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>

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]