fix(mcp): make local-endpoint budgets configurable - #1856

Summary

• tool_answer/config.py: _SYNTHESIS_MAX_TOKENS (1024, hardcoded) now reads REPOWISE_SYNTHESIS_MAX_TOKENS. A reasoning model spends its whole budget on hidden thinking before emitting an answer token, so the call comes back empty at the old cap; measured against nanbeige, 1024 tokens returned nothing, 8192 produced a correct, cited answer.
• _answer_pipeline.py / _helpers.py: _EMBED_TIMEOUT_S (8.0, hardcoded) now reads REPOWISE_EMBED_TIMEOUT_S via a new embed_timeout_s(), mirroring the sibling vector_search_timeout_s(). A cold-loading local embedding model blows the old 8s budget; get_answer still exits 0 with the semantic leg silently dropped, visible only at WARNING.
• One PR because it's one root cause: constants tuned for a warm remote service, with no override, that degrade silently on a local one. Both follow the existing pattern for REPOWISE_EMBEDDING_TIMEOUT / REPOWISE_VECTOR_SEARCH_TIMEOUT_S: an unparseable or non-positive override warns and keeps the default.

Related Issues

None.

Test Plan

• uv run pytest tests/unit/ -q — 14242 passed, 12 skipped, 2 xfailed. 2 failures, both pre-existing on unmodified main and unrelated to this change: a git-rebase test that depends on local git config, and a timezone-sensitive episode-dating test.
• uv run ruff check .
• uv run repowise risk main..HEAD:

  Change risk for main..HEAD: touches files that have broken before · 90th percentile of this repo's fix-bearing files
    fix(mcp): make local-endpoint budgets configurable
    +269 / -12 lines · 7 files · 3 dirs · 2 subsystems · entropy 2.19 · author exp 0  (fix)
  
  File                                                                       Lines   Prior fixes
  packages/server/src/repowise/server/mcp_server/tool_answer/config.py         36           11.7
  packages/server/src/repowise/server/mcp_server/_helpers.py                   36            6.4
  tests/unit/server/mcp/test_answer_synthesis_timeout.py                       47            5.8
  packages/server/src/repowise/server/mcp_server/_answer_pipeline.py           19            4.9
  packages/server/src/repowise/server/mcp_server/tool_answer/synthesis.py       6            2.8
  
  Diff shape: Typical · 57th percentile of recent commits by size and spread
    About as risky as a typical commit in this repo.
  

  This flags real history: these are exactly the answer-pipeline files this fix targets, and they carry a documented run of prior fixes (config.py alone shows a recency-weighted 11.7). That matches the PR's own premise — this is a bug-prone area — rather than indicating a new problem in this diff; the full-suite result above is the direct check on that.
  (impacted-tests / health --file need a completed repowise init against this clone; that indexing run did not finish in time and is omitted rather than faked.)

Checklist

• My code follows the project's code style
• I have added tests for new functionality
• All existing tests still pass
• I have updated documentation if needed

Written with AI assistance; measurements and tests were run and verified locally.