# ZWNJ becomes a space (splitting compounds into separate terms) and the
+# Arabic letterforms fold to their Persian equivalents. Every entry is a
+# single codepoint mapping to a single codepoint over disjoint sources, so
+# applying them in one pass is equivalent to emet's chained `str.replace`
+# calls -- provided NFC runs first, since NFC is what composes the hamza
+# forms this table then folds.
+_FOLDING: dict[int, int] = {
+    _ZWNJ: _SPACE,
+    _ARABIC_YEH: _PERSIAN_YEH,
+    _ARABIC_KAF: _PERSIAN_KEHEH,
+    _TEH_MARBUTA: _HEH,
+    _HAMZA_ON_WAW: _WAW,
+    _ALEF_HAMZA_BELOW: _ALEF,
+    _ALEF_HAMZA_ABOVE: _ALEF,
+}
+_FOLDING.update(_EASTERN_DIGITS)
+
+# Common Persian/Arabic stopwords (function words + FAQ noise), plus the
+# English function words that appear in a mixed-script corpus. Kept small and
+# explicit -- deliberately not a full hazm list. `_HEH_ALEF` is the plural
+# suffix "ha"; written as a codepoint pair because both of its letters are
+# Latin-confusable, which is exactly the case Ruff's RUF001 flags.
+_HEH_ALEF = chr(_HEH) + chr(_ALEF)
+
+_PERSIAN_STOPWORDS: frozenset[str] = frozenset(
+    {
+        "و",
+        "در",
+        "به",
+        "از",
+        "که",
+        "این",
+        "را",
+        "با",
+        "برای",
+        "آن",
+        "یک",
+        "است",
+        "شد",
+        "شده",
+        "می",
+        "های",
+        _HEH_ALEF,
+        "یا",
+        "تا",
+        "بر",
+        "اگر",
+        "هم",
+        "نیز",
+        "ولی",
+        "اما",
+        "چه",
+        "چون",
+        "روی",
+        "پس",
+        "پیش",
+        "هر",
+        "هیچ",
+        "بود",
+        "باشد",
+        "هست",
+        "نیست",
+        "کند",
+        "کرد",
+        "کردن",
+        "شود",
+        "the",
+        "a",
+        "an",
+        "of",
+        "to",
+        "and",
+        "in",
+        "on",
+        "for",
+        "is",
+        "are",
+    }
+)
+
+# Word characters minus underscore. Note this KEEPS digits: an insurance
+# corpus is full of policy numbers, dates, and amounts, and those are exactly
+# the tokens a lexical index should be able to match on.
+_TOKEN_RE = re.compile(r"[^\W_]+", re.UNICODE)
+
+FA_NORM_STOP = "fa_norm_stop"
+
+
+def _normalize_fa(text: str) -> str:
+    return unicodedata.normalize("NFC", text).translate(_FOLDING)
+
+
+def _tokenize_raw(text: str) -> list[str]:
+    return [m.group(0).lower() for m in _TOKEN_RE.finditer(text)]
+
+
+def analyze(text: str, analyzer: str = FA_NORM_STOP) -> list[str]:
+    """Tokenize `text` into sparse-index terms.
+
+    Only `fa_norm_stop` is implemented -- emet's other three variants
+    (`raw`, `fa_norm`, `fa_norm_stem`) lost the benchmark and exist there as
+    experiment arms, not as configurations this service should run.
+    """
+    if analyzer != FA_NORM_STOP:
+        raise ValueError(f"Unknown analyzer '{analyzer}'")
+    tokens = _tokenize_raw(_normalize_fa(text))
+    return [token for token in tokens if token not in _PERSIAN_STOPWORDS]

src/infrastructure/embedding/bm25.py

@@ -0,0 +1,93 @@
+"""The `bm25-fa-norm-stop` sparse embedder (ADR-0001, ADR-0005).
+
+Ported from the `emet` evaluation lab
+(`src/chatbot_gh/adapters/sparse/bm25_embedder.py`), the configuration that
+won its Farsi analyzer benchmark. Not Qdrant's hosted `Qdrant/bm25` FastEmbed
+model, whose documented language support omits Farsi (ADR-0005).
+
+**The BM25 work is split across two systems.** This adapter applies the
+term-frequency saturation half client-side -- the `k` and `b` parameters,
+including document-length normalization. IDF is *not* computed here: Qdrant
+supplies it from collection-wide statistics when the sparse vector field is
+created with `modifier="idf"`.
+
+That split is load-bearing. A Qdrant collection created without
+`modifier="idf"` will silently score these vectors as saturated term
+frequencies with no IDF weighting at all -- no error, just materially worse
+lexical retrieval. The collection bootstrap (plan 001 Phase 5) must set it.
+"""
+
+from collections import Counter
+from collections.abc import Sequence
+from hashlib import blake2b
+
+from src.application.ingestion.models import SparseVector
+from src.config import SparseEmbeddingSettings
+from src.infrastructure.embedding.analyzers import analyze
+
+# Qdrant sparse indices must be non-negative and fit a signed 32-bit int.
+_INDEX_SPACE = 2**31 - 1
+
+
+def _token_index(token: str) -> int:
+    """Map a term to its sparse-vector index.
+
+    A hash rather than a vocabulary table, so the mapping needs no shared
+    state and stays identical across processes, restarts, and — critically —
+    between ingest-time and query-time encoding. `blake2b` rather than
+    `hash()`, which is PYTHONHASHSEED-salted and therefore differs per
+    process.
+    """
+    digest = blake2b(token.encode("utf-8"), digest_size=8).digest()
+    return int.from_bytes(digest, "big") % _INDEX_SPACE
+
+
+def text_to_sparse_vector(
+    text: str, *, settings: SparseEmbeddingSettings, query: bool = False
+) -> SparseVector:
+    """Encode one text as a BM25-saturated sparse vector (IDF applied by Qdrant).
+
+    Document and query sides differ in exactly one term: documents carry the
+    `b` length normalization, queries do not (standard BM25 practice — a
+    query's own length should not discount its terms).
+    """
+    tokens = analyze(text, settings.analyzer)
+    if not tokens:
+        return SparseVector(indices=[], values=[])
+
+    frequencies = Counter(tokens)
+    doc_length = float(len(tokens))
+    k = settings.k
+    b = settings.b
+
+    indices: list[int] = []
+    values: list[float] = []
+    # Sorted so the emitted vector is deterministic for a given text, which
+    # keeps re-ingestion byte-stable and makes the output testable.
+    for token, freq in sorted(frequencies.items()):
+        if query:
+            weight = freq * (k + 1.0) / (freq + k)
+        else:
+            weight = freq * (k + 1.0) / (freq + k * (1.0 - b + b * doc_length / settings.avg_len))
+        indices.append(_token_index(token))
+        values.append(float(weight))
+
+    return SparseVector(indices=indices, values=values)
+
+
+class Bm25SparseEmbedder:
+    """A `SparseEmbedder` (see `src/application/ports/embedding.py`).
+
+    Pure CPU work with no network calls, so it is blocking: callers offload it
+    via `anyio.to_thread.run_sync` with the ingestion `CapacityLimiter`
+    (ADR-0017), never awaiting it directly on the event loop.
+    """
+
+    name = "sparse"
+
+    def __init__(self, settings: SparseEmbeddingSettings) -> None:
+        self.model_version = f"bm25-{settings.analyzer}"
+        self._settings = settings
+
+    def embed_batch(self, texts: Sequence[str], *, query: bool = False) -> list[SparseVector]:
+        return [text_to_sparse_vector(text, settings=self._settings, query=query) for text in texts]

src/infrastructure/embedding/openai_compatible.py

@@ -0,0 +1,82 @@
+"""Dense embedding adapter for OpenAI-compatible `/embeddings` endpoints.
+
+Backs both `dense_nomic` (self-hosted `nomic-embed-text-v2-moe` behind
+Ollama's OpenAI-compatible shim) and `dense_openai` (OpenAI's hosted API) —
+both speak the same request/response shape, so one adapter serves both named
+vectors with different config (ADR-0001).
+
+Uses `httpx` directly rather than the `openai` SDK. The `emet` benchmark this
+configuration comes from uses the SDK, but it is a synchronous batch tool;
+ADR-0017 requires async, semaphore-bounded batches here, and the request shape
+is small enough that the SDK earns nothing.
+
+`httpx.AsyncClient` is application-lifetime (ADR-0012): built once in the
+FastAPI lifespan and passed in, never constructed per call.
+"""
+
+from collections.abc import Sequence
+from urllib.parse import urlparse
+
+import httpx
+
+# Ollama's default port, plus the alternate the benchmarked deployment uses.
+_OLLAMA_PORTS = frozenset({11434, 11435})
+
+
+def is_ollama_base_url(base_url: str) -> bool:
+    """Whether `base_url` looks like an Ollama OpenAI-compatible endpoint.
+
+    Ollama unloads an idle model, and reloading `nomic-embed-text-v2-moe`
+    costs well over two minutes — longer than `INGESTION_TIMEOUT_SECONDS`, so
+    a cold upload would 504. `keep_alive` is how the model is kept resident,
+    and it is an Ollama extension, hence the sniffing.
+    """
+    parsed = urlparse(base_url)
+    host = parsed.hostname or ""
+    return parsed.port in _OLLAMA_PORTS or "ollama" in host.lower()
+
+
+class OpenAICompatibleEmbedder:
+    """A `DenseEmbedder` (see `src/application/ports/embedding.py`) over one
+    OpenAI-compatible `/embeddings` endpoint.
+    """
+
+    def __init__(
+        self,
+        client: httpx.AsyncClient,
+        *,
+        name: str,
+        model: str,
+        dimensions: int | None = None,
+        document_prefix: str = "",
+        keep_alive: str | None = None,
+    ) -> None:
+        self.name = name
+        self.model_version = model
+        self._client = client
+        self._model = model
+        self._dimensions = dimensions
+        self._document_prefix = document_prefix
+        self._keep_alive = keep_alive
+
+    async def embed_batch(self, texts: Sequence[str]) -> list[list[float]]:
+        inputs = (
+            [f"{self._document_prefix}{text}" for text in texts]
+            if self._document_prefix
+            else list(texts)
+        )
+        payload: dict[str, object] = {"model": self._model, "input": inputs}
+        if self._dimensions is not None:
+            payload["dimensions"] = self._dimensions
+        if self._keep_alive is not None:
+            payload["keep_alive"] = self._keep_alive
+
+        response = await self._client.post("/embeddings", json=payload)
+        response.raise_for_status()
+        body = response.json()
+
+        # Sort by `index` rather than trusting response order: the contract
+        # guarantees the field, not the ordering, and a silently permuted
+        # batch would attach every vector to the wrong chunk.
+        data = sorted(body["data"], key=lambda item: item["index"])
+        return [item["embedding"] for item in data]

src/infrastructure/minio/storage.py

@@ -0,0 +1,33 @@
+"""MinIO adapter for the `ObjectStorage` port (ADR-0013, ADR-0017).
+
+The `minio` SDK is synchronous, so every call runs through
+`anyio.to_thread.run_sync` bounded by the ingestion `CapacityLimiter` — the
+same rule ADR-0017 applies to parsing/chunking. Calling the SDK directly from
+`async def` would block every concurrent request in the process.
+"""
+
+import io
+from functools import partial
+
+from anyio import CapacityLimiter, to_thread
+from minio import Minio
+
+
+class MinioObjectStorage:
+    def __init__(self, client: Minio, *, bucket: str, limiter: CapacityLimiter) -> None:
+        self._client = client
+        self._bucket = bucket
+        self._limiter = limiter
+
+    async def put_object(self, *, key: str, data: bytes, content_type: str) -> None:
+        await to_thread.run_sync(
+            partial(
+                self._client.put_object,
+                self._bucket,
+                key,
+                io.BytesIO(data),
+                length=len(data),
+                content_type=content_type,
+            ),
+            limiter=self._limiter,
+        )

src/infrastructure/observability/logging.py

@@ -1,14 +1,50 @@
+"""Logging configuration: structlog + stdlib, dual local sinks (ADR-0011).
+
+Console and an optional file are independent, simultaneous handlers on the
+same logger, not a single renderer chosen by a flag -- the same structlog
+event fans out to both. The console handler is always human-readable
+(`ConsoleRenderer`); the file handler, when enabled via `LOG_FILE_PATH`,
+always renders JSON regardless of `LOG_JSON_FORMAT`, so a saved log stays
+machine-parseable even when the terminal next to it is not.
+
+`LOG_JSON_FORMAT` controls *stdout's* renderer only: production sets it `true`
+so the container log collector gets JSON; local development leaves it `false`
+for the colored console. `LOG_FILE_PATH` is expected to be unset in
+production -- stdout/stderr collection is preferred there over a log file
+inside the container.
+"""
+
 import logging
 import logging.config
 import sys
+from collections.abc import Callable
 
 import structlog
 
-from src.config import LoggingSettings
+from src.config import AppLimitSettings, LoggingSettings
 
