The readiness fix had unit coverage for _chat_provider_ready and the fallback
upgrade helpers, but nothing asserting the thing that actually went wrong in
production: that a not-ready round leaves the day untouched. "It returned
None" and "it wrote a fallback and pinned the whole day" are both plausible
readings of a green pure-function test.
Three cases through _ensure_daily_plan itself: not ready writes no plan and
counts one wait; the round after the configured limit is the first one allowed
to degrade, and it must actually produce a plan; a ready provider generates
immediately and clears the wait counter, so a later genuine outage gets its
full budget.
Companion suite 246 passed, joint E2E 146/146.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>