Distributed render: probe stage crashes on transient browser errors without retry

[miga-heygen]

Problem

The distributed render plan() stage crashes when headless Chrome encounters a transient frame detachment error during the browser probe, with no retry logic. The plan tarball is never written, and downstream chunk workers fail.

Error: errorCode=INTERNAL_ERROR message=Navigating frame was detached

This is a Puppeteer/Chrome DevTools Protocol error that occurs when Chrome's main frame navigates away or is detached during an operation — typically due to memory pressure, compositor race conditions, or Chrome warmup timing.

Impact

• All chunk workers fail (plan artifact missing)
• The entire distributed render fails after retry pass
• The error is transient — a fresh browser session would succeed

Root Cause

probeStage.ts calls initializeSession() (which calls page.goto() in frameCapture.ts) with zero retry logic for transient browser errors. The error propagates straight through: initializeSession → probeStage → plan() → adapter → failure.

Known transient Puppeteer errors that should be retried:

• Navigating frame was detached
• Target closed
• Protocol error ... Target closed
• Session closed
• Navigation failed because browser has disconnected

Fix Plan

1. Retry logic in probeStage.ts: Wrap the browser session lifecycle (create → initialize → probe) with retry-on-transient-error. On failure: close the crashed session, create a fresh browser, retry once.
2. Typed error classification: Add isTransientBrowserError() utility in the engine for classifying retryable vs. fatal browser errors.
3. Observability: Structured logging at the retry boundary — error class, attempt number, elapsed time.
4. Tests: Unit tests for the error classifier + integration test for probe retry.