Claude usage limits: a rate_limit_event after a failed probe is published as a complete window list

Reading apps/server/src/provider/providerUsageLimits.ts at 052c7ae: when the boot probe fails, the published limits are { windows: [], unavailable: { reason: "probeFailed" } }. A rate_limit_event during a turn then goes through applyUsageLimitsUpdate, which merges its one window onto the empty list and rebuilds the snapshot without unavailable. The result looks like a full read with only seven_day, and resolveUsageLimitsAfterProbe keeps it over later failed probes because the marker is gone.

Seen on 0.0.43-nightly.20260916.1825 after the update restarted the app: the Claude cache held only the weekly window for about 6 minutes, until a refresh-all re-ran the probe. A second Claude instance with no active turn had no windows the whole time.

Possible fix: keep unavailable, or mark the snapshot partial, when merging onto a snapshot that had it. Related: hydrateCachedProvider in providerStatusCache.ts does not carry usageLimits across a restart, so every restart starts from nothing.

Start in apps/server/src/provider/providerUsageLimits.ts at applyUsageLimitsUpdate and resolveUsageLimitsAfterProbe, reproducing the failed-probe state before a rate_limit_event arrives. Review hydrateCachedProvider in providerStatusCache.ts for the restart behavior. Done means a partial update cannot make failed probe data look complete, and the unavailable marker or equivalent partial state remains distinguishable.