Fix of Model HTTP retry/backoff + real background-job cancellation

[anon13377331]

Summary

Fixes #88 (Model HTTP retry/backoff + real background-job cancellation).

• Bounded retry with backoff for model HTTP calls. OpenAiResponsesAgentAdapter now funnels all Responses API calls through a single postResponses path (removing three duplicated HTTP blocks) that retries 429/5xx up to 3 attempts with exponential backoff plus jitter, honors Retry-After (both delta-seconds and HTTP-date forms), and fails fast when Retry-After exceeds the 30s in-call budget. Backoff waits poll for cancellation on a monotonic clock, and non-transient statuses (e.g. 400) are not retried. Timeouts and error-message format are unchanged.
• Real background-job cancellation. OpenPhoneAgentJobScheduler.isCancelled() no longer hardcodes false: each run gets a JobCancellationSignal backed by the job store, keyed by a run-ownership token (the markRunning timestamp). background_job_stop (or stuck-run repair) makes the signal fire; a watchdog thread then repeatedly calls adapter.cancel() so even a thread blocked on a stuck upstream read is torn down. All AgentJobStore read-modify-write cycles are serialized on a process-wide lock (callers construct independent instances over one SharedPreferences file, so instance-level synchronized provided no exclusion), and markCompleted/markFailed/markDispatched refuse to record for a run that is no longer live — a stopped job can't be resurrected or notify as finished.
• RuntimeException-safe job path. The job thread wraps the entire run; a broker outage or other unchecked exception records a failed run with the store's failure backoff instead of killing the assistant process. Runs ending in infrastructure errors (network_error, adapter_error) now follow failure backoff and repeated-failure alerting rather than counting as completions. scheduleNext keeps the alarm armed while any job is marked running so a run stranded by process death is repaired without waiting for a boot.
• Docs: docs/AGENT_RUNTIME_V1.md gains a "Cancellation and Upstream Failures" section; changelog updated under Unreleased.

Validation

• ./scripts/check.sh — passes, including the assistant javac gate (86 files against android-36) and the broker smoke test
• git diff --check — clean
• Device test, if relevant: not run — no ADB-connected device or emulator is available in the authoring environment. Requesting a maintainer-triggered run-gcp-lab emulator run for on-device validation (this change affects background-job behavior).
• Screenshot, trajectory, audit export, or release artifact evidence, if relevant: behavioral evidence from a local JVM harness driving the real adapter against a stub HTTP server — 15 assertions covering 429→success retry with jittered backoff, Retry-After preference, fail-fast on Retry-After beyond the budget, bounded attempts on persistent 500, no retry on 400, and cancel-during-backoff stopping within ~1s and producing the "stopped" reply. CI tier-1 equivalents (docs-site build, broker pytest) were also run locally; the one broker pytest failure observed is a pre-existing flaky test bug unrelated to this change (test_bit_flipped_signature_rejected flips the first signature character but conditions on the last).

Risk

• No API keys, private SSH keys, signing keys, generated build artifacts, personal device data, or proprietary vendor files are included.
• Licensing impact reviewed for any new third-party material — none added.
• Privacy/security impact reviewed for assistant, watcher, background-run, audit, model-provider, or device-control changes: no new data leaves the device; the change only adds retry timing, cancellation checks, and failure accounting around existing calls. The background-tool confirmation gate (background.confirmation_required) is unchanged.
• Docs updated for user-visible behavior, build flow, or release process changes.

Contributor Agreement

(For the author to confirm — these are personal attestations.)

• I have read and agree to CONTRIBUTING.md, including the contributor assignment/license/no-compensation terms.
• I have the right to submit this contribution under those terms.
• This contribution does not include material I am not allowed to submit.

---

[vercel]

@anon13377331 is attempting to deploy a commit to the adam-7512's projects Team on Vercel.

A member of the Team first needs to authorize it.

[anon13377331]

please tell me if there is anything else I can help with on my side (currently waiting for approval of the workflow run-gcp-lab)