fix(workflows): recover exhausted autonomous work after crashes

[duc15052006-dotcom]

Closes the NOTE 21-2 crash window where a worker can die on its final queue attempt after the durable workflow state already moved forward.

What changes:

• add a work-queue exhausted cleanup claim lane that leases attempts >= maxAttempts without incrementing attempts;
• keep normal execution capped: exhausted items are cleanup/reconciliation only, never dispatched again as ordinary side effects;
• add exact ready-state failure by ready timestamp + attempt;
• add exact autonomous running-state failure by durable dispatch stamp + attempt;
• reconcile exhausted ready items that crashed before or after ready->running CAS;
• reconcile exhausted wait items that crashed before or after waiting->running CAS;
• finish stale cleanup items harmlessly and release transient cleanup failures for another cleanup pass;
• prove an old wait stamp cannot fail a newer resumed state even when the workflow attempt number is unchanged;
• add queue/store/bridge regression coverage, docs and release-preflight invariants.

No second scheduler, credential path, browser bypass or host filesystem path is added.

Base: exact verify/release-candidate HEAD 985229b.
Do not merge aggregate PR #30.