[Catalog] Fix description-vs-capability integrity + add a claims lint #66

[Catalog] Fix description-vs-capability integrity + add a claims lint

Several published agents advertise capabilities their capabilities block does not have — the catalog makes promises the code can't keep. For a store this is an integrity issue, not polish.

Confirmed mismatches (live 2026-08-01)

• Creator OS — desc: "publishing through your own logged-in browser sessions via the local runner" — but runtime:null, workflow:null (no browser runtime).
• QA Automation — desc: "runs them headlessly after every deploy … observable runner" — but runtime:null, workflow:null.
• Job Application Assistant — desc references "approved local browser runner task job.apply_basic" — verify that task type actually exists in the runner (runner declares job.apply_agent, not job.apply_basic).

Do

1. For each: either wire the capability or rewrite the description to match reality. (Given the focus posture, rewriting is likely right for the aspirational ones.)
2. Add a publish-time / CI lint that flags catalog copy claiming a runtime capability (browser/posting/headless/local runner/cron) when capabilities.runtime/workflow don't back it. Keep it heuristic + overridable, but loud.

Acceptance

• No published agent's description claims a runtime capability it lacks.
• Lint flags a deliberately-mismatched fixture in tests.

Part of the catalog audit.