feat(ux): KI-056 — natural-language acknowledgers + spouse capture + Q3 paren consistency

fix(deploy): KI-055 — drop rag/vectors symlink from repo (HF Space BUILD_ERROR)

refactor + feat: KI-050b/053/054 — rename completeness + eval speed + tax questions

feat(admin): KI-052 — 3-tab admin panel (Profiles, Performance, LLM Chain)

fix(docker): KI-051 — update COPY paths after bucket reorg (data → 40-data, docs → 70-docs). Previous attempts hit Edit's read-first rule and didn't actually persist. Live HF Space rebuild was about to fail on COPY data and COPY docs since both source dirs no longer exist at the renamed paths.

refactor: KI-050 — complete data/ → 40-data/ rename across all Python refs

perf(retrieval): KI-049 — wider top-k for table-cell questions

feat(admin): KI-048 — /api/admin/profiles + /api/admin/performance

refactor: KI-047 — bucket reorg (Option A safe subset)

fix: KI-046 — explicit refusal on adversarial/fanciful out-of-corpus questions

feat: KI-045 — natural-conversation classifier in fact-find

feat: KI-044 PCM pre-roll via AudioWorklet + 11 folder READMEs

fix(voice): KI-041/042 — VAD sensitivity + Live OFF by default + dots on first turn

feat: KI-039 + KI-040 — single Clear-chat (keep profile) + named profiles

fix(voice): KI-039 — single Clear-chat button (full reset)

fix: KI-036/037/038 — greeting flow, strict paraphrase verifier, waiting dots

docs: full sweep — fix stale model names + 48→62 fields + merge SUBMISSION

perf: KI-034 retrieval cache + KI-035 reorder FAST_BRAIN_CHAIN

fix: KI-033 — move fact-find normalizer + profile extractor to NIM chain

feat(voice): KI-032 — LLM-paraphrase fact-find questions (Option B + verifier)

fix(voice): KI-030 — bot TTS now plays via in-DOM audio so barge-in works

fix(voice): KI-029 — make push-to-talk button visibly findable

feat(voice): KI-028 — clickable Live on/off toggle (green ↔ red)

feat(voice): KI-027 — Live is the only voice mode; one-shot PTT pauses it

fix: KI-026 — voice mode mutual exclusion (Live blocks legacy mic paths)

perf: KI-025 — load-balance brain calls across NIM Qwen ↔ Groq Llama

perf(eval): KI-024 — parallelize 96-Q gold eval with --workers semaphore

fix: KI-023 — word-boundary intent triggers + KI-018 regression test

fix: KI-018..022 — routing, profile telemetry, NIM timeouts, judge JSON, chat clear

feat(rag): tools/ingest_kb_summaries.py — embed kb/policies/*.md (KI-019)

docs: ADR-025 (single-Chroma-collection rationale) + KI-018/019 backlog

feat(reviews): KI-017 — enriched per-facet review chunks (10 → 60)

docs: 2026-05-14 quality sprint scoreboard

fix(persona): KI-013 reinforcement — demographic-aware recommendation rules

fix(fact-find): KI-015 — invite corrections in readback before recommending

fix(fact-find): KI-012/013/014 — flow corrections from user testing

fix(orchestrator): KI-012 critical — transition to free-form after fact-find

fix(observability): KI-001..006 — log silent failures + fail-CLOSED judge

fix(fact-find): keyword fast-path + re-ask cap (KI-011 critical)

feat(audit): 100-persona multi-scenario simulation framework

fix(ux): cold-start retries + retry-intent handling + insurer-prefixed citations

chore: print absolute output path (PDF is outside the repo now)

chore: build_readme_pdf.py outputs to ~/Desktop, not the repo

docs: add tools/build_readme_pdf.py (PDF gitignored — HF rejects binaries)

feat(voice): full-duplex Live conversation mode with VAD barge-in (D-025)

fix(fact-find): LLM-normalize answers + reject non-answers + safe readback

fix(stt): transcode browser webm/opus → wav before Sarvam (D-024)

fix(readme): trim short_description to <=60 chars (HF YAML constraint)

docs: reorganize into 8 buckets + 24 ADRs + master README (D-023)

feat: live time-since / time-until-next-probe counter in admin panel