docs: Cluster A (count drift) + Cluster B (deleted-module refs) sweep

fix(upload-extract): status parity on LLM-fail path + docs updated

feat(upload): multi-pass per-section extraction + heuristic-floor expansion (KI-332)

feat(single_brain): wire view_context → ACTIVE POLICY DIVE-IN block (KI-330)

fix(upload-extract): hash-cache short-circuit now matches scorecard endpoint

fix(upload-extract): read sc.grade not sc.overall_grade

fix(upload-extract): use _catalogue_scorecard primary path for status

fix(upload-extract): status completeness now matches scorecard endpoint

chore: surface llm_used + llm_response_chars on extraction-status endpoint

feat(upload-extract): Tier-2 content-hash cache — same PDF, no re-extraction

feat(upload-extract): Tier-1 Gemini stability + new ADR-044 + README rewrites

fix(upload-extract): drop responseMimeType=application/json — was producing unparseable payloads

fix(upload-extract): LLM output now merges INTO heuristic record (not replace)

fix(upload-extract): Gemini-primary done right — max_tokens 2048→8192 to stop truncation

revert(upload-extract): roll back Gemini-primary switch — surfaced as live failures

feat(upload-extract): switch primary LLM from NIM chain → Gemini 2.5-flash (JSON mode)

hotfix: import Header from fastapi — caused RUNTIME_ERROR on 24349be

feat(upload): backfill LLM extraction for old uploads + admin re-extract endpoint

fix(upload): missing asyncio import — module-level Lock() fired NameError on import

feat(upload): wait-for-extraction flow + status endpoint + extended voice guard

feat: uploaded-PDF parity with catalogued 148 (ADR-044)

feat!: remove cross-session profile recall (ADR-043) — net −3700 LOC

fix(privacy v3/v4): two-fact recall gate + extended same-turn extractors (closes ADR-042 follow-ups)

fix(privacy v2): same-turn age contradiction in apply_pending_recall

fix: 4 hardening fixes — sticky-retry, canned-msg copy, admin refresh wiring, recall privacy

#41 + #43: fix family-history re-ask + unverifiable-UIN warning

#47: UIN-based net-new dedup for user-uploaded policy PDFs

#45: persist chat-only profiles so returning-user recall actually works

#25/#26: returning-user recall any-turn + mid-conversation state recovery

feat(#31): deterministic profile-aware {strengths,caveat} policy summary on all 3 surfaces

fix: deductible gating (#29), existing-cover steering + dangling-turn guard (#30), compare-modal full reviews (#32)

fix(batch): voice interim, TTS truncation, RULE2.5 re-ask, faithfulness regression, markdown, compare-reviews, policy-name

fix(brain): don't re-ask recalled slots after returning-user confirm

fix(brain): wire returning-user-by-name recall into single_brain (ADR-041/KI-196)

fix(voice/brain): remove hold-space, fix silent TTS + policy-Q&A refusal

feat(#77 / #52-residual-5): admin prune for persisted uploaded docs

fix(#61): named-policy bypass not gated on intent (single_brain hardcodes it)

fix(#61): resolve the version-specific chunk-bearing card, not the empty base

fix(#61): named-policy Q&A bypasses the recommendation profile-gate

Revert "Revert "feat(#52): PDF upload → persistent JSON + marketplace card + restart-survival""

Revert "feat(#52): PDF upload → persistent JSON + marketplace card + restart-survival"

feat(#52): PDF upload → persistent JSON + marketplace card + restart-survival

fix(#63): header premium pill is the TYPICAL cohort range, labelled as such

fix(#64): preserve EXACT budget ₹ — slider shows what the user said

fix(#75): compare_policies curated fallback — no scorecard surface 404s

fix(#55+#56+#53/#54-engine): TTS full natural readout + voice warm-stream/pre-roll

fix(#60+#62): bulk-scorecard curated fallback + stop internal scoring jargon leaking

fix(#40+voice+upload): SSOT scorecard parity + Sarvam 30s-truncation + uploaded-PDF retrieval gate

fix(pricing+scoring): rebuild premium engine to 0 logical violations + classification, faithfulness, links, writing

fix(reco): cap 3, exclude non-extracted policies, unify premium, ask PED