fix(skills): align search guidance with hybrid retrieval and teach answer-card phrasing

[vraspar]

Retrieval went hybrid in prod (lexical + dense, RRF-fused, calibration hybrid-v1), but the vendored skills still taught the old model. tenjin-search told agents "search is lexical, not semantic: it matches words, not meaning", which pushes exactly the wrong behavior: compressing a question into keywords throws away the signal the dense leg runs on. This is the CLI-side half of guidance parity with the server docs in BackTrackCo/tenjin#464.

What changed:

• skills/tenjin-search/SKILL.md: lookup matches wording and meaning; send the complete question as one natural-language sentence, do not compress to keywords. Privacy hygiene (strip private identifiers, internal service names, secrets, do not search if it cannot be generalized) and the MISS-is-fine rule are untouched.
• Demand-to-supply loop: the candidate-parking example now passes --question "<the question you looked up>". Verified in src/commands/candidate.ts (stores it in the candidate meta, capped at the server's 200-char bound) and src/commands/publish.ts (prefills it into the card's questionsAnswered as a fallback, still losing to an explicit --question or frontmatter). So the exact phrasing a searcher used becomes a match target on the published card.
• skills/tenjin-publish/SKILL.md: the answer-card draft rule was one vague sentence. It now carries the same phrasing guidance as the server docs: 5 to 10 questionsAnswered entries at 200 characters max, varied in register (natural symptom sentence, terse keyword line or verbatim error string, why/how question), same-register rephrasings add nothing, tasksSupported is for tasks and not questions, and scope is embedded too so write it dense and factual.
• README.md: the lookup paragraph said "derive the smallest public phrasing of your task", which is the same compress-to-keywords advice. Now: strip the private parts, then send what is left as one complete sentence.
• Patch changeset, since skills ship inside the package.

On the vendored zero-install skill (skills/tenjin/SKILL.md): it is byte-identical to live tenjin.blog/skills.md, so pnpm sync:skill produces zero diff, and it already carries the Resource card section, the lookup and submit_feedback MCP tools, and preview in the discovery search description. Commit b3e91d1 (#25) resynced it on 2026-07-24. One stale claim rides along: line 103 still describes lookup matching as "honest lexical, not a semantic score", so this package ships one skill saying lookup matches wording and meaning and another saying it does not. That line cannot be corrected in this repo without breaking skill-drift CI, since the file is a mirror. It gets fixed upstream under BackTrackCo/tenjin#464, and this repo picks it up on the next pnpm sync:skill after that ships. src/fixtures/openapi.fixture.json has the same shape of residual: it is a vendored server contract snapshot consumed by src/contract.test.ts and still pins calibration to a const lexical-v1, and it refreshes on the next contract sync, not here.

On why the earlier drift sat unnoticed: skill-drift.yml is configured correctly (daily schedule, workflow_dispatch, and a PR trigger on skills/** + scripts/sync-skill.mjs) and it did catch the drift. Scheduled runs on main failed on 2026-07-20 and 2026-07-21. The gap is alerting, not detection: a scheduled failure on main gates no PR, opens no issue, and only emails the workflow file's last committer, so main stayed red for days. Adding a failure-path step that opens a tracking issue needs issues: write on the job, so it is left for a follow-up rather than escalating workflow permissions in a docs PR.

No lookup to search renaming here, to stay out of BackTrackCo/tenjin#463's way.

Gates: typecheck, lint, prettier check, and the full vitest suite (709 passed, 9 skipped) all green.

🤖 Generated with Claude Code