[sugarforever]

Active Android agent runs were cancelled after 30 minutes, including concurrent conversations sharing the same service generation. Remove that total-duration timer so admitted work continues until completion, explicit cancellation, or a system/runtime failure.

Keep finite wake-lock acquisitions: renew a ten-minute partial wake lock every five minutes while the service is active, cancel renewal and release on shutdown, and persist wake_lock_failed if renewal fails. Preserve the startup handshake, notification permission requirement, Stop action, cancellation deadline, recurring-task behavior and no-auto-restart policy. Historical budget-expiry diagnostics remain readable. Update notification text, manifest explanation and documentation.

Validation:

• npm run verify passed: 268 tests passed, 37 skipped; 3 bootstrap tests; types; portable smoke; Android/iOS Hermes exports.
• Six tests compile and execute the production Java lifecycle helper with a controlled clock, including two-hour execution, renewal, Stop, stale callbacks, restart and acquisition failures.
• Independent review: no actionable findings. Prettier and diff checks passed.
• Expo Doctor: 19/20; existing locked Expo patch versions are behind the currently recommended patches. Dependencies unchanged.
• Android APK compilation is running in Actions: https://github.com/sugarforever/moose/actions/runs/40151454027 . Local SDK has platform-tools only.
• Physical-device execution beyond 30 minutes and paid model requests have not been performed. The user will validate on an Android phone.

This is the implementation follow-up to the architecture/validation documentation in #27. No schema, credential, provider or runtime scheduling changes.