Modal worker does not retry an exhausted generated preparation task

Summary

On release/v0.1.0, a Modal worker cannot recover a retained inner run when the selected cloud task's exact generated preparation task exhausted its retries.

The validated selected-task DTO explicitly binds all three task IDs:

id:            node:<attempt>
preparationId: prepare:<attempt>
verifierId:    verify:<attempt>

After normal resume re-emits the retained failure, selectedInnerTaskRetriesExhausted() accepts only a blocker whose node ID equals id. A blocker for the DTO-bound preparationId is treated as unrelated, so the worker rethrows the original resume failure and never invokes the already-bounded retry path.

This occurs before agent inference.

Minimal generic reproduction

Use only the existing temporary Git fixture:

1. Select node:attempt-one with validated preparationId: prepare:attempt-one.
2. Make normal inner resume fail.
3. Make the diagnostic return one blocker: { kind: "retries-exhausted", nodeId: "prepare:attempt-one" }.
4. Make retry-task succeed only when --node-id prepare:attempt-one is used.
5. Assert the worker resumes by retrying that exact DTO-bound preparation task.

Run:

pnpm --filter @ultrafuzz/modal exec vitest run test/node-provider.test.ts \
  -t 'retries the exact generated preparation task for the selected inner task' \
  --reporter=verbose

Observed:

1 failed | 204 skipped
cloud worker phase resume-workflow failed with code 1

The command log contains up and why, but no retry-task.

Expected behavior

The recovery selector should return the exact exhausted blocker ID only when it equals one of the validated selected-task DTO's id, preparationId, or verifierId. The retry command should target that exact returned ID. Every unrelated blocker must remain terminal.

Duplicate search

No issue matched the preparation-task retry-selection gap. Issues #745 and #746 cover earlier dependency relocation and handoff-identity boundaries, not selection of a DTO-bound generated prerequisite after resume.

Start with test/node-provider.test.ts and run the provided Vitest command using the existing temporary Git fixture. Trace the recovery selector after normal inner resume fails, focusing on the validated selected-task DTO and its id, preparationId, and verifierId. Done means the test passes, the exact exhausted DTO-bound ID is retried, and unrelated blockers remain terminal.