perf: TopKHeap vector search, evidence-based edge sorting, graph traversal hardening

[MO2k4]

Summary

• Vector search: Add TopKHeap (min-heap) for O(N log K) brute-force search instead of full array sort; add vectorCache for in-memory vector storage; add queryEmbeddingCache LRU (max 50) in VectorManager
• Graph traversal: Add sortEdgesByEvidence() with EDGE_KIND_PRIORITY for deterministic, meaningful edge ordering; cap maxDepth at 20; add FIND_PATH_MAX_VISITED=10000 limit; rewrite findPath to parent-pointer BFS; use separate visitedAncestors/visitedDescendants in type hierarchy
• Graph queries: Use picomatch for safe glob matching in findByQualifiedName; batch-fetch nodes via getNodesByIds in getNodeContext/getFileDependencies/getFileDependents; optimize cycle detection with depsCache and pathSet
• DB queries: Re-add getNodesByIds and getNodesByKinds (now have callers in graph/queries.ts)
• Replace console.log/console.warn with logDebug/logWarn in vector search

Details

TopKHeap

The brute-force vector search previously sorted the entire results array to find top-K matches. TopKHeap maintains a min-heap of size K, yielding O(N log K) instead of O(N log N). For typical searches (K=10, N=10000+), this is a significant improvement.

Evidence-based edge sorting

Edges are now sorted by an evidence priority that reflects how informative each edge kind is for code understanding:

calls: 10, extends: 9, implements: 8, imports: 7, exports: 6,
type_of: 5, returns: 4, instantiates: 3, overrides: 3,
decorates: 2, references: 1, contains: 0

Graph traversal hardening

• maxDepth capped at 20 (was Infinity) to prevent unbounded traversal
• findPath rewritten from recursive DFS to iterative BFS with parent-pointer backtracking and a 10,000-node visit limit
• Type hierarchy uses separate visited sets for ancestors vs descendants to avoid incorrectly skipping nodes

Batch node fetching

getNodesByIds fetches nodes in chunks of 999 (SQLite parameter limit) with a Map return for O(1) lookup. Used in 4 call sites in graph/queries.ts. getNodesByKinds uses an IN clause for filtered queries, used in getProjectGraph.

Test plan

• npm run build compiles without errors
• npm test passes with same baseline (28 pre-existing failures, 326 passing)
• 5 files changed: src/db/queries.ts, src/graph/queries.ts, src/graph/traversal.ts, src/vectors/manager.ts, src/vectors/search.ts
• 1 new export (sortEdgesByEvidence) — has 2 callers in traversal.ts
• getNodesByIds — 4 callers in graph/queries.ts
• getNodesByKinds — 1 caller in graph/queries.ts
• Sentry preserved (2 captureException in db/queries.ts, same as upstream)
• No upstream features removed