docs(agents): correct context and prefer proven integration coverage
Summary
Correct misleading agent routing, glossary and validation guidance; prefer provider integration for command behavior and E2E for native behavior while preserving independent unit/type guards and #4050's full-unit fallback. Remove one duplicate alert-wait timeout mock after retained integration detection was proven.
Six files changed; no production behavior changes or scope expansion. Closes #4370; part of #4368.
Validation
Base: 5a836c05d5abfe3daa7556af4a1906a1e1029e17; head: 37119e3cb92f782c6ce3b9c617251d4cf0d79a8c.
The initial diff received Claude Opus sentence review. Required Claude and Cubic findings were fixed; context budgets pass. The coordinator checked the corrections.
The same focused command in an isolated baseline checkout with the candidate removed passed 10 tests; dropping the caller's alert-wait budget caused the retained provider integration scenario to fail (1 failed, 9 passed); restoration passed 10. This proves detection, with no suite-speedup claim. Boundary/error/lifetime tests remain.
pnpm check:affected --run --head HEAD passed all seven selected checks with exactCommit: true. Review dispositions, deletion proof and receipt excerpt. Full coverage and provider CI remain GitHub obligations. No device run applies: only guidance and a redundant mocked test changed.