Bound MCP tool discovery payload

[Widthdom]

Summary

• Make no-argument tools/list return a deterministic compact catalog for all enabled tools while retaining authoritative invocation schemas.
• Omit documentation-only schema prose, output schemas, examples, and expanded workflow metadata from default discovery; keep full definitions available through format: "full", exact names, and standard cursor pagination.
• Bound initialize guidance, keep it tool-filter aware, add section-level UTF-8/token telemetry without arguments, and preserve safety/authority caveats in compact descriptions.
• Add bilingual developer/testing documentation and changelog.d/unreleased/5059.fixed.md, including regression coverage related to Audit MCP, LSP, JSON-RPC, and rate-limit protocol boundaries #4336.

Root cause and impact

The default MCP catalog serialized long bilingual descriptions, examples, capability metadata, and recursive output schemas for every tool, while initialization repeated workflow guidance. That consumed a large part of an agent context before its first query.

The final real stdio exchange measures 3,016 UTF-8 bytes for initialize and 57,561 UTF-8 bytes for the default tools/list response, below the documented 64-KiB catalog budget. All 25 tool names and complete call-time input constraints remain discoverable; clients can request selected full definitions on demand.

Validation

• dotnet build CodeIndex.sln -c Release --no-restore -p:UseSharedCompilation=false — 0 warnings, 0 errors
• Full Release suite before review fixes:
  • net9.0: 10,762 passed, 413 skipped, 0 failed
  • net8.0: 11,229 passed, 7 skipped, 0 failed
• Post-review/post-main-merge related tests: 11 passed on net8.0 and 11 passed on net9.0
• dotnet run --project tools/CodeIndex.Changelog -p:UseSharedCompilation=false -- check — 8 fragments valid
• Real stdio MCP probe — initialize 3,016 bytes; default tools/list 57,561 bytes; compact safety/authority guidance present
• git diff --check origin/main...HEAD

Adversarial review

Two mandatory Codex adversarial-review rounds were completed. The first found two issues (the suggestion source-code warning and filtered prompt guidance); both were fixed. The second found missing bounded authority caveats for analysis tools; the compact catalog now retains and tests those caveats.

Follow-up candidates

None identified within the requested scope.

Fixes #5059