* fix(usage): correct last-known quota retention and presentation

Unknown service status clears stale outage prose instead of keeping
the previous incident visible. Extra-lane quotas persist across
restarts. Expired quotas hold no measurement with explicit not-stale
state. Restore banner shows only for actually stale data. Per-provider
resets preserve healthy providers' disk cache. Empty usage results
keep CLI error detail. Stale supplement sections hide until
revalidated. Measurements older than 24 hours never stamp fresh.
Overview rows keep account/status context beside the last-known note.
Long update timestamps ellipsize on one line.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* fix(usage): preserve extra windows and stale cost state

Address the Codex and CodeRabbit cache review findings. Restore bounded extra-window arrays through the QML contract, keep token costs hidden for stale usage across lifecycle updates, and use receipt time for future live timestamps. Update usage and development guidance; TODO scope and CLI contracts are unchanged.

* fix(usage): exclude quota-less successes from retention

Preserve successful empty, credits-only, and unavailable-quota responses when their supplemental timestamp exceeds the quota retention window. Add cache and real-app expiry regressions. Update usage, development, and changelog documentation; configuration, translations, and TODO scope remain unchanged.

---------

Co-authored-by: Claude Code <noreply@anthropic.com>