probe: u20 search-quality evaluation (do not merge)

Draft
AF
2d ago
+1058-268

Draft probe. Runs the search-quality gate on audit slice u20's reconstructed candidate to learn whether any real-query row moves. Not for landing; closed after the gate reports.


Summary by cubic

Probe PR that runs the search-quality gate on audit slice u20's reconstructed semantic-search and configure fixes to learn whether any real-query benchmark row moves. Not for landing; it will be closed after the gate reports.

Bug Fixes

  • Bounded snippet reads now preserve valid text and replace invalid UTF-8 instead of dropping whole reads.
  • pyproject.toml tool detection now parses TOML rather than scanning raw markers.
  • Exact memo eviction no longer resets unrelated active verifiers, and stale readers cannot erase replacement payloads.
  • Test-file classification uses root-relative paths and tolerates nested checkouts and shared test support directories.