 
-def configure_logging(settings: LoggingSettings) -> None:
+def _bind_environment(settings: AppLimitSettings) -> Callable[..., dict[str, object]]:
+    """A static processor, not a contextvar: `env`/`service_version` don't
+    vary per request, and a contextvar bound before the first request would
+    be wiped by `RequestIdMiddleware`'s `clear_contextvars()` on that request.
+    Closing over `settings` at configure time makes every event carry them
+    instead, regardless of request context (ADR-0011).
+    """
+
+    def processor(
+        logger: object, method_name: str, event_dict: dict[str, object]
+    ) -> dict[str, object]:
+        event_dict["env"] = settings.env
+        event_dict["service_version"] = settings.service_version
+        return event_dict
+
+    return processor
+
+
+def configure_logging(settings: LoggingSettings, app_settings: AppLimitSettings) -> None:
     shared_processors = [
+        _bind_environment(app_settings),
         structlog.contextvars.merge_contextvars,
         structlog.stdlib.add_log_level,
         structlog.stdlib.add_logger_name,
@@ -27,55 +63,84 @@ def configure_logging(settings: LoggingSettings) -> None:
         cache_logger_on_first_use=True,
     )
 
-    renderer = (
+    console_renderer = (
         structlog.processors.JSONRenderer()
         if settings.json_format
         else structlog.dev.ConsoleRenderer(colors=True)
     )
 
-    logging.config.dictConfig(
-        {
-            "version": 1,
-            "disable_existing_loggers": False,
-            "formatters": {
-                "default": {
+    formatters = {
+        "console": {
             "()": structlog.stdlib.ProcessorFormatter,
             "processors": [
                 structlog.stdlib.ProcessorFormatter.remove_processors_meta,
-                        renderer,
+                console_renderer,
             ],
             "foreign_pre_chain": [
                 structlog.stdlib.ExtraAdder(),
                 *shared_processors,
             ],
         },
-            },
-            "handlers": {
+    }
+    handlers: dict[str, dict[str, object]] = {
         "console": {
             "class": "logging.StreamHandler",
             "level": settings.level,
-                    "formatter": "default",
+            "formatter": "console",
             "stream": sys.stdout,
         },
-            },
+    }
+    root_handlers = ["console"]
+
+    if settings.file_path is not None:
+        # File handler always renders JSON, independent of the console
+        # renderer chosen above -- a saved log stays machine-parseable even
+        # when stdout is the colored, human-readable renderer.
+        formatters["file"] = {
+            "()": structlog.stdlib.ProcessorFormatter,
+            "processors": [
+                structlog.stdlib.ProcessorFormatter.remove_processors_meta,
+                structlog.processors.JSONRenderer(),
+            ],
+            "foreign_pre_chain": [
+                structlog.stdlib.ExtraAdder(),
+                *shared_processors,
+            ],
+        }
+        handlers["file"] = {
+            "class": "logging.handlers.RotatingFileHandler",
+            "level": settings.level,
+            "formatter": "file",
+            "filename": settings.file_path,
+            "maxBytes": settings.file_max_bytes,
+            "backupCount": settings.file_backup_count,
+        }
+        root_handlers.append("file")
+
+    logging.config.dictConfig(
+        {
+            "version": 1,
+            "disable_existing_loggers": False,
+            "formatters": formatters,
+            "handlers": handlers,
             "loggers": {
                 "": {
-                    "handlers": ["console"],
+                    "handlers": root_handlers,
                     "level": settings.level,
                     "propagate": False,
                 },
                 "uvicorn": {
-                    "handlers": ["console"],
+                    "handlers": root_handlers,
                     "level": settings.level,
                     "propagate": False,
                 },
                 "uvicorn.access": {
-                    "handlers": ["console"],
+                    "handlers": root_handlers,
                     "level": settings.level,
                     "propagate": False,
                 },
                 "sqlalchemy.engine": {
-                    "handlers": ["console"],
+                    "handlers": root_handlers,
                     "level": "WARNING",
                     "propagate": False,
                 },

src/infrastructure/postgres/models/__init__.py

@@ -4,6 +4,7 @@ from src.infrastructure.postgres.models.ingestion_job import IngestionJob
 from src.infrastructure.postgres.models.ingestion_job_event import IngestionJobEvent
 from src.infrastructure.postgres.models.source_file import SourceFile
 from src.infrastructure.postgres.models.tenant import Tenant
+from src.infrastructure.postgres.models.tenant_domain import TenantDomain
 
 __all__ = [
     "ApiKey",
@@ -12,4 +13,5 @@ __all__ = [
     "IngestionJobEvent",
     "SourceFile",
     "Tenant",
+    "TenantDomain",
 ]

src/infrastructure/postgres/models/tenant_domain.py

@@ -0,0 +1,52 @@
+import uuid
+from datetime import datetime
+
+from sqlalchemy import CheckConstraint, DateTime, ForeignKey, String, UniqueConstraint, func
+from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy.orm import Mapped, mapped_column
+
+from src.infrastructure.postgres.models.base import Base
+
+TENANT_DOMAIN_STATUSES = ("active", "disabled")
+
+
+class TenantDomain(Base):
+    """A domain a tenant is allowed to ingest into (ADR-0009).
+
+    Tenants do not share a domain list — one may run 14 insurance lines and
+    another 6 — so this is a per-tenant table rather than an enum or a global
+    lookup.
+
+    Its purpose is to stop an arbitrary caller-supplied `domain` from silently
+    creating a new Qdrant partition. `domain` is denormalized into every point's
+    payload and into `source_files`, and a typo like `fier` for `fire` produces
+    no error anywhere: the file indexes into a partition retrieval never queries,
+    so it is invisible rather than failed.
+
+    `domain` is the immutable key. Renaming it would mean rewriting every point
+    payload that carries it, which is a migration, not an edit — `display_name`
+    is the mutable human-facing label instead.
+    """
+
+    __tablename__ = "tenant_domains"
+    __table_args__ = (
+        UniqueConstraint("tenant_id", "domain", name="uq_tenant_domains_tenant_id_domain"),
+        CheckConstraint(f"status IN {TENANT_DOMAIN_STATUSES}", name="ck_tenant_domains_status"),
+    )
+
+    id: Mapped[uuid.UUID] = mapped_column(primary_key=True)
+    tenant_id: Mapped[uuid.UUID] = mapped_column(
+        ForeignKey("tenants.id", ondelete="CASCADE"), index=True
+    )
+    domain: Mapped[str] = mapped_column(String(80))
+    display_name: Mapped[str] = mapped_column(String(200))
+    status: Mapped[str] = mapped_column(String(20), default="active", server_default="active")
+    metadata_: Mapped[dict[str, object]] = mapped_column(
+        "metadata", JSONB, default=dict, server_default="{}"
+    )
+
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now(), onupdate=func.now()
+    )
+    disabled_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), default=None)

src/infrastructure/postgres/repositories/api_keys.py

@@ -0,0 +1,45 @@
+"""API-key lookups and issuance (ADR-0008, ADR-0009).
+
+Plain functions over an `AsyncSession` the caller owns. No function here
+commits, rolls back, or closes the session (ADR-0012). Secret comparison
+happens in `src/application/auth`, not here — this module only fetches rows
+by their non-secret `key_prefix`.
+"""
+
+import uuid
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.infrastructure.postgres.models.api_key import ApiKey
+
+
+async def get_by_prefix(session: AsyncSession, key_prefix: str) -> ApiKey | None:
+    result = await session.execute(select(ApiKey).where(ApiKey.key_prefix == key_prefix))
+    return result.scalar_one_or_none()
+
+
+def create(
+    session: AsyncSession,
+    *,
+    tenant_id: uuid.UUID,
+    name: str,
+    key_prefix: str,
+    key_hash: str,
+    scopes: list[str],
+    actor_type: str = "backend",
+    created_by: str | None = None,
+) -> ApiKey:
+    """Persist an issued key. The caller hashes the secret; this never sees it."""
+    api_key = ApiKey(
+        id=uuid.uuid4(),
+        tenant_id=tenant_id,
+        name=name,
+        key_prefix=key_prefix,
+        key_hash=key_hash,
+        scopes=scopes,
+        actor_type=actor_type,
+        created_by=created_by,
+    )
+    session.add(api_key)
+    return api_key

src/infrastructure/postgres/repositories/ingestion_jobs.py

@@ -0,0 +1,119 @@
+"""`ingestion_jobs`/`ingestion_job_events` persistence (ADR-0009, ADR-0017).
+
+Plain functions over an `AsyncSession` the caller owns. No function here
+commits, rolls back, or closes the session (ADR-0012) — the two-transaction
+shape in `src/application/files/upload.py` depends on that. Every read is
+tenant-scoped by a required `tenant_id` argument.
+"""
+
+import uuid
+from datetime import UTC, datetime
+
+from sqlalchemy import desc, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.infrastructure.postgres.models.ingestion_job import IngestionJob
+from src.infrastructure.postgres.models.ingestion_job_event import IngestionJobEvent
+
+
+async def get_by_id(
+    session: AsyncSession, *, tenant_id: uuid.UUID, ingestion_job_id: uuid.UUID
+) -> IngestionJob | None:
+    result = await session.execute(
+        select(IngestionJob).where(
+            IngestionJob.id == ingestion_job_id, IngestionJob.tenant_id == tenant_id
+        )
+    )
+    return result.scalar_one_or_none()
+
+
+async def get_latest_for_source_file(
+    session: AsyncSession, *, tenant_id: uuid.UUID, source_file_id: uuid.UUID
+) -> IngestionJob | None:
+    result = await session.execute(
+        select(IngestionJob)
+        .where(
+            IngestionJob.tenant_id == tenant_id,
+            IngestionJob.source_file_id == source_file_id,
+        )
+        .order_by(desc(IngestionJob.created_at))
+        .limit(1)
+    )
+    return result.scalar_one_or_none()
+
+
+def create_running(
+    session: AsyncSession,
+    *,
+    tenant_id: uuid.UUID,
+    source_file_id: uuid.UUID,
+    requested_by_api_key_id: uuid.UUID | None,
+    chunking_strategy: str,
+) -> IngestionJob:
+    job = IngestionJob(
+        id=uuid.uuid4(),
+        tenant_id=tenant_id,
+        source_file_id=source_file_id,
+        requested_by_api_key_id=requested_by_api_key_id,
+        status="running",
+        chunking_strategy=chunking_strategy,
+        started_at=datetime.now(UTC),
+    )
+    session.add(job)
+    return job
+
+
+async def mark_terminal(
+    session: AsyncSession,
+    *,
+    tenant_id: uuid.UUID,
+    ingestion_job_id: uuid.UUID,
+    status: str,
+    points_created: int = 0,
+    points_updated: int = 0,
+    points_soft_deleted: int = 0,
+    points_skipped: int = 0,
+    error_code: str | None = None,
+    error_message: str | None = None,
+) -> IngestionJob | None:
+    """Move a job from `running` to a terminal status.
+
+    Never reads/writes a job whose current status is already terminal — a
+    terminal job must not transition back to `running` or to a different
+    terminal status (ADR-0017).
+    """
+    job = await get_by_id(session, tenant_id=tenant_id, ingestion_job_id=ingestion_job_id)
+    if job is None or job.status != "running":
+        return None
+    job.status = status
+    job.completed_at = datetime.now(UTC)
+    job.points_created = points_created
+    job.points_updated = points_updated
+    job.points_soft_deleted = points_soft_deleted
+    job.points_skipped = points_skipped
+    job.error_code = error_code
+    job.error_message = error_message
+    return job
+
+
+def append_event(
+    session: AsyncSession,
+    *,
+    tenant_id: uuid.UUID,
+    ingestion_job_id: uuid.UUID,
+    level: str,
+    stage: str,
+    message: str,
+    details: dict[str, object] | None = None,
+) -> IngestionJobEvent:
+    event = IngestionJobEvent(
+        id=uuid.uuid4(),
+        tenant_id=tenant_id,
+        ingestion_job_id=ingestion_job_id,
+        level=level,
+        stage=stage,
+        message=message,
+        details=details or {},
+    )
+    session.add(event)
+    return event

