feat(search): add opt-in TESTS-signal to graph-stage (E17)

[ManSio]

Summary

Adds an opt-in TESTS-signal to symbol search (graph-stage): covering tests
(incoming TESTS edges from the dynamic-trace bootstrap) are appended to found
function definitions as a separate, lower-ranked result group.

Behind MSCODEBASE_TESTS_SIGNAL (env), off by default — production behavior
is unchanged.

What changed

• SearchConfig.tests_signal (env flag, default false).
• SymbolIndexAdapter.get_tests_for_symbol: resolves incoming TESTS edges for a
  function, capped (hub functions like safe_mkdir have 234 tests).
• Searcher._append_tests_signal: appends up to 3 tests per function,
  graph_score = 0.4 (vs 1.0 for definitions), sentinel
  chunk_index = -(20_000_000 + line) to avoid RRF collisions.
• 7 unit tests in tests/test_graph_stage_e4.py.
• Reproducible harnesses shipped: e17_ab_tests_signal.py (narrow A/B),
  e17_wide_panel.py (35 queries + language coverage), e17_redteam.py (5 attacks).

Findings (measured, not predicted)

• hit@1 unchanged (94.3% both arms, 35-query wide panel). The signal adds
  LLM context only — it does not improve retrieval ranking.
• Language coverage is the critical limitation: Python 34.0% of functions
  covered by TESTS edges; Other (Go/Rust) 0%; TypeScript 0%. The dynamic
  trace is Python-only.
• Red team: 5/5 repelled — concurrency (1000 calls, 0 errors), boundaries
  (234-test hub = 16.11ms), abuse (nonexistent symbol → empty), TOCTOU (graph
  closed mid-search → graceful), dependency failure (bad path → empty).

How to reproduce

python -X utf8 experiments/bootstrap/e17_wide_panel.py
python -X utf8 experiments/bootstrap/e17_redteam.py
python -X utf8 experiments/bootstrap/e17_ab_tests_signal.py

Notes

• Article draft (full history Exp 7 → 16 → E17): docs/blog/bootstrap-pipeline.md.
• graph_score = 0.4 is an empirical constant, unverified against reranker/BM25
  weight interactions — see "What Could Go Wrong" in the article.
• Tests: 34 passed locally; ruff clean.
• verified_from_clean_state: this PR is what enables the clean-state check.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 72629b36-52d0-47b6-98cc-b529bfd80aae

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}