feat(ingestion): add bounded, benchmark-aligned embedding execution

Why:
- Plan 001 Phase 4 needs batched, concurrency-bounded embedding wired into
  the inline upload path, with process-wide capacity/timeout/chunk-limit
  guards (ADR-0017).
- The BM25 analyzer and dense-model config are ported from the `emet`
  evaluation lab, which benchmarked them against the real Farsi corpus
  (bm25-fa-norm-stop; nomic-embed-text-v2-moe at 768-dim; text-embedding-3-large
  at native 3072-dim), closing open items in ADR-0001/ADR-0005.

Changes:
- New: embedding ports, orchestration (embed_chunks), request-bounds
  helpers, and dense/sparse adapters (analyzers.py, bm25.py,
  openai_compatible.py).
- upload.py now parses/chunks/embeds inline behind INGESTION_MAX_CONCURRENCY
  (503), INGESTION_TIMEOUT_SECONDS (504), and the chunk-count ceiling (413);
  every failure path still writes a terminal job row.
- Lifespan builds and warms both dense embedders at startup (fail-soft) and
  creates the sparse embedder and concurrency semaphore.
- httpx moves from dev to main dependencies (adapters use it directly).

Impact:
- Qdrant point upserts are still Phase 5 -- chunks_indexed stays 0.
- New EMBEDDING_* env vars documented in .env.example; safe defaults.

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

.env.example

@@ -45,6 +45,45 @@ INGESTION_EMBED_CONCURRENCY=4
 QDRANT_URL=http://127.0.0.1:6343
 QDRANT_API_KEY=
 
+# 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
 # is silently truncated by the model, so the cap is enforced before embedding.