src/infrastructure/postgres/repositories/source_files.py

@@ -0,0 +1,86 @@
+"""`source_files` persistence (ADR-0009).
+
+Plain functions over an `AsyncSession` the caller owns. No function here
+commits, rolls back, or closes the session (ADR-0012). Every read is
+tenant-scoped by a required `tenant_id` argument, so a missing filter is a
+signature error rather than a cross-tenant leak.
+"""
+
+import uuid
+from datetime import datetime
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.infrastructure.postgres.models.source_file import SourceFile
+
+
+async def get_by_id(
+    session: AsyncSession, *, tenant_id: uuid.UUID, source_file_id: uuid.UUID
+) -> SourceFile | None:
+    result = await session.execute(
+        select(SourceFile).where(SourceFile.id == source_file_id, SourceFile.tenant_id == tenant_id)
+    )
+    return result.scalar_one_or_none()
+
+
+async def find_active_by_content_hash(
+    session: AsyncSession, *, tenant_id: uuid.UUID, domain: str, content_sha256: str
+) -> SourceFile | None:
+    result = await session.execute(
+        select(SourceFile).where(
+            SourceFile.tenant_id == tenant_id,
+            SourceFile.domain == domain,
+            SourceFile.content_sha256 == content_sha256,
+            SourceFile.status == "active",
+        )
+    )
+    return result.scalar_one_or_none()
+
+
+def create(
+    session: AsyncSession,
+    *,
+    source_file_id: uuid.UUID,
+    tenant_id: uuid.UUID,
+    domain: str,
+    source_filename: str,
+    source_type: str,
+    content_sha256: str,
+    byte_size: int,
+    storage_uri: str,
+    created_by_api_key_id: uuid.UUID | None,
+) -> SourceFile:
+    """`source_file_id` is caller-generated: the upload service derives the
+    MinIO object key from it before this row exists, so the id has to be
+    chosen up front rather than assigned by the database.
+    """
+    source_file = SourceFile(
+        id=source_file_id,
+        tenant_id=tenant_id,
+        domain=domain,
+        source_filename=source_filename,
+        source_type=source_type,
+        content_sha256=content_sha256,
+        byte_size=byte_size,
+        storage_uri=storage_uri,
+        created_by_api_key_id=created_by_api_key_id,
+    )
+    session.add(source_file)
+    return source_file
+
+
+def mark_soft_deleted(source_file: SourceFile, *, deleted_at: datetime) -> None:
+    """Retire a file: `status='soft_deleted'` plus `deleted_at` (ADR-0009).
+
+    Takes the already-loaded row rather than an id, because the caller fetched
+    it under its tenant filter and re-fetching here would be a second place
+    that could forget that filter.
+
+    Retiring the row matters beyond bookkeeping: `find_active_by_content_hash`
+    matches only `active` files, so a re-upload of the same bytes after a delete
+    creates a fresh file and re-ingests it, instead of taking the duplicate path
+    and returning a file whose points have all been deactivated.
+    """
+    source_file.status = "soft_deleted"
+    source_file.deleted_at = deleted_at

src/infrastructure/postgres/repositories/tenant_domains.py

@@ -0,0 +1,73 @@
+"""`tenant_domains` persistence (ADR-0009).
+
+Plain functions over an `AsyncSession` the caller owns. No function here
+commits, rolls back, or closes the session (ADR-0012). Every read and write is
+tenant-scoped by a required `tenant_id` argument, so a missing filter is a
+signature error rather than a cross-tenant leak.
+"""
+
+import uuid
+from datetime import UTC, datetime
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.infrastructure.postgres.models.tenant_domain import TenantDomain
+
+
+async def get(session: AsyncSession, *, tenant_id: uuid.UUID, domain: str) -> TenantDomain | None:
+    result = await session.execute(
+        select(TenantDomain).where(
+            TenantDomain.tenant_id == tenant_id, TenantDomain.domain == domain
+        )
+    )
+    return result.scalar_one_or_none()
+
+
+async def list_for_tenant(
+    session: AsyncSession, *, tenant_id: uuid.UUID, include_disabled: bool = False
+) -> list[TenantDomain]:
+    statement = select(TenantDomain).where(TenantDomain.tenant_id == tenant_id)
+    if not include_disabled:
+        statement = statement.where(TenantDomain.status == "active")
+    result = await session.execute(statement.order_by(TenantDomain.domain))
+    return list(result.scalars().all())
+
+
+def create(
+    session: AsyncSession,
+    *,
+    tenant_id: uuid.UUID,
+    domain: str,
+    display_name: str,
+    metadata: dict[str, object] | None = None,
+) -> TenantDomain:
+    tenant_domain = TenantDomain(
+        id=uuid.uuid4(),
+        tenant_id=tenant_id,
+        domain=domain,
+        display_name=display_name,
+        metadata_=metadata or {},
+    )
+    session.add(tenant_domain)
+    return tenant_domain
+
+
+def update_display_name(tenant_domain: TenantDomain, *, display_name: str) -> TenantDomain:
+    """`domain` itself is deliberately not updatable.
+
+    It is denormalized into every Qdrant point payload and into `source_files`,
+    so changing the key would mean rewriting all of them — a migration, not an
+    edit. The label is what callers actually want to change.
+    """
+    tenant_domain.display_name = display_name
+    return tenant_domain
+
+
+def set_status(tenant_domain: TenantDomain, *, status: str) -> TenantDomain:
+    """Disable/re-enable a domain. Existing points are untouched either way —
+    disabling blocks new uploads, it is not a delete (ADR-0002).
+    """
+    tenant_domain.status = status
+    tenant_domain.disabled_at = datetime.now(UTC) if status == "disabled" else None
+    return tenant_domain

src/infrastructure/postgres/repositories/tenants.py

@@ -0,0 +1,27 @@
+"""Tenant lookups and creation (ADR-0009).
+
+Plain functions over an `AsyncSession` the caller owns. No function here
+commits, rolls back, or closes the session (ADR-0012).
+"""
+
+import uuid
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from src.infrastructure.postgres.models.tenant import Tenant
+
+
+async def get_by_id(session: AsyncSession, tenant_id: uuid.UUID) -> Tenant | None:
+    return await session.get(Tenant, tenant_id)
+
+
+async def get_by_slug(session: AsyncSession, slug: str) -> Tenant | None:
+    result = await session.execute(select(Tenant).where(Tenant.slug == slug))
+    return result.scalar_one_or_none()
+
+
+def create(session: AsyncSession, *, slug: str, name: str) -> Tenant:
+    tenant = Tenant(id=uuid.uuid4(), slug=slug, name=name)
+    session.add(tenant)
+    return tenant

src/infrastructure/qdrant/client.py

@@ -9,10 +9,22 @@ def create_client(settings: QdrantSettings) -> AsyncQdrantClient:
     return AsyncQdrantClient(url=settings.url, api_key=settings.api_key)
 
 
-async def ping(client: AsyncQdrantClient, timeout: float) -> bool:
+async def ping(client: AsyncQdrantClient, timeout: float, *, collection: str) -> bool:
+    """Whether Qdrant is reachable **and** the `chunks` collection exists.
+
+    Reachability alone is not readiness here. The collection is created by a
+    deployment step (`python -m src.cli.qdrant_bootstrap`, see ADR-0001
+    "Collection provisioning"), so a process can boot against a healthy Qdrant
+    that has no collection at all. Without this check that misconfiguration
+    stays invisible until the first upload fails with a `502` — after the
+    request has already paid for the MinIO write and the embedding round trips.
+
+    This is the Qdrant analogue of an unapplied Alembic migration, and it
+    belongs in `/readyz` for the same reason: it is a dependency-readiness
+    condition, not a process-health one.
+    """
     try:
         async with asyncio.timeout(timeout):
-            await client.get_collections()
+            return await client.collection_exists(collection)
     except Exception:
         return False
-    return True

src/infrastructure/qdrant/collection.py

@@ -0,0 +1,200 @@
+"""The `chunks` collection schema and its provisioning (ADR-0001).
+
+**This runs as a deployment step, never at FastAPI startup.** Creating a
+collection is DDL, and the project already rules DDL out of the request/boot
+path: ADR-0009 requires Alembic for Postgres schema, and ADR-0012 makes
+LangGraph's `.setup()` a deployment step. Doing it in the lifespan would also
+couple boot to Qdrant being reachable (that is `/readyz`'s job), race across
+replicas, and hide a misconfigured collection until traffic arrives.
+
+Entry point for operators: `uv run python -m src.cli.qdrant_bootstrap`.
+
+Two properties of this schema are load-bearing and fail *silently* if wrong,
+which is why `ensure_chunks_collection` verifies rather than skips:
+
+- **`sparse` must carry `modifier=IDF`.** `src/infrastructure/embedding/bm25.py`
+  computes only BM25's term-frequency saturation; Qdrant supplies IDF from
+  collection-wide statistics. Without the modifier there is no error and no
+  warning — lexical retrieval just quietly loses its IDF term (ADR-0005).
+- **The dense dimensions are pinned**: `dense_nomic` 768,
+  `dense_openai` 3072 (native, `dimensions` deliberately unset). They are
+  constants here rather than settings because they are model facts; changing
+  one is a re-embedding migration, not a config tweak (ADR-0001).
+
+All four named vectors are defined at creation even though `late_interaction`
+stays unpopulated until ADR-0003's rerank work (ADR-0017 does not compute it at
+ingest). Sparse and multivector fields cannot be added to an existing
+collection without recreating it, so deferring them is the one thing this
+schema cannot afford.
+"""
+
+import structlog
+from qdrant_client import AsyncQdrantClient, models
+
+logger = structlog.get_logger(__name__)
+
+DENSE_NOMIC_VECTOR = "dense_nomic"
+DENSE_OPENAI_VECTOR = "dense_openai"
+SPARSE_VECTOR = "sparse"
+LATE_INTERACTION_VECTOR = "late_interaction"
+
+DENSE_NOMIC_DIMENSIONS = 768
+DENSE_OPENAI_DIMENSIONS = 3072
+# jina-colbert-v2's per-token output dimension (ADR-0005).
+LATE_INTERACTION_DIMENSIONS = 128
+
+
+class CollectionSchemaMismatchError(RuntimeError):
+    """An existing collection does not match the schema this code expects.
+
+    Raised loudly instead of returning: silently accepting a collection whose
+    dense size or sparse modifier differs is precisely the failure this
+    explicit bootstrap step exists to prevent.
+    """
+
+
+def _vectors_config() -> dict[str, models.VectorParams]:
+    return {
+        DENSE_NOMIC_VECTOR: models.VectorParams(
+            size=DENSE_NOMIC_DIMENSIONS, distance=models.Distance.COSINE
+        ),
+        DENSE_OPENAI_VECTOR: models.VectorParams(
+            size=DENSE_OPENAI_DIMENSIONS, distance=models.Distance.COSINE
+        ),
+        # Rerank-only: never independently ANN-searched, so its HNSW graph is
+        # disabled (m=0), and stored on disk so its larger footprint does not
+        # degrade dense/sparse query latency (ADR-0001).
+        LATE_INTERACTION_VECTOR: models.VectorParams(
+            size=LATE_INTERACTION_DIMENSIONS,
+            distance=models.Distance.COSINE,
+            multivector_config=models.MultiVectorConfig(
+                comparator=models.MultiVectorComparator.MAX_SIM
+            ),
+            hnsw_config=models.HnswConfigDiff(m=0),
+            on_disk=True,
+        ),
+    }
+
+
+def _sparse_vectors_config() -> dict[str, models.SparseVectorParams]:
+    return {SPARSE_VECTOR: models.SparseVectorParams(modifier=models.Modifier.IDF)}
+
+
+# `content`'s full-text index backs ADR-0002's keyword search. The
+# `multilingual` tokenizer is the one that segments Persian correctly; `word`
+# splits on non-alphanumerics, which mis-handles ZWNJ-joined compounds. No
+# stemmer or stopword list is configured: `content` is already letter-folded by
+# `normalize_persian_text` at ingest (ADR-0018), and the *ranked* Farsi lexical
+# path is the benchmarked BM25 sparse vector, not this index. This one exists
+# for exact keyword/filter matching, which ADR-0002 keeps deliberately distinct
+# from retrieval.
+_CONTENT_INDEX = models.TextIndexParams(
+    type=models.TextIndexType.TEXT,
+    tokenizer=models.TokenizerType.MULTILINGUAL,
+    lowercase=True,
+)
+
+# (field name, schema). `order_id` is float because Qdrant's `Range` conditions
+# and `order_by` only support numeric/datetime payloads -- a keyword key could
+# only be sorted client-side after fetching every chunk (ADR-0001).
+_PayloadIndexSchema = models.PayloadSchemaType | models.KeywordIndexParams | models.TextIndexParams
+_PAYLOAD_INDEXES: tuple[tuple[str, _PayloadIndexSchema], ...] = (
+    # `is_tenant` co-locates a tenant's vectors on disk for sequential reads,
+    # which is the whole point of payload-partitioned multitenancy.
+    (
+        "tenant_id",
+        models.KeywordIndexParams(type=models.KeywordIndexType.KEYWORD, is_tenant=True),
+    ),
+    ("domain", models.PayloadSchemaType.KEYWORD),
+    ("file_id", models.PayloadSchemaType.KEYWORD),
+    ("order_id", models.PayloadSchemaType.FLOAT),
+    ("previous_chunk_id", models.PayloadSchemaType.KEYWORD),
+    ("next_chunk_id", models.PayloadSchemaType.KEYWORD),
+    ("content", _CONTENT_INDEX),
+    # Every read path filters on `is_active` (ADR-0002 implies `is_active: true`
+    # unless the caller opts in), and the re-ingestion sweep and `/v1/points`
+    # both range over `chunk_index`. Both were unindexed while ingestion was the
+    # only reader; plan 002 makes them hot.
+    ("is_active", models.PayloadSchemaType.BOOL),
+    ("chunk_index", models.PayloadSchemaType.INTEGER),
+)
+
+
+def _verify_existing(collection: str, info: models.CollectionInfo) -> None:
+    params = info.config.params
+    vectors = params.vectors
+    if not isinstance(vectors, dict):
+        raise CollectionSchemaMismatchError(
+            f"collection {collection!r} has an unnamed dense vector; ADR-0001 requires "
+            f"named vectors and this cannot be fixed without recreating the collection"
+        )
+
+    for name, expected_size in (
+        (DENSE_NOMIC_VECTOR, DENSE_NOMIC_DIMENSIONS),
+        (DENSE_OPENAI_VECTOR, DENSE_OPENAI_DIMENSIONS),
+        (LATE_INTERACTION_VECTOR, LATE_INTERACTION_DIMENSIONS),
+    ):
+        existing = vectors.get(name)
+        if existing is None:
+            raise CollectionSchemaMismatchError(
+                f"collection {collection!r} is missing the {name!r} vector"
+            )
+        if existing.size != expected_size:
+            raise CollectionSchemaMismatchError(
+                f"collection {collection!r} has {name!r} at {existing.size} dimensions, "
+                f"expected {expected_size}; re-dimensioning is a re-embedding migration"
+            )
+
+    sparse = (params.sparse_vectors or {}).get(SPARSE_VECTOR)
+    if sparse is None:
+        raise CollectionSchemaMismatchError(
+            f"collection {collection!r} is missing the {SPARSE_VECTOR!r} vector; sparse "
+            f"vectors cannot be added without recreating the collection"
+        )
+    if sparse.modifier != models.Modifier.IDF:
+        raise CollectionSchemaMismatchError(
+            f"collection {collection!r} has {SPARSE_VECTOR!r} with modifier "
+            f"{sparse.modifier!r}, expected 'idf'; without it Qdrant applies no IDF "
+            f"and lexical retrieval silently degrades (ADR-0005)"
+        )
+
+
+async def ensure_chunks_collection(client: AsyncQdrantClient, *, collection: str) -> bool:
+    """Create the `chunks` collection and its payload indexes if absent.
+
+    Idempotent: an existing collection is verified against the pinned schema
+    and left alone. Returns whether it created the collection.
+
+    Raises `CollectionSchemaMismatchError` if an existing collection diverges.
+    """
+    if await client.collection_exists(collection):
+        _verify_existing(collection, await client.get_collection(collection))
+        logger.info("qdrant.collection.verified", collection=collection)
+        # Payload indexes are additive and idempotent, so (re)creating them
+        # here is what lets an index be added to an already-live collection.
+        await _create_payload_indexes(client, collection=collection)
+        return False
+
+    await client.create_collection(
+        collection_name=collection,
+        vectors_config=_vectors_config(),
+        sparse_vectors_config=_sparse_vectors_config(),
+        # m=0 disables the global index; payload_m builds per-tenant graphs
+        # instead, per Qdrant's multitenant guidance (ADR-0001).
+        hnsw_config=models.HnswConfigDiff(m=0, payload_m=16),
+    )
+    logger.info("qdrant.collection.created", collection=collection)
+    await _create_payload_indexes(client, collection=collection)
+    return True
+
+
+async def _create_payload_indexes(client: AsyncQdrantClient, *, collection: str) -> None:
+    for field_name, field_schema in _PAYLOAD_INDEXES:
+        await client.create_payload_index(
+            collection_name=collection, field_name=field_name, field_schema=field_schema
+        )
+    logger.info(
+        "qdrant.collection.payload_indexes.ensured",
+        collection=collection,
+        fields=[name for name, _ in _PAYLOAD_INDEXES],
+    )

