Reduce rebuild memory amplification

[Widthdom]

Summary

• Reuse same-line C# recovery decisions across declaration-pattern candidates instead of repeating regex work for every pattern.
• Temporarily disable SQLite memory mapping only during high-churn CLI/MCP bulk-load scopes, then restore the configured value on success, failure, and cancellation paths.
• Add allocation and SQLite lifecycle regression tests, plus bilingual developer documentation and a changelog fragment.

Before/after profile

Same-machine Release rebuild command:

dotnet ./src/CodeIndex/bin/Release/net8.0/cdidx.dll index . --rebuild --yes --memory-trace --json --no-progress

Metric	Before (fd3c303)	After (71b4e920)	Change
Elapsed	127,818 ms	128,568 ms	+0.6%
Total allocated	22,456,726,496 bytes	19,239,696,416 bytes	-14.3%
Peak working set	1,218,101,248 bytes	928,333,824 bytes	-23.8%
Peak managed heap	161,139,736 bytes	158,261,816 bytes	-1.8%

The after run indexed 1,261 files, persisted all 47,607 extracted symbols and all 477,700 extracted references, left every graph/index trust flag ready, and reported zero reference-cap hits. Relative to the baseline recorded in #5056, the after run reduced total allocations by 13.7%, peak working set by 31.9%, and elapsed time by 4.3%.

A stable synthetic regression fixture covering 1,000 dense C# generic properties measured 6,495,928 allocated bytes after the change versus 7,015,928 before it (-7.4%), with a 6,800,000-byte ceiling.

Validation

• dotnet restore CodeIndex.sln
• dotnet build CodeIndex.sln -c Release --no-restore -p:UseSharedCompilation=false — net8.0/net9.0, 0 warnings, 0 errors
• Full Release suite before the final upstream merge:
  • net8.0: 11,211 passed, 7 skipped, 0 failed
  • net9.0: 10,744 passed, 413 skipped, 0 failed
• Post-merge focused CLI/MCP indexing suite:
  • net8.0: 218 passed, 1 skipped, 0 failed
  • net9.0: 216 passed, 3 skipped, 0 failed
• Post-merge new regression tests:
  • net8.0: 2 passed, 0 failed
  • net9.0: 1 passed, 1 expected framework skip, 0 failed
• dotnet format CodeIndex.sln --verify-no-changes --no-restore
• dotnet run --project tools/CodeIndex.Changelog -- check
• Fresh local index status: 1,263 files, 47,681 symbols, 479,583 references; all completion/trust flags ready, zero cap hits, indexed HEAD matches the workspace
• Codex adversarial review: no actionable defects found. The reviewer independently passed build, targeted tests, changelog validation, formatting verification, and diff checks. Its additional unfiltered Debug test attempt encountered an unrelated macOS temporary-keychain environment failure.

Documentation and changelog

• Updated both English and Japanese SQLite tuning sections in DEVELOPER_GUIDE.md.
• Added changelog.d/unreleased/5056.changed.md.

Follow-ups

None.

Fixes #5056