fix: batch Prime background job polling

[DamianB-BitFlipper]

Summary

Prime runtime command execution now delegates to the Prime SDK run_background_job helper instead of manually polling each job with get_background_job. Because runtimes already share one SDK client per event loop, concurrent VM job checks can now coalesce into batches of up to 100 rather than producing one gateway /read-file request per job.

The call preserves the previous one-second polling cadence and uses the existing effectively-unbounded 30-day safety deadline. Rollout cancellation remains the practical execution timeout, and result/error mapping is unchanged.

Verification

• uv run --isolated pytest tests/: 915 passed, 76 credential-gated tests skipped
• focused delegation test: passed
• Ruff, formatting, and pre-push ty checks: passed
• all-files Markdown lint remains blocked by pre-existing inline HTML findings in the untouched ScaleSWE README

---

Note

Medium Risk
Changes how every Prime run() waits on VM jobs; behavior is intended to match prior polling but relies on SDK batching and timeout semantics.

Overview
Prime sandbox command execution now uses the shared SDK run_background_job helper instead of starting a job and polling get_background_job in a loop. That lets concurrent polls on the per–event-loop client batch (up to ~100 jobs), cutting per-job gateway traffic while keeping 1s poll_interval and the same long safety deadline.

The idle-timeout sandbox lifetime fallback constant is renamed to EFFECTIVELY_UNBOUNDED_SECONDS with a clearer docstring; container timeout_minutes still uses that 30-day bound. Rollout cancellation remains the real timeout; ProgramResult mapping is unchanged.

^{Reviewed by Cursor Bugbot for commit b2e7304. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Replace manual polling with SDK run_background_job in PrimeRuntime.run

• Replaces the manual start_background_job + get_background_job while-loop with 1s sleeps in PrimeRuntime.run with a single SDK run_background_job call using timeout=EFFECTIVELY_UNBOUNDED_SECONDS and poll_interval=1.
• Renames IDLE_FALLBACK_LIFETIME to EFFECTIVELY_UNBOUNDED_SECONDS in prime.py and updates all references including the sandbox timeout calculation.
• Risk: polling behavior and timeout enforcement now depend on the SDK's run_background_job implementation rather than the previous manual loop.

^{Macroscope summarized b2e7304.}

[macroscopeapp]

Approvability

Verdict: Approved at b2e7304

Macroscope's review found this PR approvable — This localized Prime runtime fix replaces manual one-second job polling with the existing SDK helper, preserving command execution and result handling while enabling shared-client batching. The only added bound is a 30-day safety deadline, leaving normal rollout timeouts and existing behavior effectively unchanged.

^{You can add or adjust custom eligibility rules. Learn more.}