perf(db): batch node lookups, fix insertNode cache, auto-ANALYZE after writes

[andreinknv]

Summary

Three DB-layer improvements bundled in one PR. All low-risk, high-leverage.

1. Batch getNodesByIds — fix N+1 in graph traversal

QueryBuilder.getNodesByIds(ids[]) returns Map<id, Node> in one SQL round-trip (chunked at 500 for SQLite param-limit safety; cache-aware so already-cached entries are served from memory and only the misses hit SQL).

Replaces 9 N+1 loops in src/graph/traversal.ts:

• getCallersRecursive / getCalleesRecursive
• getTypeAncestors / getTypeDescendants
• findUsages
• getImpactRecursive (both inner loops: contains-children + dependents)
• findPath BFS frontier
• traverseBFS / traverseDFS neighbor expansion
• getChildren

Each previously did getNodeById per edge inside a loop; for a function with N callers at depth D, that was N^D point reads. Now: one IN-list query per traversal step. Expected 10–50× speedup on deep / fan-out-heavy traversals (impact analysis on popular utilities, call graphs of central modules).

2. insertNode cache invalidation — fix correctness bug

QueryBuilder.insertNode uses INSERT OR REPLACE INTO nodes. The LRU nodeCache was invalidated by updateNode and deleteNode but NOT by insertNode, so a re-indexed node (replacing a cached row) would still serve the pre-replace version on next getNodeById until LRU eviction pushed it out.

Now invalidates nodeCache.delete(node.id) at the top of insertNode (after validation, before SQL — so failed-validation early-returns don't churn the cache).

3. Auto-maintenance after bulk writes

DatabaseConnection.runMaintenance() runs:

• PRAGMA optimize — incremental ANALYZE; only re-analyzes tables whose row counts changed materially since the last ANALYZE. Without it, the SQLite query planner has zero statistics on freshly-bulk-loaded tables and can pick wrong indexes.
• PRAGMA wal_checkpoint(PASSIVE) — fold WAL pages back into the main DB file so the WAL doesn't grow unboundedly between automatic checkpoints (which fire at 1000 pages).

Both non-blocking and silently swallowed on failure — best-effort, never load-bearing. Wired into indexAll (when files were indexed) and sync (when files changed). No-op when nothing happened.

Files changed

File	Change
src/db/queries.ts	Add getNodesByIds; invalidate cache in insertNode
src/db/index.ts	Add runMaintenance() helper
src/graph/traversal.ts	Replace 9 N+1 loops with batch lookups
src/index.ts	Call runMaintenance after indexAll / sync
__tests__/db-perf.test.ts (NEW)	9 regression tests covering batch lookup, chunking, cache hits, cache invalidation, maintenance no-throw on closed DB

Test plan

• npm test: 388/389 pass on macOS (one pre-existing fs.watch flake)
• npx tsc --noEmit clean
• Independent reviewer pass before pushing — APPROVE; one info finding addressed (getChildren was the 9th N+1 site, now batched too)

🤖 Generated with Claude Code