[Borisserz]

Summary

Fixes the "providers list goes empty after prolonged uptime" report. The fix makes polling recover by itself instead of needing a restart.

Root cause. UsageStore stops the poll timer on NSWorkspace.screensDidSleepNotification (to save battery) and restarts it only on screensDidWake. Nothing else could clear pollingSuspended: didWake, network reconnects, and the midnight NSCalendarDayChanged handler each ran one refresh(), but none of them rescheduled the timer. So if a screensDidWake was ever missed, polling stayed off for the rest of the process lifetime.

This also explains the exact symptom. The overnight NSCalendarDayChanged refresh runs at 00:00 with displays asleep. No tokens exist yet for the new day, so it correctly writes providers: [] / menuTitle: "0" with an empty lastError. With the timer gone, nothing replaces that snapshot. It stays frozen while Claude Code keeps producing usage, and a restart fixes it because init reschedules the timer. The popover doesn't refresh on open, so it shows the same frozen state.

Fix.

• While suspended, a low-frequency probe (every 5 minutes, 20% tolerance) checks the real display state with CGDisplayIsAsleep(CGMainDisplayID()). If the screens are awake, it resumes polling and refreshes once.
• didWake and every refresh() entry run the same check. Any trigger (system wake, network reconnect, manual refresh) heals a stale suspension right away instead of waiting for the probe.
• When displays really are asleep, nothing changes: polling stays suspended, and only the probe timer runs, which costs one cheap CG call every 5 minutes.
• Each recovery writes polling resumed without screensDidWake (<reason>) to ~/Library/Logs/PokeTokenBar.log, so a missed wake shows up in field logs.

About usage-cache.json. The reporter saw it as an unreadable binary blob. That's expected: the cache is zlib-compressed JSON (LocalUsageCache.saveIfNeeded), so it isn't related to this bug.

Fixes #350

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Documentation
• Other:

Checklist

• swift build and swift test pass locally (scripts/test-gate.sh: 1389 tests, 0 failures, logic-core coverage 93.35%)
• PR title and description are written in English
• UI changes are described above (before/after — images optional) — no UI changes
• No copyrighted assets, secrets, or private tooling references are committed (see CONTRIBUTING)
• Tests were added or updated for this change

Test plan

New UsageStoreTests cases. Display state is injected through the new displaysAsleep init parameter, which defaults to CGDisplayIsAsleep.

• testSuspendedPollingProbeResumesWhenDisplaysAreAwake: the trigger branch (suspended, screens awake, no screensDidWake). The probe resumes the timer, removes itself, and the refresh fills the snapshot.
• testSuspendedPollingProbeStaysSuspendedWhileDisplaysSleep: displays asleep, so polling stays suspended.
• testSystemWakeResumesSuspendedPollingWhenDisplaysAreAwake: the didWake path alone.
• testRefreshWhileDisplaysAwakeClearsAStaleSuspension: the refresh() path alone, without the probe or wake.
• testRefreshWhileDisplaysAsleepKeepsPollingSuspended: the midnight refresh with screens off doesn't resume polling.
• Fault injection: making the stale-suspension check always return false makes the 3 recovery tests fail (9 assertions). The 2 negative tests still pass, as expected.
• Recorded the defect class in docs/reference/defect-log.md (energy section).