perf(search): trigram index for substring name search

[gtarpenning]

Summary

name LIKE '%x%' cannot use a B-tree, so every substring lookup scans all nodes. The prompt hook is a fresh process per prompt and runs several of these each time, so the scan cost is paid on every prompt. This adds a trigram FTS5 index (nodes_tri) and uses it to narrow the candidates.

Independent of the cross-process memo PR; either can merge first. #2115 and #2260 touch nearby code in queries.ts.

What changed

• Migration 12 creates nodes_tri(name, qualified_name), an external-content FTS5 table over nodes with tokenize='trigram', plus insert/delete/update triggers beside the nodes_fts ones. Fresh databases get it from schema.sql.
• findNodesByNameSubstring and searchNodesLike add rowid IN (SELECT rowid FROM nodes_tri ...) when the token is 3+ characters and has no _ or %. The original LIKE stays in the query, so the index only narrows candidates and result sets are unchanged. Shorter or wildcard tokens take the old path.
• Bulk loads drop and rebuild the trigram index with the FTS one. The trigger list is now read from schema.sql instead of a hard-coded name list, so the heal after an interrupted bulk load covers both.
• Without FTS5 ([Bug] v1.5.0 requires FTS5 which is missing in official Node.js builds #1532), migration 12 is skipped, the table is absent, and queries use LIKE. A later bulk load on an FTS5 build creates the table.

Measured

Warm query time on a large TS/Go/Python monorepo index (about 1.5 GB, 350k nodes, 1.2M edges):

Query	Before	After
Substring lookup (name LIKE)	40-80 ms	1-5 ms
Name or qualified-name lookup	55-85 ms	2-11 ms

Result sets were identical for every token tried (compared by rowid). Tokens under 3 characters stay on the scan path, where the index is no faster.

• Size: nodes_tri adds about 2.5-3.7% to the database (55 MB on 1.5 GB).
• Migration: about 1.5 s on that index; about 9 s per 1M synthetic nodes.
• Write cost: a 300-file sync took 64.9 s with the triggers and 59-63 s without, within run-to-run noise because resolution dominates.

Compatibility

• Older binaries open a v12 database fine: the triggers are SQLite-level, so their writes keep nodes_tri current, and they never query it.
• A database migrated on a build without FTS5 has no nodes_tri until the next bulk load on a build that has it; substring search stays correct throughout.

Testing

• npx tsc --noEmit
• npx vitest run __tests__/trigram-substring.test.ts __tests__/fts5-fallback.test.ts: results match a LIKE scan; a node removed from nodes_tri alone disappears from long-token results, so the index is consulted (the test fails with the prefilter reverted); a v11 database without FTS5 opens and falls back to LIKE.
• npx vitest run: 5748 passed, 20 failed. The 20 are in daemon-older-version, daemon-registry and index-command, and fail the same way on unmodified main.

🤖 Generated with Claude Code