test(db): assert vector via ::text cast, not the adapter return type

refactor(scripts): extract shared _common.py connection + embedder (phase 5b)

perf(db): collapse tagged_lookup N+1 and batch ingest (phase 4b)

test(db): add Postgres+pgvector integration harness (phase 4a)

fix(metering): surface `available` in /usage so the badge matches the 429

feat(metering): wire enforcement into endpoints + GET /usage + /health

feat(metering): quota core — identity, daily counters, ledger (flag off)

feat(metering): thread token usage through the pipeline into QueryResponse

feat(metering): capture real token usage from provider APIs

docs(plan): 3.14 measured — both hyde prompt variants die; the writer lacks the rulebook's defined vocabulary

feat(api): stream /query over SSE — token/restart/final contract (#79)

feat(health): expose hyde config — model from the provider, skip flag from settings (#78)

feat(scripts): 2.2 flip gate — hyde_model ALIVE, gemma-4-31b writes passages as well as the 120B

refactor(rag): strip the dead semantic cache (#77)

feat(scripts): 2.3 flip gate — semantic cache DIES on a real adversarial pair

feat(scripts): 2.1 flip gate probe — skip_hyde_when_routed ALIVE 3/3 claims

refactor(rag): strip the dead concise_reasoning flag (#76)

refactor(rag): remove the dead relaxed flag, make the routing gate one function (#75)

feat(rag): relaxed routing threshold behind a flag — 3W/0L on retrieval (3.11.1) (#74)

fix(probe): the instruments were lying — make them measure production (Fase 6) (#73)

perf(prompt): cap the Reasoning section on simple lookups (2.6, flag off) (#71)

perf(rag): stop paying for HyDE on routed queries + dedicated hyde_model (#70)

feat(rag): semantic answer cache for non-hard queries (plan 2.3, flag off) (#69)

fix(rag): strip possessive so card names in the possessive detect (eval-029) (#72)

fix(rag): hard-routing timeout + bare champion detection (eval-024) + eval-027 data fix (#68)

feat(rag): flip hard_query_routing and keyword_family_extra defaults (4.2+4.3, 3.5) (#63)

fix(rag): hard provider decoupled from the main provider — routing was dead in prod (#59)

feat(rag): hard-query routing — thinking model + full-rulebook stuffing (4.2+4.3, flag off) (#58)

feat(retrieval): keyword family completion beyond top_k (3.5, flag-gated) (#57)

feat(ingest): contextual-retrieval machinery + version-scoped chunk ids (#56)

fix(eval): strict rule-code lineage — recall no longer paper-hits by family (#55)

fix(ingest): match family-sentence rule titles, capture first bold only (#54)

feat(config): cap generation output tokens (max_output_tokens=1024) (#51)

feat(prompt): add few-shot chaining examples to system prompt (v6) (#50)

fix(cache): never cache degraded responses (#49)

feat(rag): fine rule-title sections + reserved keyword slot (hard-bucket retrieval) (#48)

feat(config): enable cross-encoder reranker by default (#46)

fix(config): replace retired gemini-2.0-flash default with gemini-flash-lite-latest (#45)

feat(rag): add multi-card reasoning scaffold threaded via extra_system (#43)

feat(rag): wire cross-encoder reranker behind enable_reranker (#42)

feat(rag): implement GeminiProvider.hyde() for dual-arm retrieval (#41)

feat(frontend): detect HF Space cold start and auto-retry once awake

fix(generation): retry Gemini 429s to match openai-compat path

fix(pipeline): guard empty Answer section with retry and fallback

feat(api): re-resolve corpus_version on demand when empty at startup

refactor(db): extract resolve_corpus_version, share with startup

refactor(pipeline): pass corpus_version explicitly, stop mutating settings

fix(pipeline): whole-word match for exact-card confidence

feat(ratelimit): allow shared Redis storage for counters

feat(config): require proxy_shared_secret in production