fix: remote recall calls setup.sql's gitmem_scar_search (GIT-114)

[nTEG-dev]

What

Recall's store fallback now calls the function setup.sql actually creates (GIT-114).

• Default prefix (gitmem_): calls gitmem_scar_search with similarity_threshold. If the store answers PGRST202 (no such function), it falls back to match_gitmem_learnings_weighted with match_threshold. The first function that exists is remembered for the process; any other error is a real failure and isn't retried under another name.
• Non-default prefix: only match_<table>_weighted. gitmem_scar_search reads gitmem_learnings by name, so under orchestra_ it would search the wrong table.
• Missing columns: gitmem_scar_search returns only id, title, description, severity, scar_type, counter_arguments, decay_multiplier and similarity. The fields recall renders (learning_type, applies_when, why_this_matters, action_protocol, …) are fetched for the matched ids in one select. If that select fails, the matched scars are still returned without them.
• Removed the dead semanticSearch (match_gitmem_learnings, no callers).
• Drift analyzer extended to RPC names (scanRpcNames, parseFunctions). Every RPC src/ can call on the customer's store must be defined by v1.8.0's setup.sql or allowlisted by ticket: match_gitmem_learnings_weighted (GIT-114, fallback only) and match_transcript_chunks (GIT-106, dev only). The licensing RPC is a fixed URL to another project and is classified external.

Why

When the local vector index isn't ready, recall asks the store. That covers every server start until warm-up finishes, a new or empty store, a failed load, and GITMEM_SEARCH_MODE=remote. GIT-93 named the RPC after the function on nTEG's store; no setup.sql defines it. So every customer's cold recall got PGRST202 and returned nothing.

Decision to review

The fallback keeps nTEG's store working. The ticket says to use gitmem_scar_search. Per the GIT-93 survey, nTEG's store has match_*_weighted and no *_scar_search. So instead of switching outright, the code tries the setup.sql function first and falls back only on PGRST202. I didn't probe production to confirm which functions it has (rule: never touch production); the fallback makes that unnecessary.

How to test

Venue (vujndhcumgroztfkkdus, reset to v1.8.0:schema/setup.sql; production not touched). flow has two new probes, each on a fresh server: recall with GITMEM_SEARCH_MODE=remote, and recall called immediately on a cold server.

	origin/main 1.11.0	This branch
POST /rpc/match_gitmem_learnings_weighted	404 (PGRST202)	not called
Remote-mode recall	no scar; exit 1	seeded scar via gitmem_scar_search
Cold-start recall	no scar	seeded scar; took the Local cache not ready fallback
Exit	1	0, no expected failures seen

Unit tests: 1365/1365 (+9). Four cover the fallback, no retry on a real error, the follow-up select and tolerating its failure; five cover RPC drift, including a planted-drift self-test.

Existing tests changed

scar-search-rpc-name.test.ts:

• "tracks the prefix rather than hardcoding one deployment's table" asserted match_gitmem_learnings_weighted for the default prefix, which is the bug itself. It now asserts gitmem_scar_search with similarity_threshold.
• "semanticSearch calls match_" was removed along with the dead function.
• The three orchestra_ tests are unchanged: nTEG's prefix still calls match_orchestra_learnings_weighted.

Checklist

• Unit tests pass (1365)
• Venue: fails on main, passes on the branch (v1.8.0)
• No schema change

🤖 Generated with Claude Code