Resume terminal failed workflows with pending ready nodes

Summary

A durable Modal recovery can get stuck when the workflow backend reports a terminal failed run even though the Ultrafuzz durable state has no failed nodes and still has pending ready work to dispatch.

Evidence

R43 Aave v4 invariant-only run aave-v4-v0012-main-issue270-r43-invariant-only on main 4af20bf9b9e5833cc2e35e74a1080a46697779c6 reached all eight property-lens outputs (21 succeeded / 11 pending) and then the original sandbox exited resume-required.

The detached overseer automatically launched recovery attempts without manual run control:

• generation 1 attempt 2 sandbox sb-hSxH59GnxxCelK60lzHU5G, stopped at 2026-08-06T09:07:50Z, no progress, worker diagnostic terminal-run-non-resumable
• generation 2 attempt 3 sandbox sb-AD9BEDVQg96eHfFy14sq7b, stopped at 2026-08-06T09:23:27Z, no progress, same diagnostic

The durable state after attempt 3 is terminal failed with counts 21 succeeded / 11 pending. property-specification-fanin is pending, wait_reason: ready, next_eligible_action: dispatch, while there are no failed nodes. Events show both resume submissions were accepted, followed by immediate syncs with workflow_status: failed / workflow_state: failed and no fan-in dispatch.

Preserved evidence:

• /home/ubuntu/ultrafuzz-private/aave-v4-v0012-n1-20260802/r43/r43-private-inspection-20260806T0909Z.json
• /home/ubuntu/ultrafuzz-private/aave-v4-v0012-n1-20260802/r43/r43-private-inspection-20260806T0919Z.json
• /home/ubuntu/ultrafuzz-private/aave-v4-v0012-n1-20260802/r43/r43-detached-overseer.log
• /home/ubuntu/ultrafuzz-private/aave-v4-v0012-n1-20260802/r43/r43-recovery-state.json

Impact

The invariant-only Aave pipeline cannot naturally advance from property JSON/Markdown to fan-in, invariant setup, Recon, findings, and report, even though the next node is ready. Repeated Modal recoveries preserve the evidence but do not make progress.

Expected fix

Add a regression for terminal failed workflow state with no failed workflow tasks and pending ready durable nodes, then make resume recovery reset/fork/replay or otherwise rehydrate the backend so ready pending nodes dispatch instead of immediately re-finalizing as failed.

Start with the preserved inspection JSON files and detached-overseer log to confirm the terminal failed state, pending ready node, and accepted resume submissions. Trace the resume recovery, workflow-status sync, and ready-node dispatch entry points, then add the regression described in the issue. Done means a terminal failed workflow with no failed tasks and pending ready nodes resumes and dispatches the next node instead of immediately finalizing as failed.