ci: exercise deferred agent dispatch and surface background failures

[Zongwei9888]

Two automation tests still assumed the Agent had consumed input when its durable Run became RUNNING. A 100 ms dispatch delay reproduced both failures: the overlap test counted Agent starts too early, and the interruption/continuation test reused an unconsumed scripted step.

Wait for the scripted Agent before interrupting it, and check the overlap execution count after completion. Reuse the same startup synchronization in the legacy-client race test. Run all 32 Goal-run scenarios with both immediate and deferred Agent dispatch in the normal suite, so every supported Python version exercises both schedules.

Improve failure visibility without retries or relaxed deadlines: name each test in CI logs, emit Python thread stacks for tests taking over 60 seconds, and fail on unhandled background-thread/unraisable exceptions. Update the CI guide with the synchronization contract. No production runtime code changes.

Validation:

• Before the fixes: delayed dispatch reproduced 2 failures; 30 scenarios passed.
• After the fixes: 64 immediate/deferred scenarios passed, plus 5 consecutive complete runs (320 scenarios).
• Full local suite with the stricter configuration: 1,872 passed, 10 skipped.
• Separate probes verified background-thread and unraisable exceptions fail pytest; slow-test stack output preserves a successful verdict when the test itself succeeds.
• Pre-commit, actionlint, and diff checks passed.