test(knowledge): freeze hybrid retrieval contract

[Ldsystem]

Summary

• add a focused keep-summarizing corpus for exact identifiers, paraphrases, mixed-source candidates, and noise
• freeze vector provenance, deduplication, observable reciprocal-rank fusion, fallback diagnostics, stale-index rejection, and governance boundaries
• preserve missing G1/G2/G3 behavior as strict expected-RED coverage without breaking the repository's required CI
• keep all production wiring out of this slice

CI repair

• split the FTS-only fixture from the sqlite-vec-dependent fixture so portable fallback, stale-index, lexical-control, and governance checks do not skip
• replace the production helper-name requirement with a black-box cmd_query RRF test using deterministic FTS/vector results
• replace bare StopIteration failures with candidate-ID diagnostics
• mark only genuinely missing production behavior as strict xfail; an unexpected green result fails until the production slice updates the contract state

Validation

• focused contract: 2 passed, 11 xfailed
• full Ubuntu-equivalent command: 566 passed, 11 xfailed
• macOS-portable command: 565 passed, 1 deselected, 11 xfailed
• git diff --check: passed
• production and workflow files: unchanged

Review status

Cursor accepted the original fixture scope. ChatGPT requested two repairs—remove RRF helper-name coupling and prevent sqlite-vec unavailability from dropping portable coverage. Commit 5675b34 implements both; fresh independent review is pending.

This remains a test-only RED contract slice. G1/G2/G3 production implementation continues separately.

Refs WOR-58.