src/infrastructure/qdrant/point_repository.py

@@ -0,0 +1,247 @@
+"""Qdrant adapter for the `PointRepository` port (ADR-0002, ADR-0015).
+
+Every Qdrant filter for the `/v1/points` surface is built here. Routers and
+application services never import `qdrant_client` — they pass a `tenant_id` and
+get `Point` models back.
+
+Two mechanics are worth reading before changing anything:
+
+**Reads go through `scroll`, not `retrieve`.** `retrieve` fetches by id and
+takes no filter, which would force the tenant check to happen *after* Qdrant
+answered — exactly the "check it in Python afterwards" shape ADR-0002's
+isolation rule exists to prevent. `scroll` with a `HasIdCondition` plus the
+tenant condition pushes the check server-side, so a foreign id returns an empty
+page rather than a row this code has to remember to reject.
+
+**Ordered listing paginates by `order_id` value, not by offset.** Qdrant does
+not return a `next_page_offset` when `order_by` is set, and an offset-based
+cursor would skip or repeat rows when a concurrent insert shifts positions
+underneath the reader. Ranging on `order_id > cursor` is stable instead: a point
+inserted ahead of the cursor was already returned, and one inserted after it
+shows up on a later page. This relies on `order_id` being unique within a file,
+which ADR-0002 guarantees by rejecting a reorder whose gap would collapse onto a
+neighbour value.
+"""
+
+import uuid
+from collections.abc import Sequence
+
+from qdrant_client import AsyncQdrantClient, models
+
+from src.application.points.point import Point
+from src.application.ports.point_repository import PayloadPatch, PointPage
+
+# Qdrant's scroll returns `(records, next_page_offset)`; a record's id is a
+# `str | int` union in the SDK, and every id this service writes is a UUID.
+type _ScrollRecord = models.Record
+
+
+def _uuid_of(record: _ScrollRecord) -> uuid.UUID:
+    return uuid.UUID(str(record.id))
+
+
+def _to_point(record: _ScrollRecord, *, with_vectors: bool) -> Point:
+    vectors = record.vector if with_vectors and isinstance(record.vector, dict) else None
+    return Point.from_payload(_uuid_of(record), record.payload or {}, vectors=vectors)
+
+
+def _conditions(
+    *,
+    tenant_id: uuid.UUID,
+    domain: str | None = None,
+    file_id: uuid.UUID | None = None,
+    include_inactive: bool = False,
+) -> list[models.Condition]:
+    """The base filter every point query carries.
+
+    `tenant_id` is unconditional and first. `is_active` is added unless the
+    caller explicitly opted into inactive points, which is ADR-0002's
+    "`is_active: true` implied" rule expressed once rather than per method.
+    """
+    conditions: list[models.Condition] = [
+        models.FieldCondition(key="tenant_id", match=models.MatchValue(value=str(tenant_id)))
+    ]
+    if domain is not None:
+        conditions.append(
+            models.FieldCondition(key="domain", match=models.MatchValue(value=domain))
+        )
+    if file_id is not None:
+        conditions.append(
+            models.FieldCondition(key="file_id", match=models.MatchValue(value=str(file_id)))
+        )
+    if not include_inactive:
+        conditions.append(
+            models.FieldCondition(key="is_active", match=models.MatchValue(value=True))
+        )
+    return conditions
+
+
+class QdrantPointRepository:
+    """A `PointRepository` (see `src/application/ports/point_repository.py`)."""
+
+    def __init__(self, client: AsyncQdrantClient, *, collection: str) -> None:
+        self._client = client
+        self._collection = collection
+
+    async def get(
+        self, *, tenant_id: uuid.UUID, point_id: uuid.UUID, with_vectors: bool = False
+    ) -> Point | None:
+        records, _ = await self._client.scroll(
+            collection_name=self._collection,
+            scroll_filter=models.Filter(
+                must=[
+                    *_conditions(tenant_id=tenant_id, include_inactive=True),
+                    models.HasIdCondition(has_id=[str(point_id)]),
+                ]
+            ),
+            limit=1,
+            with_payload=True,
+            with_vectors=with_vectors,
+        )
+        if not records:
+            return None
+        return _to_point(records[0], with_vectors=with_vectors)
+
+    async def get_many(
+        self, *, tenant_id: uuid.UUID, point_ids: Sequence[uuid.UUID]
+    ) -> tuple[Point, ...]:
+        if not point_ids:
+            return ()
+        records, _ = await self._client.scroll(
+            collection_name=self._collection,
+            scroll_filter=models.Filter(
+                must=[
+                    *_conditions(tenant_id=tenant_id, include_inactive=True),
+                    models.HasIdCondition(has_id=[str(point_id) for point_id in point_ids]),
+                ]
+            ),
+            limit=len(point_ids),
+            with_payload=True,
+            with_vectors=False,
+        )
+        return tuple(_to_point(record, with_vectors=False) for record in records)
+
+    async def list_by_file(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        file_id: uuid.UUID,
+        limit: int,
+        cursor: str | None = None,
+        include_inactive: bool = False,
+    ) -> PointPage:
+        conditions = _conditions(
+            tenant_id=tenant_id, file_id=file_id, include_inactive=include_inactive
+        )
+        if cursor is not None:
+            conditions.append(
+                models.FieldCondition(key="order_id", range=models.Range(gt=float(cursor)))
+            )
+
+        records, _ = await self._client.scroll(
+            collection_name=self._collection,
+            scroll_filter=models.Filter(must=conditions),
+            order_by=models.OrderBy(key="order_id", direction=models.Direction.ASC),
+            limit=limit,
+            with_payload=True,
+            with_vectors=False,
+        )
+        points = tuple(_to_point(record, with_vectors=False) for record in records)
+        # A short page means the listing is exhausted. A full page might be, but
+        # claiming so would need an extra round trip; handing back a cursor that
+        # yields an empty final page is the cheaper honest answer.
+        next_cursor = repr(points[-1].order_id) if len(points) == limit else None
+        return PointPage(points=points, next_cursor=next_cursor)
+
+    async def count(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> int:
+        result = await self._client.count(
+            collection_name=self._collection,
+            count_filter=models.Filter(
+                must=_conditions(
+                    tenant_id=tenant_id,
+                    domain=domain,
+                    file_id=file_id,
+                    include_inactive=include_inactive,
+                )
+            ),
+            exact=True,
+        )
+        return result.count
+
+    async def keyword_search(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        query: str,
+        limit: int,
+        cursor: str | None = None,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> PointPage:
+        conditions = _conditions(
+            tenant_id=tenant_id,
+            domain=domain,
+            file_id=file_id,
+            include_inactive=include_inactive,
+        )
+        conditions.append(models.FieldCondition(key="content", match=models.MatchText(text=query)))
+
+        # No `order_by` here, so Qdrant does return a page offset: the full-text
+        # index filters rather than scores, and imposing `order_id` ordering
+        # across files would be meaningless (`order_id` is per-file).
+        records, next_offset = await self._client.scroll(
+            collection_name=self._collection,
+            scroll_filter=models.Filter(must=conditions),
+            limit=limit,
+            offset=cursor,
+            with_payload=True,
+            with_vectors=False,
+        )
+        return PointPage(
+            points=tuple(_to_point(record, with_vectors=False) for record in records),
+            next_cursor=str(next_offset) if next_offset is not None else None,
+        )
+
+    async def apply_patches(self, *, tenant_id: uuid.UUID, patches: Sequence[PayloadPatch]) -> None:
+        if not patches:
+            return
+        await self._client.batch_update_points(
+            collection_name=self._collection,
+            update_operations=[
+                models.SetPayloadOperation(
+                    set_payload=models.SetPayload(
+                        payload=patch.payload,
+                        filter=models.Filter(must=self._patch_conditions(tenant_id, patch)),
+                    )
+                )
+                for patch in patches
+            ],
+        )
+
+    @staticmethod
+    def _patch_conditions(tenant_id: uuid.UUID, patch: PayloadPatch) -> list[models.Condition]:
+        """Address one point by id, under this tenant, at an expected version.
+
+        `include_inactive=True`: soft-deleting and relinking both have to reach
+        points the default read filter hides, and a patch is addressed by an
+        explicit id rather than discovered by a listing.
+        """
+        conditions: list[models.Condition] = [
+            *_conditions(tenant_id=tenant_id, include_inactive=True),
+            models.HasIdCondition(has_id=[str(patch.point_id)]),
+        ]
+        if patch.expected_version is not None:
+            conditions.append(
+                models.FieldCondition(
+                    key="version", match=models.MatchValue(value=patch.expected_version)
+                )
+            )
+        return conditions

