Stabilize automatic compute generation coverage

[dovvnloading]

What changed

• Replace the timing-sensitive subprocess dependency in the generation-observation test with a deterministic scratch coordinator seam.
• Keep the existing end-to-end local-profile test as coverage for the real isolated scratch worker.
• Assert the generated observation receives the verified arithmetic result and the coordinator lifecycle completes.

Root cause

Automatic compute intentionally falls back to ordinary chat when an isolated worker does not complete within its 1.5-second request budget. The old test required the result on every run while launching a real Windows child process, so occasional startup delays produced a false CI failure despite the expected fallback behavior.

Validation

• python -m ruff check .
• python -m pytest -q tests/test_open_source_execution.py
• python -m pytest -q (374 passed, 1 skipped)
• The repaired regression passed 20 consecutive runs.