perf(nav): reduce code-rank path canonicalization and membership overhead

[Copilot]

Each rankCode run was calling realpathSync on every reference file path — including repeats — and using Array.includes for O(n²) target deduplication in createService.

Changes

packages/nav/src/service.ts

• Set-based deduplication: fileNames.includes(resolved) → fileNameSet.has(resolved), making target membership O(1) after setup
• createCachedBoundaryChecker(boundary) (new export): returns a (path: string) => boolean closure backed by a Map cache — each distinct path hits realpathSync at most once per checker lifetime

packages/nav/src/code-rank/rank.ts

• Creates one isWithinBoundary checker per rankCode call; passes it into collectReferenceStats as a plain predicate
• collectReferenceStats signature drops the boundary object in favor of (filePath: string) => boolean — simpler and cache-aware

// Before: realpathSync called on every reference.fileName
if (!isPathWithinWorkspaceBoundary(boundary, reference.fileName)) { ... }

// After: cached per rankCode run
const isWithinBoundary = createCachedBoundaryChecker(boundary);
// ...
if (!isWithinBoundary(reference.fileName)) { ... }

Benchmark

• bench/suites/fixtures.ts: adds createCodeRankFixture — hub-and-spoke TS workspace (one hub with N×M exports, N spoke files each calling their slice)
• bench/suite.ts: new nav/code-rank group benchmarking 50-file × 4-export fixture

Tests

• service.test.ts: unit tests for createCachedBoundaryChecker — inside paths, outside paths, symlinks resolving outside the boundary
• code-rank.test.ts: symlinked files resolving outside the git boundary are excluded from referencingFiles; symbol deduplication produces no duplicate entries

• Fixes perf(nav): reduce code-rank path and membership overhead #9