Relax eval inference preflight timeouts

[d42me]

Summary

• add explicit inference timeout handling so eval preflight failures surface cleanly instead of crashing with a raw httpx timeout
• increase the Prime eval preflight timeout to 5 minutes and make timeouted model/billing probes warning-only so slow thinking models can still start
• add regression coverage for the longer preflight timeout and timeout fallback paths

Testing

• pytest packages/prime/tests/test_eval_billing.py -q
• pytest packages/prime/tests/test_hosted_eval.py -q

---

Note

Medium Risk
Changes eval preflight behavior by passing custom HTTP timeouts into InferenceClient and treating timeout failures as warnings, which could allow eval runs to proceed with an invalid/unwarmed model. Risk is limited to CLI UX/flow and is covered by added tests.

Overview
Relaxes prime eval inference preflight by introducing a dedicated EVAL_PREFLIGHT_TIMEOUT and wiring it through to InferenceClient, rather than using the default client timeout.

Model validation and billing-probe preflights now catch httpx.TimeoutException and continue with a warning instead of crashing/failing the run.

Updates tests to stub InferenceClient construction, assert the preflight timeout is passed through, and adds regression coverage for the timeout-continue paths and custom client timeout support.

^{Reviewed by Cursor Bugbot for commit e4d6139. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e4d6139. Configure here.}