Lock down Darwin Labs runner lifecycle and operational regressions

[sjohnston1972]

Recent production runs exposed silent managed-runner dispatch failures, unsafe requeue attempts, and 30-second stale-target waits.

Acceptance criteria:

• A draft containing runs, evidence, analysis, or selection cannot be started in place.
• Missing credentials and GitHub dispatch failures return a visible error while preserving recoverability.
• Retry always uses a new immutable experiment identity.
• Browser actions have a bounded timeout and runner/agent failures remain distinguishable.
• Tests cover start/retry/dispatch guards and action timeout behavior.
• A documented smoke check verifies a managed GitHub runner claims and completes real-target work.

[sjohnston1972]

Implemented in v0.25.18: immutable-run start guard, visible dispatch failure with recoverable awaiting_runner state, retry under a new experiment identity, 5-second browser-action timeout, bounded errors, and direct handler/runner tests. Local real-browser Lab E2E completed in 50 seconds. Remaining acceptance item: authenticated managed GitHub runner smoke against production.