Perf: reduce initial full-index lookup and C# extraction costs

[Widthdom]

Summary

Initial full indexing spends avoidable work sorting reference-source candidates and repeatedly scanning and allocating C# declaration text. These changes reduce that work while preserving existing extraction records and recovering previously missed constructors.

• Rank each canonical, display, and legacy ASCII source lookup through temporary indexes, then compare at most three containing candidates. This applies to every language using the fresh bulk writer.
• Stop confirmed C# method lookahead as soon as body tokens reject an accessor. Preserve delayed and attributed accessors, recover affected constructors, and advance the shared C#/Razor/Blazor/CSHTML extractor contract to 19 so old stamped rows are refreshed.
• Bound static-lambda declaration probes and skip impossible method-prefix regex matches. Cover the shared C# language keys, dense declarations, and typed, tuple, and function-pointer lambda controls.

Database layout, transaction boundaries, cancellation, and rollback behavior are unchanged.

Performance evidence

A fixed 1,554-file snapshot at 0d39d0658 was indexed into a new database for each run on macOS ARM64, Release .NET 8, with --parallelism 2 --memory-trace.

Run	Elapsed	Total managed allocations
Before this PR	76.0 s	18.46 GB
Optimized measurement	68.9 s	7.83 GB
Optimized repeat (2602d764e)	69.6 s	7.84 GB

These are individual observations, not a general speed guarantee or a measurement of peak resident-memory reduction. All runs completed every file without warnings or extraction errors. All pre-existing symbol identities and extraction metadata were preserved; seven constructors were recovered, with the corresponding declaration-reference and resolution updates. The repeat at 2602d764e matched the optimized measurement across all 58,861 symbol records and 555,136 references after normalizing internal database IDs, including reference contexts and source/target symbol records.

Validation

• Release symbol/reference extraction and fresh-writer suites: .NET 8 passed 2,569 tests (2 existing skips); .NET 9 passed 2,539 (32 existing skips).
• After the final optimization, the targeted regex, fresh-writer, C# prepass/static-interface, unchanged-file reuse, and old-index upgrade suite passed 199 tests on .NET 8 (1 existing skip) and 196 on .NET 9 (4 existing skips).
• Fresh reference resolution, authoritative fresh writer, initial full-index, and reference-source regressions passed 68 tests on .NET 8 (2 existing skips) and 67 on .NET 9 (3 existing skips). This includes the SQL-shape assertion updated for the projected containment rank; its old expectation was reproduced as a failure on both frameworks before the correction.
• Builds, focused dotnet format --verify-no-changes, git diff --check, and changelog validation passed.
• Root and workspace index checks passed at the final commit; database integrity returned integrity_ok.
• Independent adversarial review of the performance changes and SQL-assertion correction: No blocking/actionable issues found. The reviewer also passed 40 focused tests, reran the corrected SQL-shape regression on .NET 9, and verified ordinary full-scan recovery of an unstamped legacy C# index using a Razor fixture.

Documentation and changelog

Updated docs/initial-index-performance.md and TESTING_GUIDE.md, including benchmark limits and regression coverage. Added bilingual fragments:

• changelog.d/unreleased/+initial-index-ranked-source-probes.changed.md
• changelog.d/unreleased/+initial-index-method-lookahead.changed.md
• changelog.d/unreleased/+initial-index-declaration-probes.changed.md

This work is not tied to an issue and does not auto-close any issues. A follow-up candidate is profiling the remaining C# prepass cost on larger repositories.

Collaborator checklist

• The change follows the existing code style.
• The change includes tests when behavior changes.
• The change includes a changelog fragment when user-visible behavior changes.
• Documentation has been updated when needed.