Reduce initial full-index work on large codebases

[Widthdom]

Summary

Initial full indexing spends substantial work repeating declaration probes, receiver-scope analysis and reference-source SQL for large repositories. This PR reduces that work across 10 commits:

• Specialize fresh reference-source lookups when canonical names are sufficient, preserving display/legacy fallbacks. Bind fresh reference flags as SQL zero literals (12 parameters per row instead of 14; native batches increase from 36 to 42 rows) and skip ineligible known-zero mutual-recursion rows.
• Prune impossible C# declaration, capture and enum probes; defer cumulative lookahead copies, scan appended semicolon state once, and cache only requested callable receiver scopes. C#, Razor, Blazor and CSHTML share these improvements.
• Skip ordinary code spans during C#, Kotlin and Scala structural masking while preserving nested strings, interpolation, comments and source coordinates.

Names, extracted records, resolution rules, database layout, transaction boundaries and rollback behavior remain unchanged.

Validation

• dotnet test CodeIndex.sln -c Release -p:UseSharedCompilation=false: .NET 8 12,733 passed / 9 skipped; .NET 9 12,239 passed / 435 skipped. Zero failures, errors, timeouts or aborted tests.
• Full symbol/reference extractor suites: .NET 8 2,599 passed; .NET 9 2,570 passed / 29 skipped. Deterministic work/allocation budgets and complete-record comparisons cover the shared language paths.
• Both final empty-database runs over the fixed 1,577-file snapshot match the saved pre-change output in logical files, chunks, 59,387 symbols, 562,615 references, reference lines, issues, candidates (including complete per-reference candidate sets), hotspot records, schema and stable metadata after normalizing generated IDs/indexing times. SQLite integrity checks pass. FTS posting payloads were not compared.
• Changed-file formatting, changelog validation and diff checks pass. The canonical root index and both workspace members are healthy at the final HEAD.
• Independent adversarial review of origin/main..HEAD: No blocking/actionable issues found.

Measurements

On the same C#-heavy source snapshot, the first two batches each had their own alternating Release .NET 8 before/after measurements: mean 43.6 → 37.7 seconds (~13.6%) and 37.40 → 34.51 seconds (~7.7%). These are separate comparisons, not a combined final-branch speedup.

The final batch has standalone Debug .NET 8 observations of 39.014 / 40.010 seconds, with no matched Release before/after timing; no percentage improvement is claimed for it. Details and allocation observations are in docs/initial-index-performance.md.

Documentation and changelog

Updated TESTING_GUIDE.md and the bilingual performance documentation. Added these bilingual fragments:

• changelog.d/unreleased/+fresh-canonical-source-probes.changed.md
• changelog.d/unreleased/+csharp-receiver-scope-reuse.changed.md
• changelog.d/unreleased/+csharp-declaration-prefix-pruning.changed.md
• changelog.d/unreleased/+fresh-reference-flag-work.changed.md
• changelog.d/unreleased/+csharp-reference-probe-pruning.changed.md
• changelog.d/unreleased/+csharp-header-probe-pruning.changed.md
• changelog.d/unreleased/+csharp-lookahead-copy-reuse.changed.md
• changelog.d/unreleased/+csharp-callable-receiver-cache.changed.md
• changelog.d/unreleased/+structural-mask-code-spans.changed.md

Follow-up candidates

Collect matched Release timings for the final batch and extend measurements to large corpora with a broader language mix.