src/infrastructure/qdrant/points.py

@@ -0,0 +1,97 @@
+"""Qdrant adapter for the `PointStorage` port (ADR-0001, ADR-0002, ADR-0015).
+
+The `qdrant_client` SDK appears here and nowhere in `application/`. This module
+translates the SDK-free `ChunkPoint` into `PointStruct`s and builds every
+filter — routers and application services never construct Qdrant filters.
+
+`AsyncQdrantClient` is genuinely async, so unlike the `minio` adapter nothing
+here needs a thread offload.
+"""
+
+import uuid
+from collections.abc import Sequence
+from datetime import datetime
+
+from qdrant_client import AsyncQdrantClient, models
+
+from src.application.points.models import ChunkPoint
+from src.infrastructure.qdrant.collection import SPARSE_VECTOR
+
+
+def _tenant_file_filter(
+    tenant_id: uuid.UUID, file_id: uuid.UUID, *, from_chunk_index: int
+) -> models.Filter:
+    """Points of one file, at or past `from_chunk_index`, within one tenant.
+
+    `tenant_id` is always a condition, never optional: a `file_id` alone is not
+    authority to mutate anything (ADR-0002's isolation rule applies to every
+    code path, not just reads).
+    """
+    return models.Filter(
+        must=[
+            models.FieldCondition(key="tenant_id", match=models.MatchValue(value=str(tenant_id))),
+            models.FieldCondition(key="file_id", match=models.MatchValue(value=str(file_id))),
+            models.FieldCondition(key="chunk_index", range=models.Range(gte=from_chunk_index)),
+        ]
+    )
+
+
+class QdrantPointStorage:
+    """A `PointStorage` (see `src/application/ports/point_storage.py`)."""
+
+    def __init__(self, client: AsyncQdrantClient, *, collection: str) -> None:
+        self._client = client
+        self._collection = collection
+
+    async def upsert_points(self, points: Sequence[ChunkPoint]) -> None:
+        if not points:
+            return
+        await self._client.upsert(
+            collection_name=self._collection,
+            points=[
+                models.PointStruct(
+                    id=str(point.point_id),
+                    vector={
+                        **point.dense,
+                        SPARSE_VECTOR: models.SparseVector(
+                            indices=point.sparse.indices, values=point.sparse.values
+                        ),
+                    },
+                    payload=point.payload,
+                )
+                for point in points
+            ],
+        )
+
+    async def deactivate_points_from_index(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        file_id: uuid.UUID,
+        from_chunk_index: int,
+        deleted_at: datetime,
+        updated_by: str,
+    ) -> int:
+        """Soft-delete via `set_payload` — the points stay for audit (ADR-0002).
+
+        Counts first so the caller can report how many points the sweep
+        touched; `set_payload` itself reports only an operation status.
+        """
+        point_filter = _tenant_file_filter(tenant_id, file_id, from_chunk_index=from_chunk_index)
+        stale = await self._client.count(
+            collection_name=self._collection, count_filter=point_filter, exact=True
+        )
+        if stale.count == 0:
+            return 0
+
+        await self._client.set_payload(
+            collection_name=self._collection,
+            payload={
+                "is_active": False,
+                "deleted_at": deleted_at.isoformat(),
+                "updated_at": deleted_at.isoformat(),
+                "updated_by": updated_by,
+            },
+            points=models.FilterSelector(filter=point_filter),
+        )
+        return stale.count

src/main.py

@@ -1,5 +1,7 @@
 from fastapi import FastAPI
 
+from src.api.errors import register_exception_handlers
+from src.api.middleware import RequestIdMiddleware
 from src.api.router import router as v1_router
 from src.api.routers.health import router as health_router
 from src.bootstrap.lifespan import create_lifespan
@@ -8,6 +10,8 @@ from src.config import Settings
 
 def create_app(settings: Settings | None = None) -> FastAPI:
     app = FastAPI(lifespan=create_lifespan(settings))
+    app.add_middleware(RequestIdMiddleware)
+    register_exception_handlers(app)
     app.include_router(health_router)
     app.include_router(v1_router, prefix="/v1")
     return app

tests/conftest.py

@@ -1,7 +1,8 @@
-from collections.abc import AsyncIterator
+from collections.abc import AsyncIterator, Iterator
 
 import pytest
 import pytest_asyncio
+import structlog
 from asgi_lifespan import LifespanManager
 from fastapi import FastAPI
 from httpx import ASGITransport, AsyncClient
@@ -9,15 +10,50 @@ from httpx import ASGITransport, AsyncClient
 from src.config import Settings
 from src.main import create_app
 
+# Disposable Postgres/MinIO/Qdrant containers (ADR-0016). Registered here
+# because `pytest_plugins` is only honoured in the root conftest, and both
+# `tests/integration/*/` and `tests/e2e/` need the same containers. The
+# fixtures are session-scoped but lazy, so unit runs still need no Docker.
+pytest_plugins = ["tests.support.containers"]
+
+
+@pytest.fixture(autouse=True)
+def _reset_structlog_after_test() -> Iterator[None]:
+    """Undo any real `configure_logging()` call before the next test runs.
+
+    Any test that exercises the app's lifespan (directly, or via the `client`/
+    `api_client` fixtures below and in `test_domains_api.py`) calls the real
+    `configure_logging()`, which mutates *global* structlog/stdlib state --
+    including `cache_logger_on_first_use=True`. Left in place, that setting
+    silently breaks `structlog.testing.capture_logs()` in unrelated tests
+    later in the same pytest process: a module-level
+    `logger = structlog.get_logger(__name__)` cached under the real config no
+    longer routes through `capture_logs()`'s temporary processor swap, so
+    assertions on captured events see nothing (ADR-0016: isolate per test --
+    this generalizes to global config mutations, not just data).
+    """
+    yield
+    structlog.reset_defaults()
+
 
 @pytest.fixture
 def settings() -> Settings:
+    # Every external dependency points at a closed port so unit tests never
+    # reach real infrastructure. For the embedders this matters twice over:
+    # the real defaults are a colleague's Ollama box and OpenAI's paid API,
+    # and ADR-0016 forbids routine runs calling either. Connection-refused is
+    # immediate, so the lifespan's fail-soft warm-up costs nothing here --
+    # and these tests passing at all is what proves it is fail-soft.
     return Settings(
         postgres={"host": "127.0.0.1", "port": 1},
         minio={"endpoint": "127.0.0.1:1"},
         ingestion={"timeout_seconds": 1.0},
         qdrant={"url": "http://127.0.0.1:1"},
         app={"readiness_check_timeout_seconds": 0.5},
+        embedding={
+            "nomic": {"base_url": "http://127.0.0.1:1/v1", "timeout_seconds": 0.5},
+            "openai": {"base_url": "http://127.0.0.1:1/v1", "timeout_seconds": 0.5},
+        },
     )
 
 
@@ -26,6 +62,24 @@ def app(settings: Settings) -> FastAPI:
     return create_app(settings)
 
 
+@pytest.fixture(autouse=True)
+def _no_real_logging_configuration(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Prevent the app lifespan from calling the real `configure_logging()`.
+
+    It sets `cache_logger_on_first_use=True` (ADR-0011), which permanently
+    monkeypatches the `.bind` method on whichever module-level
+    `logger = structlog.get_logger(__name__)` instance is used first --
+    `structlog.reset_defaults()` only resets *global* config, not that
+    per-instance mutation, so real configuration leaking into one test would
+    silently break `structlog.testing.capture_logs()` in every test that runs
+    afterward in the same process (ADR-0016: isolate per test). Tests that
+    spin up the full app via `LifespanManager` (`client`, `api_client`) are
+    testing HTTP behavior, not logging output, so they don't need it for
+    real.
+    """
+    monkeypatch.setattr("src.bootstrap.lifespan.configure_logging", lambda *a, **k: None)
+
+
 @pytest_asyncio.fixture
 async def client(app: FastAPI) -> AsyncIterator[AsyncClient]:
     async with (

