test(e2e): fail-fast runaway detector for OC long-running test

The OC long-running test hangs on CI with 1700+ identical mock requests
in a tight loop on a single sendPrompt call. Previously this consumed
the full 600s sendPrompt budget per turn, dragging the entire run to
the 40-min job timeout and providing no diagnostic data when the run
was eventually cancelled.

Changes:
- Runaway detector: when a single turn generates >100 unexpected mock
  requests AFTER the OC-PARTS diagnostic has fired, log the runaway
  flag. Currently observational — once the cause is understood we
  can also abort the sendPrompt deterministically.
- Per-prompt budget: 600s → 180s. The OC-PARTS diagnostic fires at
  +25 mockReqs (~17s into the loop at 3 req/sec), so 180s is plenty
  of headroom for the diagnostic to capture state before timeout.
- Net effect: failing OC long-running runs surface in ~3min instead
  of ~10min, freeing CI cycles for faster iteration.