fix(producer): retry probe navigation timeouts

[miguel-heygen]

Summary

Treat Chromium navigation timeouts during browser probe/init as transient browser failures so runProbeStage retries once with a fresh browser session.

This preserves the preferred Linux BeginFrame path. It does not force screenshot mode.

Why

Render-streaming probe/init failures can surface as browser lifecycle errors such as:

• Navigating frame was detached
• Protocol error (Runtime.evaluate): Target closed
• Navigation timeout of 60000 ms exceeded

Main already retries the first two classes through isTransientBrowserError, but navigation timeout was explicitly excluded. This PR closes that gap so probe-stage navigation timeouts use the same fresh-browser retry path before the producer emits a terminal render error.

Validation

• Adds classifier coverage for Navigation timeout of 60000 ms exceeded.
• Adds producer probe-stage coverage showing a transient navigation timeout retries with a fresh browser session and succeeds.
• Operational prod-path validation was performed separately and is tracked in the internal incident notes; workflow/run identifiers are intentionally omitted from this public PR description.

Verification

• bun test packages/engine/src/services/frameCapture-transientErrors.test.ts packages/producer/src/services/render/stages/probeStage.test.ts — 30 pass, 0 fail
• bunx oxfmt --check packages/engine/src/services/frameCapture.ts packages/engine/src/services/frameCapture-transientErrors.test.ts packages/producer/src/services/render/stages/probeStage.test.ts
• bunx oxlint packages/engine/src/services/frameCapture.ts packages/engine/src/services/frameCapture-transientErrors.test.ts packages/producer/src/services/render/stages/probeStage.test.ts
• git diff --check HEAD~1..HEAD
• pre-commit hooks passed: lint, format, fallow audit, typecheck, commitlint

Note: I also tried the broader render-stage test bundle including captureStreamingStage.test.ts; that test file has unrelated mock drift in isolation (getCapturePerfSummary export/session options mocks), so I did not treat it as verification for this PR.

[miguel-heygen]

Closing this for now because it is retry classification / mitigation, not the root-cause fix for the Chromium renderer crash. We will reopen or replace with a root fix once the crash trigger is proven.

[miguel-heygen]

Reopening based on the deeper prod-path repro. We now have exact-image A/B evidence: the deployed prod sidecar image (0.7.2, Chrome 150) fails the exact failed streamed-preview source under /v1/render-stream with the observed browser_probe signatures (6/8 failures: Navigating frame was detached / Target closed), while the 0.7.6 candidate image passes the identical replay 8/8. The candidate classifier still returns false for Navigation timeout of 60000 ms exceeded, so this follow-up remains required to cover all observed production signatures. This should be treated as part of the root fix, not just a Temporal retry bandaid: it lets the sidecar recover inside probe by relaunching a fresh browser session before emitting SSE error.