Codify live retrieval RPCs, reclaim dead vector indexes, unify schema health

docs/process-hardening.md

@@ -49,3 +49,14 @@ This document turns the current process review into phased, durable repo practic
 - Process scripts do not commit, push, deploy, mutate Supabase data, or run dependency updates.
 - `npm run check:indexing` includes local OCR prerequisites (`fitz`/PyMuPDF, `pytesseract`, and the Tesseract binary). A failure at that prerequisite step is local machine setup debt, not evidence that indexed production data or search behavior regressed.
 - Supabase performance-advisor `unused_index` INFO items are monitored, not automatically fixed. Do not remove search/RAG support indexes until live query evidence, local explain/verification, and rollback planning show the index is safe to drop.
+
+## Retrieval RPC drift & indexing hygiene (2026-07-01)
+
+- The four app-path hybrid retrieval RPCs (`match_document_chunks_hybrid`, `match_document_embedding_fields_hybrid`, `match_document_index_units_hybrid`, `match_document_memory_cards_hybrid` + its `_v2` core) had live-only performance fixes applied via raw SQL that were never captured in migrations, so a `supabase db reset` / branch DB reproduced the slow pre-fix shapes. Migration `20260701140631_codify_live_retrieval_rpcs` codifies the live definitions (validated byte-equivalent to live via whitespace-stripped `pg_get_functiondef` md5 before applying — a confirmed no-op on live), and `supabase/schema.sql` was reconciled to match. A clean replay now reproduces production retrieval.
+- **Rule: never change a retrieval RPC (or any function) on the live project with raw `execute_sql`.** Go through a committed migration plus a `supabase/schema.sql` update. Raw-SQL edits are exactly how this drift accumulated.
+- `search_schema_health()` runs an execution smoke (invokes each hybrid RPC with a zero vector) that surfaces through `npm run check:indexing`; it fails if an RPC regresses to an error state (e.g. the historical `42702` ambiguous-id break). This is the standing guard against the original bug class.
+- **Migration `20260702014803_drop_legacy_vector_indexes` (applied 2026-07-02 with explicit user approval)** reclaimed ~4.4 GB of dead/duplicate vector indexes (embedding_fields ivfflat 3.66 GB @ 8 scans, chunks ivfflat 610 MB, index_units HNSW 640 MB @ 0 scans, plus dead btrees). Verified post-apply: all targets gone, `detect_legacy_ivfflat_indexes()` empty, DB 13 GB -> 8.6 GB, `search_schema_health()` ok. The documented follow-ups are done: `supabase/schema.sql` now declares the live-kept embedding-fields indexes (`owner_id_idx`, `source_chunk_id_idx`, `search_tsv_chunk_gin_idx`, `owner_document_created_idx`, `meta_rag_indexing_version_idx`) instead of the dropped ones, no longer creates the index_units HNSW index, and `tests/supabase-schema.test.ts` asserts the new shape. There is intentionally no HNSW index on `document_index_units.embedding` — re-add only if that RPC gains a vector-first candidate path.
+- **`search_schema_health()` two-lineage divergence: RESOLVED** by `20260702021604_reconcile_search_schema_health_superset` (applied live 2026-07-02, verified `ok:true`). The single definition now carries the comprehensive signature checks (incl. `match_document_memory_cards_hybrid_v2`), the full 22-entry required-index list (post-drop: no index_units HNSW, memory_cards HNSW added; every entry verified present live before shipping), the legacy-ivfflat report, AND the hybrid-RPC execution smoke. schema.sql matches exactly (the migration is extracted from it).
+- **Known follow-up debts (documented, not actioned):**
+ - Live migration history has duplicate-version churn (two each of `api_rate_limits`, `audit_logs`, `rag_queries_retention`, `audit_logs_service_role_policy`, `indexing_reliability_recovery`) from the same raw-apply habit. Do not rewrite history; treat as a caution for future applies.
+ - Auth server is capped at 10 absolute DB connections (Supabase advisor); switch to percentage-based allocation in the dashboard before scaling instance size (not settable via SQL/MCP).

supabase/migrations/20260701000000_document_label_taxonomy_v2.sql

@@ -0,0 +1,38 @@
+-- Expand document labels for smart search scope and normalize the known ECT duplicate.
+
+alter table public.document_labels
+ drop constraint if exists document_labels_label_type_check;
+
+alter table public.document_labels
+ add constraint document_labels_label_type_check
+ check (label_type in (
+ 'site',
+ 'topic',
+ 'document_type',
+ 'medication',
+ 'risk',
+ 'setting',
+ 'workflow',
+ 'population',
+ 'service',
+ 'clinical_action',
+ 'care_phase',
+ 'document_intent',
+ 'content_feature',
+ 'custom'
+ ));
+
+delete from public.document_labels duplicate
+using public.document_labels canonical
+where duplicate.label_type = 'topic'
+ and canonical.label_type = 'topic'
+ and duplicate.label = 'electroconvulsive therapy'
+ and canonical.label = 'electroconvulsive-therapy'
+ and duplicate.document_id = canonical.document_id
+ and duplicate.source = canonical.source
+ and duplicate.id <> canonical.id;
+
+update public.document_labels
+set label = 'electroconvulsive-therapy'
+where label_type = 'topic'
+ and label = 'electroconvulsive therapy';