Perf: reduce initial full-index extraction and reference work

[Widthdom]

Summary

Initial full indexing repeatedly evaluates impossible declaration regexes, resolves identical reference sources, and scans unrelated same-name symbols or earlier occurrences on dense lines. These changes reduce that work while preserving the logical index contents and existing resolution rules.

• Reject impossible C# declaration suffixes and add audited punctuation gates for C#, Java, Kotlin, C/C++, JavaScript and TypeScript, including C# aliases.
• Resolve repeated fresh reference-source identities once per qualifying insertion batch, and seek candidates through the existing file/name and container/name indexes. These writer improvements apply across languages and preserve scoped and retained-graph refreshes.
• Scan C# lambda captures once per body, reuse structured scope keys, and use proven recorded reference positions to avoid repeated line scans. Preserve Unicode/escaped identifiers, overlapping matches, ties and trimmed-column fallbacks.

Performance and compatibility

The benchmark uses the fixed a33c5a8eb source snapshot: 1,577 files, 59,387 symbols and 562,615 references. Each version was measured twice on macOS ARM64, Release .NET 8, with a new database and --parallelism 2 --memory-trace; heavy tests were stopped during measurement.

Independent comparison	Mean elapsed time	Change
Declaration suffixes, shared source lookups and lambda captures	72.9 s → 62.3 s	About 15% lower
Punctuation gates, scoped candidate seeks and reference positions, relative to 8533dff4f	60.2 s → 44.7 s	About 26% lower

Each percentage describes its own measurement session. The second comparison reduced total managed allocations from 8.09–8.10 GB to 7.48 GB (about 8%); the first comparison increased allocations from 7.93–7.96 GB to 8.10 GB.

All runs completed without warnings, errors or extraction cap hits. Logical file, chunk, symbol, reference-line, issue, reference and resolution-candidate records match after normalizing generated IDs and indexing timestamps. The fixture is C#-heavy; these measurements are not a general speed guarantee. Scoped candidate seeks add a small fixed cost for unique names or names already confined to eligible scopes; the measured tradeoff is documented.

Validation

• Whole solution at 8d90c292b: dotnet test CodeIndex.sln -c Release --no-restore -p:UseSharedCompilation=false --settings tests/CodeIndex.Tests/CodeIndex.Tests.runsettings --blame-crash --blame-hang --blame-hang-timeout 5m.
• .NET 8: 12,682 passed, 9 skipped, 0 failed. .NET 9: 12,188 passed, 435 skipped, 0 failed.
• Cross-language output-parity tests and regression budgets for SQLite VM instructions, regex attempts, occurrence searches and allocations; no new wall-clock thresholds.
• Bilingual changelog fragment validation and git diff --check passed.
• Formatting correction 0d3027832: ./dev.sh lint passed (the same script invoked by CI's make lint), and both affected test classes passed all 57 cases on each of .NET 8 and .NET 9. Independent review confirmed that only test whitespace changed; production code and test semantics are unchanged.
• Independent adversarial review of origin/main..HEAD: No blocking/actionable issues found.
• Canonical index and workspace checks passed at the final commit: 1,585 matching files, both workspace members healthy.

Documentation and changelog

Updated TESTING_GUIDE.md and the English/Japanese measurement and compatibility notes in docs/initial-index-performance.md.

Bilingual fragments:

• changelog.d/unreleased/+initial-index-confirmation-suffixes.changed.md
• changelog.d/unreleased/+initial-index-shared-reference-sources.changed.md
• changelog.d/unreleased/+initial-index-lambda-captures.changed.md
• changelog.d/unreleased/+initial-index-scope-candidate-seeks.changed.md
• changelog.d/unreleased/+initial-index-declaration-punctuation.changed.md
• changelog.d/unreleased/+initial-index-reference-occurrences.changed.md

No issue auto-close references or separate follow-up candidates are included.