tests/e2e/conftest.py

@@ -0,0 +1,244 @@
+"""The vertical slice against real Postgres, MinIO, and Qdrant (ADR-0016).
+
+These are the ADR-0016 end-to-end tests: "a small vertical-slice acceptance
+contract through the real application composition". They run on
+Testcontainers, in the default `uv run pytest` run, because ADR-0016 makes
+Testcontainers "the standard automated integration-test resource mechanism"
+and reserves Docker Compose for manual validation and the separate serialized
+smoke test of the *running web process* (`tests/e2e/test_compose_smoke.py`).
+
+What is real here: routing, auth, scopes, the domain allowlist, the ADR-0008
+error envelope, the two-transaction upload, MinIO object writes, BM25 sparse
+embedding, and Qdrant upserts into a per-test collection created by the
+production `ensure_chunks_collection`.
+
+What is faked, and only this: the two dense embedders. They are paid/remote
+network calls, and ADR-0016 forbids routine runs from touching a live
+provider. Their fakes return the pinned 768/3072 dimensions, because the real
+collection rejects anything else.
+"""
+
+import uuid
+from collections.abc import AsyncIterator, Callable, Coroutine
+from contextlib import AsyncExitStack
+from dataclasses import dataclass
+from typing import Any
+
+import pytest
+import pytest_asyncio
+from asgi_lifespan import LifespanManager
+from httpx import ASGITransport, AsyncClient
+from qdrant_client import AsyncQdrantClient, models
+from sqlalchemy.ext.asyncio import AsyncEngine, AsyncSession, async_sessionmaker
+
+from src.bootstrap.dependencies import get_dense_embedders, get_sessionmaker
+from src.config import MinioSettings, PostgresSettings, QdrantSettings, Settings
+from src.infrastructure.postgres.database import create_sessionmaker
+from src.infrastructure.qdrant.collection import (
+    DENSE_NOMIC_DIMENSIONS,
+    DENSE_OPENAI_DIMENSIONS,
+    ensure_chunks_collection,
+)
+from src.main import create_app
+from tests.fakes import FakeDenseEmbedder
+from tests.support.factories import create_api_key, create_tenant, create_tenant_domain
+
+CSV_BYTES = b"question,answer\nhow do I file a claim,call the branch\nwhat is covered,see policy\n"
+
+# Past validation's OOXML magic-byte check, so the failure lands in the DOCX
+# parser rather than in upload validation -- the branch this exercises.
+CORRUPT_DOCX_BYTES = b"PK\x03\x04" + b"not really a docx" * 8
+
+
+@dataclass
+class E2EStack:
+    """One configured app instance plus the handles a test asserts against."""
+
+    client: AsyncClient
+    collection: str
+    dense_embedders: list[FakeDenseEmbedder]
+
+
+StackFactory = Callable[..., Coroutine[Any, Any, E2EStack]]
+
+
+@dataclass(frozen=True)
+class Principal:
+    """A provisioned tenant with a bearer token and a registered domain.
+
+    Plain values, not the ORM `Tenant`: tests re-read rows the app committed on
+    a shared connection, and any expiry/refresh of a held ORM instance would
+    lazy-load outside a greenlet context (`MissingGreenlet`).
+    """
+
+    tenant_id: uuid.UUID
+    tenant_slug: str
+    token: str
+    domain: str
+
+    @property
+    def headers(self) -> dict[str, str]:
+        return {"Authorization": f"Bearer {self.token}"}
+
+
+@pytest.fixture
+def e2e_settings(
+    postgres_settings: PostgresSettings,
+    minio_settings: MinioSettings,
+    qdrant_settings: QdrantSettings,
+) -> Settings:
+    """Settings wired to this test's containers.
+
+    Every nested settings object is passed explicitly. `src/config.py` cascades
+    `.env` into each nested class, so omitting one would let a developer's real
+    `EMBEDDING_NOMIC_BASE_URL` (a colleague's Ollama box) into a routine test
+    run. The embedder URLs below point at a closed port for the same reason:
+    the lifespan warms the *real* embedders before the fakes are substituted,
+    and connection-refused is both instant and free.
+    """
+    return Settings(
+        postgres=postgres_settings,
+        minio=minio_settings,
+        qdrant=qdrant_settings,
+        embedding={
+            "nomic": {"base_url": "http://127.0.0.1:1/v1", "timeout_seconds": 0.5},
+            "openai": {"base_url": "http://127.0.0.1:1/v1", "timeout_seconds": 0.5},
+        },
+        app={"readiness_check_timeout_seconds": 2.0},
+    )
+
+
+@pytest_asyncio.fixture(loop_scope="session")
+async def make_stack(
+    e2e_settings: Settings,
+    db_sessionmaker: async_sessionmaker[AsyncSession],
+    qdrant_client: AsyncQdrantClient,
+) -> AsyncIterator[StackFactory]:
+    """Build an app under test, optionally with tightened ingestion bounds.
+
+    A factory rather than a fixture because the capacity and timeout tests need
+    their own `INGESTION_*` values, and those are read at app construction.
+    """
+    async with AsyncExitStack() as exit_stack:
+
+        async def _make(
+            *,
+            ingestion: dict[str, object] | None = None,
+            collection: str | None = None,
+            dense_delay_seconds: float = 0.0,
+            bootstrap_collection: bool = True,
+            sessionmaker: async_sessionmaker[AsyncSession] | None = None,
+        ) -> E2EStack:
+            resolved_collection = collection or e2e_settings.qdrant.collection
+            if bootstrap_collection:
+                await ensure_chunks_collection(qdrant_client, collection=resolved_collection)
+
+            settings = e2e_settings.model_copy(
+                update={
+                    "qdrant": e2e_settings.qdrant.model_copy(
+                        update={"collection": resolved_collection}
+                    ),
+                    "ingestion": e2e_settings.ingestion.model_copy(update=ingestion or {}),
+                }
+            )
+
+            dense = [
+                FakeDenseEmbedder(
+                    name="dense_nomic",
+                    dimensions=DENSE_NOMIC_DIMENSIONS,
+                    value=0.1,
+                    delay_seconds=dense_delay_seconds,
+                ),
+                FakeDenseEmbedder(
+                    name="dense_openai",
+                    dimensions=DENSE_OPENAI_DIMENSIONS,
+                    value=0.2,
+                    delay_seconds=dense_delay_seconds,
+                ),
+            ]
+
+            app = create_app(settings)
+            # The session factory is the test's, so factory-created tenants are
+            # visible to the app and every write rolls back at teardown. Object
+            # storage, point storage, and the sparse embedder stay real.
+            resolved_sessionmaker = sessionmaker or db_sessionmaker
+            app.dependency_overrides[get_sessionmaker] = lambda: resolved_sessionmaker
+            app.dependency_overrides[get_dense_embedders] = lambda: dense
+
+            manager = await exit_stack.enter_async_context(LifespanManager(app))
+            client = await exit_stack.enter_async_context(
+                AsyncClient(
+                    transport=ASGITransport(app=manager.app),
+                    base_url="http://test",
+                    timeout=30.0,
+                )
+            )
+            return E2EStack(client=client, collection=resolved_collection, dense_embedders=dense)
+
+        yield _make
+
+
+@pytest_asyncio.fixture(loop_scope="session")
+async def committed_sessionmaker(
+    postgres_engine: AsyncEngine,
+) -> async_sessionmaker[AsyncSession]:
+    """A session factory whose writes really commit, one connection each.
+
+    The default `db_sessionmaker` pins every session to a single connection
+    inside one rolled-back transaction, which is what isolates a test's writes
+    -- but savepoints on a shared connection cannot interleave, so any test
+    that issues genuinely *concurrent* requests deadlocks or fails with
+    `InvalidSavepointSpecificationError`. Those tests use this instead and rely
+    on their uuid-keyed tenant for isolation (ADR-0016 allows either).
+    """
+    return create_sessionmaker(postgres_engine)
+
+
+@pytest_asyncio.fixture(loop_scope="session")
+async def principal(db_session: AsyncSession) -> Principal:
+    """A tenant with an upload-scoped key and one registered domain.
+
+    Registering the domain is not optional: `POST /v1/files` rejects an
+    unregistered one with `400 unknown_domain` before anything else runs
+    (ADR-0009).
+    """
+    return await provision_principal(db_session)
+
+
+async def provision_principal(
+    session: AsyncSession, *, domain: str = "fire", scopes: list[str] | None = None
+) -> Principal:
+    tenant = await create_tenant(session)
+    _, token = await create_api_key(session, tenant=tenant, scopes=scopes or ["files:write"])
+    await create_tenant_domain(session, tenant=tenant, domain=domain)
+    await session.commit()
+    return Principal(tenant_id=tenant.id, tenant_slug=tenant.slug, token=token, domain=domain)
+
+
+def upload_payload(
+    *, domain: str, filename: str = "faq.csv", data: bytes = CSV_BYTES
+) -> dict[str, Any]:
+    return {"files": {"file": (filename, data, "text/csv")}, "data": {"domain": domain}}
+
+
+async def count_active_points(
+    client: AsyncQdrantClient, collection: str, *, tenant_id: uuid.UUID
+) -> int:
+    """Active points for one tenant, read through the raw client.
+
+    `PointStorage` is deliberately write-only (point reads are plan 002's
+    `/v1/points`), so the verification read here uses the SDK directly.
+    """
+    result = await client.count(
+        collection_name=collection,
+        count_filter=models.Filter(
+            must=[
+                models.FieldCondition(
+                    key="tenant_id", match=models.MatchValue(value=str(tenant_id))
+                ),
+                models.FieldCondition(key="is_active", match=models.MatchValue(value=True)),
+            ]
+        ),
+        exact=True,
+    )
+    return result.count

tests/e2e/test_compose_smoke.py

@@ -0,0 +1,143 @@
+"""Operational smoke test against the *running web process* (ADR-0016).
+
+Different in kind from every other test in this repo. Everything else drives
+the app in-process through `ASGITransport`; this drives a real uvicorn process
+over a real socket, backed by the Docker Compose stack, after the two
+deployment steps (`alembic upgrade head`, `src.cli.qdrant_bootstrap`) have run
+for real. ADR-0016 reserves Compose for exactly this: "manual local validation
+and a later, serialized operational smoke test of the running web process",
+run "as a serialized pre-release or scheduled gate".
+
+It is therefore **skipped unless `SMOKE_BASE_URL` is set**, so a plain
+`uv run pytest` never touches Docker Compose. Run it through `scripts/smoke.sh`,
+which brings the stack up, provisions a tenant, starts the process, and exports
+the environment below.
+
+Because the app runs in its own process, this is also the only place in the
+suite where the real `configure_logging()` is in effect --
+`tests/conftest.py::_no_real_logging_configuration` no-ops it everywhere else
+to keep global structlog state out of other tests. So the ADR-0011 log sink is
+asserted here and nowhere else.
+"""
+
+import json
+import os
+import uuid
+from collections.abc import AsyncIterator
+from pathlib import Path
+
+import pytest
+import pytest_asyncio
+from httpx import AsyncClient
+from qdrant_client import AsyncQdrantClient, models
+
+BASE_URL = os.environ.get("SMOKE_BASE_URL")
+API_KEY = os.environ.get("SMOKE_API_KEY", "")
+DOMAIN = os.environ.get("SMOKE_DOMAIN", "smoke")
+LOG_PATH = os.environ.get("SMOKE_LOG_PATH", "")
+QDRANT_URL = os.environ.get("SMOKE_QDRANT_URL", "http://127.0.0.1:6343")
+QDRANT_COLLECTION = os.environ.get("SMOKE_QDRANT_COLLECTION", "chunks")
+
+pytestmark = [
+    pytest.mark.e2e,
+    pytest.mark.slow,
+    pytest.mark.asyncio,
+    pytest.mark.skipif(
+        not BASE_URL,
+        reason="SMOKE_BASE_URL is unset; run this through scripts/smoke.sh",
+    ),
+    # Well past the global 10s budget: this drives a real process over a
+    # socket, and the upload does real embedding work end to end.
+    pytest.mark.timeout(120),
+]
+
+# Unique per run so a re-run against the same persistent Compose volumes is a
+# new file rather than an idempotent duplicate-hash hit (ADR-0017).
+_CSV = f"question,answer\nsmoke run {uuid.uuid4().hex},indexed\n".encode()
+
+
+@pytest_asyncio.fixture
+async def client() -> AsyncIterator[AsyncClient]:
+    async with AsyncClient(base_url=BASE_URL or "", timeout=120.0) as async_client:
+        yield async_client
+
+
+async def test_running_process_reports_healthy_and_ready() -> None:
+    """Readiness must be green *after* the deployment steps, not before --
+    an unbootstrapped Qdrant is deliberately unready.
+    """
+    async with AsyncClient(base_url=BASE_URL or "", timeout=30.0) as client:
+        healthz = await client.get("/healthz")
+        readyz = await client.get("/readyz")
+
+    assert healthz.status_code == 200
+    assert healthz.json() == {"status": "ok"}
+    assert readyz.status_code == 200, readyz.text
+    assert readyz.json() == {"postgres": True, "minio": True, "qdrant": True}
+
+
+async def test_upload_through_the_running_process_indexes_retrievable_points(
+    client: AsyncClient,
+) -> None:
+    """The whole slice against the deployed shape: HTTP in, Qdrant points out."""
+    upload = await client.post(
+        "/v1/files",
+        files={"file": ("smoke.csv", _CSV, "text/csv")},
+        data={"domain": DOMAIN},
+        headers={"Authorization": f"Bearer {API_KEY}"},
+    )
+    assert upload.status_code == 201, upload.text
+    body = upload.json()
+    assert body["status"] == "succeeded"
+    assert body["chunks_indexed"] > 0
+
+    status = await client.get(
+        f"/v1/files/{body['file_id']}", headers={"Authorization": f"Bearer {API_KEY}"}
+    )
+    assert status.status_code == 200
+    assert status.json()["ingestion_status"] == "succeeded"
+
+    qdrant = AsyncQdrantClient(url=QDRANT_URL)
+    try:
+        count = await qdrant.count(
+            collection_name=QDRANT_COLLECTION,
+            count_filter=models.Filter(
+                must=[
+                    models.FieldCondition(
+                        key="file_id", match=models.MatchValue(value=body["file_id"])
+                    ),
+                    models.FieldCondition(key="is_active", match=models.MatchValue(value=True)),
+                ]
+            ),
+            exact=True,
+        )
+    finally:
+        await qdrant.close()
+
+    assert count.count == body["chunks_indexed"]
+
+
+async def test_running_process_writes_structured_ingestion_logs() -> None:
+    """The real ADR-0011 sink, which only a separate process exercises.
+
+    An upload that succeeds while emitting nothing parseable is an upload no
+    operator can investigate -- and the runbook's failure procedure is written
+    against these exact event names.
+    """
+    if not LOG_PATH:
+        pytest.skip("SMOKE_LOG_PATH is unset; scripts/smoke.sh normally provides it")
+
+    lines = Path(LOG_PATH).read_text("utf-8").splitlines()
+    events = [json.loads(line) for line in lines if line.startswith("{")]
+
+    by_name = {event.get("event") for event in events}
+    assert "ingestion.job.started" in by_name
+    assert "ingestion.job.completed" in by_name
+
+    completed = next(event for event in events if event.get("event") == "ingestion.job.completed")
+    assert completed["tenant_id"]
+    assert completed["ingestion_job_id"]
+    assert completed["points_upserted"] > 0
+    # Static process context bound once at startup (ADR-0011).
+    assert completed["env"]
+    assert completed["service_version"]

tests/e2e/test_ingestion_slice.py

@@ -0,0 +1,329 @@
+"""The ingestion vertical slice end to end (plan 001, Phase 6).
+
+Every test here asserts a reliability invariant ADR-0016 lists as a reusable
+contract, not a happy path: the bound maps to its status code, the failure
+still writes a terminal job row, the retry converges, and the tenant filter
+holds. Failures additionally assert the ADR-0008 error envelope, because the
+status code alone is not the contract a client integrates against.
+"""
+
+import asyncio
+import uuid
+
+import pytest
+from httpx import Response
+from qdrant_client import AsyncQdrantClient
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+
+from src.infrastructure.postgres.models.ingestion_job import IngestionJob
+from tests.e2e.conftest import (
+    CORRUPT_DOCX_BYTES,
+    Principal,
+    StackFactory,
+    count_active_points,
+    provision_principal,
+    upload_payload,
+)
+
+pytestmark = [
+    pytest.mark.e2e,
+    pytest.mark.postgres,
+    pytest.mark.minio,
+    pytest.mark.qdrant,
+    pytest.mark.asyncio(loop_scope="session"),
+]
+
+
+async def _latest_job(session: AsyncSession, *, tenant_id: uuid.UUID) -> IngestionJob:
+    """The tenant's most recent job row, read fresh.
+
+    `populate_existing` matters: this session shares a connection with the
+    app's sessions, so without it the identity map can answer from a row
+    loaded before the request under test committed its update.
+    """
+    result = await session.execute(
+        select(IngestionJob)
+        .where(IngestionJob.tenant_id == tenant_id)
+        .order_by(IngestionJob.created_at.desc(), IngestionJob.id)
+        .execution_options(populate_existing=True)
+    )
+    return result.scalars().first() or pytest.fail("no ingestion job was written")
+
+
+def _envelope(response: Response) -> dict[str, object]:
+    """The ADR-0008 error body. Asserting the code here, not just the HTTP
+    status, is the difference between testing the contract and testing the
+    number."""
+    error: dict[str, object] = response.json()["error"]
+    assert error["request_id"], "every error envelope carries its correlation id"
+    return error
+
+
+async def test_upload_succeeds_and_indexes_points_under_the_tenant_filter(
+    make_stack: StackFactory, principal: Principal, qdrant_client: AsyncQdrantClient
+) -> None:
+    stack = await make_stack()
+
+    response = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+
+    assert response.status_code == 201
+    body = response.json()
+    assert body["status"] == "succeeded"
+    assert body["chunks_indexed"] > 0
+    indexed = await count_active_points(
+        qdrant_client, stack.collection, tenant_id=principal.tenant_id
+    )
+    assert indexed == body["chunks_indexed"]
+
+
+async def test_get_file_reports_the_terminal_status_after_upload(
+    make_stack: StackFactory, principal: Principal
+) -> None:
+    stack = await make_stack()
+    created = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+
+    status = await stack.client.get(
+        f"/v1/files/{created.json()['file_id']}", headers=principal.headers
+    )
+
+    assert status.status_code == 200
+    assert status.json()["ingestion_status"] == "succeeded"
+    assert status.json()["domain"] == principal.domain
+
+
+async def test_upload_duplicate_content_returns_the_existing_job_without_reingesting(
+    make_stack: StackFactory, principal: Principal, qdrant_client: AsyncQdrantClient
+) -> None:
+    stack = await make_stack()
+    first = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+    points_after_first = await count_active_points(
+        qdrant_client, stack.collection, tenant_id=principal.tenant_id
+    )
+
+    second = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+
+    assert first.status_code == 201
+    # 200, not 201: nothing was created the second time (ADR-0017 idempotency).
+    assert second.status_code == 200
+    assert second.json()["file_id"] == first.json()["file_id"]
+    assert second.json()["ingestion_job_id"] == first.json()["ingestion_job_id"]
+    assert (
+        await count_active_points(qdrant_client, stack.collection, tenant_id=principal.tenant_id)
+        == points_after_first
+    )
+
+
+async def test_upload_retried_after_a_failed_job_succeeds_without_duplicate_points(
+    make_stack: StackFactory,
+    principal: Principal,
+    db_session: AsyncSession,
+    qdrant_client: AsyncQdrantClient,
+) -> None:
+    """ADR-0017, "Re-running an ingestion stays safe": a failed attempt is
+    retried by re-uploading the same bytes, and deterministic point ids make
+    the retry converge instead of duplicating.
+    """
+    stack = await make_stack()
+    stack.dense_embedders[0].fail_next = True
+
+    failed = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+    failed_job = await _latest_job(db_session, tenant_id=principal.tenant_id)
+    retried = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+
+    assert failed.status_code == 502
+    assert failed_job.status == "failed"
+    assert failed_job.error_code == "embedding_failed"
+    assert retried.status_code == 201
+    assert retried.json()["status"] == "succeeded"
+    # Same source file, new attempt -- a terminal job is never reused or
+    # transitioned back to running.
+    assert retried.json()["ingestion_job_id"] != str(failed_job.id)
+    indexed = await count_active_points(
+        qdrant_client, stack.collection, tenant_id=principal.tenant_id
+    )
+    assert indexed == retried.json()["chunks_indexed"]
+
+
+async def test_get_file_for_another_tenants_file_returns_404_not_403(
+    make_stack: StackFactory, principal: Principal, db_session: AsyncSession
+) -> None:
+    """404, never 403: a 403 would confirm the file exists to a stranger."""
+    stack = await make_stack()
+    created = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+    other = await provision_principal(db_session)
+
+    response = await stack.client.get(
+        f"/v1/files/{created.json()['file_id']}", headers=other.headers
+    )
+
+    assert response.status_code == 404
+    assert _envelope(response)["code"] == "not_found"
+
+
+async def test_upload_to_an_unregistered_domain_is_rejected_before_any_job_is_written(
+    make_stack: StackFactory, principal: Principal, db_session: AsyncSession
+) -> None:
+    stack = await make_stack()
+
+    response = await stack.client.post(
+        "/v1/files", **upload_payload(domain="never-registered"), headers=principal.headers
+    )
+
+    assert response.status_code == 400
+    assert _envelope(response)["code"] == "unknown_domain"
+    jobs = await db_session.execute(
+        select(IngestionJob).where(IngestionJob.tenant_id == principal.tenant_id)
+    )
+    assert jobs.scalars().all() == []
+
+
+async def test_upload_at_capacity_is_rejected_with_503_and_retry_after(
+    make_stack: StackFactory, committed_sessionmaker: async_sessionmaker[AsyncSession]
+) -> None:
+    """ADR-0017 rejects rather than queues, so the second concurrent upload
+    must fail fast with a backoff hint instead of waiting for a slot.
+
+    This is the one test that needs two requests genuinely in flight at once,
+    so it runs on `committed_sessionmaker` -- see that fixture for why the
+    shared-connection default cannot serve concurrent sessions.
+    """
+    async with committed_sessionmaker() as session:
+        principal = await provision_principal(session)
+    stack = await make_stack(
+        ingestion={"max_concurrency": 1},
+        dense_delay_seconds=0.4,
+        sessionmaker=committed_sessionmaker,
+    )
+
+    first, second = await asyncio.gather(
+        stack.client.post(
+            "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+        ),
+        stack.client.post(
+            "/v1/files",
+            **upload_payload(domain=principal.domain, filename="other.csv", data=b"a,b\n1,2\n"),
+            headers=principal.headers,
+        ),
+    )
+
+    statuses = sorted([first.status_code, second.status_code])
+    assert statuses == [201, 503]
+    rejected = first if first.status_code == 503 else second
+    assert _envelope(rejected)["code"] == "ingestion_at_capacity"
+    assert rejected.headers["Retry-After"] == "1"
+
+
+async def test_upload_exceeding_the_timeout_returns_504_and_writes_a_terminal_job(
+    make_stack: StackFactory, principal: Principal, db_session: AsyncSession
+) -> None:
+    """A timeout must never leave a job stuck in `running` (ADR-0016)."""
+    stack = await make_stack(ingestion={"timeout_seconds": 0.05}, dense_delay_seconds=0.5)
+
+    response = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+
+    assert response.status_code == 504
+    assert _envelope(response)["code"] == "ingestion_timeout"
+    job = await _latest_job(db_session, tenant_id=principal.tenant_id)
+    assert job.status == "failed"
+    assert job.error_code == "timeout"
+
+
+async def test_upload_of_an_unparseable_document_returns_400_and_a_failed_job(
+    make_stack: StackFactory, principal: Principal, db_session: AsyncSession
+) -> None:
+    stack = await make_stack()
+
+    response = await stack.client.post(
+        "/v1/files",
+        **upload_payload(domain=principal.domain, filename="broken.docx", data=CORRUPT_DOCX_BYTES),
+        headers=principal.headers,
+    )
+
+    assert response.status_code == 400
+    assert _envelope(response)["code"] == "validation_error"
+    job = await _latest_job(db_session, tenant_id=principal.tenant_id)
+    assert job.status == "failed"
+    assert job.error_code == "parse_failed"
+
+
+async def test_upload_returns_502_when_qdrant_indexing_fails(
+    make_stack: StackFactory, principal: Principal, db_session: AsyncSession
+) -> None:
+    """A real Qdrant error, not a fake: the collection the deployment step
+    should have created is missing, which is the failure an operator actually
+    hits when `qdrant_bootstrap` was skipped.
+    """
+    stack = await make_stack(
+        collection=f"never_bootstrapped_{uuid.uuid4().hex}", bootstrap_collection=False
+    )
+
+    response = await stack.client.post(
+        "/v1/files", **upload_payload(domain=principal.domain), headers=principal.headers
+    )
+
+    assert response.status_code == 502
+    assert _envelope(response)["code"] == "index_error"
+    job = await _latest_job(db_session, tenant_id=principal.tenant_id)
+    assert job.status == "failed"
+    assert job.error_code == "index_failed"
+
+
+async def test_upload_without_the_files_write_scope_is_forbidden(
+    make_stack: StackFactory, db_session: AsyncSession
+) -> None:
+    stack = await make_stack()
+    reader = await provision_principal(db_session, scopes=["domains:read"])
+
+    response = await stack.client.post(
+        "/v1/files", **upload_payload(domain=reader.domain), headers=reader.headers
+    )
+
+    assert response.status_code == 403
+    assert _envelope(response)["code"] == "missing_scope"
+
+
+async def test_readyz_reports_every_dependency_ready_against_real_services(
+    make_stack: StackFactory,
+) -> None:
+    """`/readyz` is dependency readiness, `/healthz` is process health."""
+    stack = await make_stack()
+
+    ready = await stack.client.get("/readyz")
+    healthy = await stack.client.get("/healthz")
+
+    assert healthy.status_code == 200
+    assert ready.status_code == 200
+    assert ready.json() == {"postgres": True, "minio": True, "qdrant": True}
+
+
+async def test_readyz_is_unready_when_the_chunks_collection_is_missing(
+    make_stack: StackFactory,
+) -> None:
+    """A reachable but unbootstrapped Qdrant is deliberately not ready --
+    uploads to it would 502 (see the indexing test above).
+    """
+    stack = await make_stack(
+        collection=f"never_bootstrapped_{uuid.uuid4().hex}", bootstrap_collection=False
+    )
+
+    response = await stack.client.get("/readyz")
+
+    assert response.status_code == 503
+    assert response.json()["qdrant"] is False

tests/fakes.py

@@ -1,5 +1,301 @@
-"""Hand-written fakes for narrow application-owned ports.
+"""Hand-written fakes for narrow application-owned ports (ADR-0016)."""
 
-No application ports exist yet (Phase 1 only wires infrastructure client
-lifecycle). Fakes are added here as ports are introduced in later phases.
+import asyncio
+import uuid
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from datetime import datetime
+
+from src.application.ingestion.models import SparseVector
+from src.application.points.models import ChunkPoint
+from src.application.points.point import Point
+from src.application.ports.point_repository import PayloadPatch, PointPage
+
+
+@dataclass
+class FakeObjectStorage:
+    """In-memory `ObjectStorage`. `fail_next` simulates one upload failure."""
+
+    objects: dict[str, bytes] = field(default_factory=dict)
+    fail_next: bool = False
+
+    async def put_object(self, *, key: str, data: bytes, content_type: str) -> None:
+        if self.fail_next:
+            self.fail_next = False
+            raise OSError("simulated object storage failure")
+        self.objects[key] = data
+
+
+@dataclass
+class FakeDenseEmbedder:
+    """A scripted `DenseEmbedder`. Returns a fixed-dimension zero vector per
+    text by default; `fail_next` simulates one batch failure.
     """
+
+    name: str
+    dimensions: int = 4
+    value: float = 0.0
+    """Component value of every returned vector. Non-zero where a real Qdrant
+    has to score the result, since a zero vector has no direction to compare."""
+    model_version: str = "fake-dense-v1"
+    calls: list[list[str]] = field(default_factory=list)
+    fail_next: bool = False
+    delay_seconds: float = 0.0
+    """Simulates a slow provider call, e.g. to exercise timeout handling."""
+
+    async def embed_batch(self, texts: Sequence[str]) -> list[list[float]]:
+        self.calls.append(list(texts))
+        if self.delay_seconds:
+            await asyncio.sleep(self.delay_seconds)
+        if self.fail_next:
+            self.fail_next = False
+            raise RuntimeError("simulated embedder failure")
+        return [[self.value] * self.dimensions for _ in texts]
+
+
+@dataclass
+class FakeSparseEmbedder:
+    """A scripted `SparseEmbedder`. Returns an empty sparse vector per text."""
+
+    name: str = "sparse"
+    model_version: str = "fake-sparse-v1"
+    calls: list[list[str]] = field(default_factory=list)
+    fail_next: bool = False
+
+    def embed_batch(self, texts: Sequence[str], *, query: bool = False) -> list[SparseVector]:
+        self.calls.append(list(texts))
+        if self.fail_next:
+            self.fail_next = False
+            raise RuntimeError("simulated embedder failure")
+        return [SparseVector(indices=[], values=[]) for _ in texts]
+
+
+@dataclass
+class FakePointStorage:
+    """An in-memory `PointStorage`.
+
+    `points` is keyed by point id, so a re-upsert of the same deterministic id
+    overwrites rather than accumulating — the property a test asserting "a
+    retry produces no duplicate points" needs the fake to actually model.
+
+    `fail_on_batch` fails the Nth (0-based) upsert batch, which is how a test
+    checks that the soft-delete sweep never runs after a partial failure.
+    """
+
+    points: dict[str, ChunkPoint] = field(default_factory=dict)
+    upsert_batches: list[int] = field(default_factory=list)
+    deactivate_calls: list[dict[str, object]] = field(default_factory=list)
+    fail_on_batch: int | None = None
+    fail_deactivate: bool = False
+    max_in_flight: int = 0
+    _in_flight: int = 0
+
+    async def upsert_points(self, points: Sequence[ChunkPoint]) -> None:
+        self._in_flight += 1
+        self.max_in_flight = max(self.max_in_flight, self._in_flight)
+        try:
+            # Yield so concurrent batches actually overlap; without this the
+            # in-flight ceiling is trivially 1 and the bound goes untested.
+            await asyncio.sleep(0)
+            index = len(self.upsert_batches)
+            self.upsert_batches.append(len(points))
+            if self.fail_on_batch is not None and index == self.fail_on_batch:
+                raise RuntimeError("simulated point storage failure")
+            for point in points:
+                self.points[str(point.point_id)] = point
+        finally:
+            self._in_flight -= 1
+
+    async def deactivate_points_from_index(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        file_id: uuid.UUID,
+        from_chunk_index: int,
+        deleted_at: datetime,
+        updated_by: str,
+    ) -> int:
+        self.deactivate_calls.append(
+            {
+                "tenant_id": tenant_id,
+                "file_id": file_id,
+                "from_chunk_index": from_chunk_index,
+                "updated_by": updated_by,
+            }
+        )
+        if self.fail_deactivate:
+            raise RuntimeError("simulated deactivate failure")
+
+        def is_stale(point: ChunkPoint) -> bool:
+            chunk_index = point.payload.get("chunk_index")
+            return (
+                point.payload.get("file_id") == str(file_id)
+                and point.payload.get("tenant_id") == str(tenant_id)
+                and point.payload.get("is_active") is True
+                and isinstance(chunk_index, int)
+                and chunk_index >= from_chunk_index
+            )
+
+        stale = [point for point in self.points.values() if is_stale(point)]
+        for point in stale:
+            point.payload["is_active"] = False
+            point.payload["deleted_at"] = deleted_at.isoformat()
+        return len(stale)
+
+
+@dataclass
+class FakePointRepository:
+    """An in-memory `PointRepository`.
+
+    Stores `Point` models keyed by point id. The filter, ordering, and version
+    semantics below are reimplemented in Python rather than stubbed, because
+    every service this fake stands behind is being tested *for* those semantics:
+    a stub that returned points in insertion order would pass a reorder test
+    that real Qdrant fails.
+
+    Three behaviours are modelled deliberately:
+
+    - `is_active` is filtered out unless `include_inactive`, matching ADR-0002's
+      implied read filter.
+    - `list_by_file` sorts by `order_id` and paginates by value, the same cursor
+      mechanism the Qdrant adapter uses, so a test can catch a cursor that skips
+      or repeats.
+    - `apply_patches` honours `expected_version` by silently matching nothing on
+      a mismatch — the no-op-not-error behaviour Qdrant's filtered `set_payload`
+      actually has, which is what makes the service's post-check necessary.
+    """
+
+    points: dict[str, Point] = field(default_factory=dict)
+    patch_batches: list[list[PayloadPatch]] = field(default_factory=list)
+    fail_next: bool = False
+
+    def add(self, point: Point) -> None:
+        """Seed a point. Test-only helper, not part of the port."""
+        self.points[str(point.point_id)] = point
+
+    def _visible(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> list[Point]:
+        return [
+            point
+            for point in self.points.values()
+            if point.tenant_id == tenant_id
+            and (domain is None or point.domain == domain)
+            and (file_id is None or point.file_id == file_id)
+            and (include_inactive or point.is_active)
+        ]
+
+    async def get(
+        self, *, tenant_id: uuid.UUID, point_id: uuid.UUID, with_vectors: bool = False
+    ) -> Point | None:
+        point = self.points.get(str(point_id))
+        if point is None or point.tenant_id != tenant_id:
+            return None
+        return point if with_vectors else point.model_copy(update={"vectors": None})
+
+    async def get_many(
+        self, *, tenant_id: uuid.UUID, point_ids: Sequence[uuid.UUID]
+    ) -> tuple[Point, ...]:
+        wanted = {str(point_id) for point_id in point_ids}
+        return tuple(
+            point
+            for key, point in self.points.items()
+            if key in wanted and point.tenant_id == tenant_id
+        )
+
+    async def list_by_file(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        file_id: uuid.UUID,
+        limit: int,
+        cursor: str | None = None,
+        include_inactive: bool = False,
+    ) -> PointPage:
+        ordered = sorted(
+            self._visible(tenant_id=tenant_id, file_id=file_id, include_inactive=include_inactive),
+            key=lambda point: point.order_id,
+        )
+        if cursor is not None:
+            ordered = [point for point in ordered if point.order_id > float(cursor)]
+
+        page = tuple(ordered[:limit])
+        next_cursor = repr(page[-1].order_id) if len(page) == limit else None
+        return PointPage(points=page, next_cursor=next_cursor)
+
+    async def count(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> int:
+        return len(
+            self._visible(
+                tenant_id=tenant_id,
+                domain=domain,
+                file_id=file_id,
+                include_inactive=include_inactive,
+            )
+        )
+
+    async def keyword_search(
+        self,
+        *,
+        tenant_id: uuid.UUID,
+        query: str,
+        limit: int,
+        cursor: str | None = None,
+        domain: str | None = None,
+        file_id: uuid.UUID | None = None,
+        include_inactive: bool = False,
+    ) -> PointPage:
+        # Whole-token containment, lowercased: an approximation of Qdrant's
+        # full-text index good enough to tell "matched" from "did not", which is
+        # all the service layer branches on. Ranking is not modelled because the
+        # real index does not rank either (ADR-0002).
+        terms = query.lower().split()
+        matches = [
+            point
+            for point in self._visible(
+                tenant_id=tenant_id,
+                domain=domain,
+                file_id=file_id,
+                include_inactive=include_inactive,
+            )
+            if all(term in point.content.lower().split() for term in terms)
+        ]
+        start = int(cursor) if cursor is not None else 0
+        page = tuple(matches[start : start + limit])
+        next_start = start + limit
+        return PointPage(
+            points=page, next_cursor=str(next_start) if next_start < len(matches) else None
+        )
+
+    async def apply_patches(self, *, tenant_id: uuid.UUID, patches: Sequence[PayloadPatch]) -> None:
+        self.patch_batches.append(list(patches))
+        if self.fail_next:
+            self.fail_next = False
+            raise RuntimeError("simulated point repository failure")
+
+        for patch in patches:
+            point = self.points.get(str(patch.point_id))
+            if point is None or point.tenant_id != tenant_id:
+                continue
+            if patch.expected_version is not None and point.version != patch.expected_version:
+                continue
+            # Re-validated rather than `model_copy`d, because a patch payload
+            # carries wire values (an ISO `deleted_at`, a stringified
+            # `previous_chunk_id`) exactly as it would reach Qdrant. Copying
+            # without validation would leave the fake holding a `str` where a
+            # read from real Qdrant returns a `datetime`, and a service bug that
+            # depends on the difference would pass here and fail in production.
+            updated = point.model_dump()
+            updated.update(patch.payload)
+            self.points[str(patch.point_id)] = Point.model_validate(updated)

tests/integration/minio/test_storage.py

@@ -0,0 +1,39 @@
+import uuid
+
+import pytest
+from anyio import CapacityLimiter
+
+from src.application.files.storage_keys import source_file_object_key
+from src.config import MinioSettings
+from src.infrastructure.minio.client import create_client
+from src.infrastructure.minio.storage import MinioObjectStorage
+
+pytestmark = [pytest.mark.integration, pytest.mark.minio, pytest.mark.asyncio]
+
+
+async def test_put_object_stores_bytes_under_server_derived_key(
+    minio_settings: MinioSettings,
+) -> None:
+    client = create_client(minio_settings)
+    storage = MinioObjectStorage(client, bucket=minio_settings.bucket, limiter=CapacityLimiter(2))
+    key = source_file_object_key(uuid.uuid4(), uuid.uuid4())
+    data = b"tenant-scoped, id-addressed bytes"
+
+    await storage.put_object(key=key, data=data, content_type="text/csv")
+
+    stored = client.get_object(minio_settings.bucket, key).read()
+    assert stored == data
+
+
+async def test_put_object_overwrites_existing_object_at_same_key(
+    minio_settings: MinioSettings,
+) -> None:
+    client = create_client(minio_settings)
+    storage = MinioObjectStorage(client, bucket=minio_settings.bucket, limiter=CapacityLimiter(2))
+    key = source_file_object_key(uuid.uuid4(), uuid.uuid4())
+
+    await storage.put_object(key=key, data=b"first", content_type="text/csv")
+    await storage.put_object(key=key, data=b"second", content_type="text/csv")
+
+    stored = client.get_object(minio_settings.bucket, key).read()
+    assert stored == b"second"

tests/integration/postgres/conftest.py

@@ -1,72 +0,0 @@
-from collections.abc import AsyncIterator, Iterator
-
-import pytest
-import pytest_asyncio
-from alembic.command import upgrade
-from alembic.config import Config
-from sqlalchemy.ext.asyncio import AsyncEngine, AsyncSession, async_sessionmaker
-from testcontainers.community.postgres import PostgresContainer
-
-from src.config import PostgresSettings
-from src.infrastructure.postgres.database import create_engine
-
-
-def _alembic_config(database_url: str) -> Config:
-    config = Config("alembic.ini")
-    config.set_main_option("sqlalchemy.url", database_url)
-    return config
-
-
-def _settings_from_url(url: str) -> PostgresSettings:
-    # testcontainers returns postgresql+asyncpg://user:pass@host:port/db ;
-    # PostgresSettings builds its own dsn from parts, so parse the parts back out.
-    without_scheme = url.split("://", 1)[1]
-    creds, hostpart = without_scheme.split("@", 1)
-    user, password = creds.split(":", 1)
-    hostport, db = hostpart.split("/", 1)
-    host, port = hostport.split(":", 1)
-    return PostgresSettings(host=host, port=int(port), user=user, password=password, db=db)
-
-
-@pytest.fixture(scope="session")
-def postgres_container() -> Iterator[PostgresContainer]:
-    with PostgresContainer("postgres:17", driver="asyncpg") as container:
-        yield container
-
-
-@pytest.fixture(scope="session")
-def postgres_url(postgres_container: PostgresContainer) -> str:
-    return postgres_container.get_connection_url()
-
-
-@pytest.fixture(scope="session")
-def migrated_postgres_url(postgres_url: str) -> str:
-    """The container's URL, after Alembic has created the schema on it once."""
-    upgrade(_alembic_config(postgres_url), "head")
-    return postgres_url
-
-
-@pytest_asyncio.fixture(scope="session")
-async def postgres_engine(migrated_postgres_url: str) -> AsyncIterator[AsyncEngine]:
-    engine = create_engine(_settings_from_url(migrated_postgres_url))
-    try:
-        yield engine
-    finally:
-        await engine.dispose()
-
-
-@pytest_asyncio.fixture
-async def db_session(postgres_engine: AsyncEngine) -> AsyncIterator[AsyncSession]:
-    """One session per test, bound to a rolled-back outer transaction.
-
-    Isolates each test's writes (ADR-0016: isolate data per test) without
-    needing a fresh container or unique keys per test.
-    """
-    async with postgres_engine.connect() as connection:
-        outer_transaction = await connection.begin()
-        sessionmaker = async_sessionmaker(
-            bind=connection, expire_on_commit=False, join_transaction_mode="create_savepoint"
-        )
-        async with sessionmaker() as session:
-            yield session
-        await outer_transaction.rollback()