Retain verified prior-generation output when prompt drift prevents exact reuse

Generic symptom

Dynamic retry recovery must not mark an old result current when its newly rendered prompt differs. Today the safe rejection also discards all useful work: the replacement model starts from scratch even when node identity, manifest generation, and output contracts are unchanged.

Minimal reproduction

1. Complete and verify one generated attempt in a synthetic dynamic workflow.
2. Archive the generation during an explicit producer retry.
3. Regenerate the same manifest and task/output identity.
4. Make a small change to runtime-rendered context so the exact prompt differs.
5. Resume.

Expected

Keep the old verified publications as immutable, clearly labeled prior-generation evidence. Do not count them as a current success. Admit them to only the same replacement attempt, with bounded paths and digests, so the replacement agent can review and update prior work instead of starting over. The current verifier must still validate newly published current outputs.

Actual

Exact-result rehydration correctly rejects the changed prompt, but the archived verified result is not available to the replacement attempt. The entire agent task runs again without its prior output.

Scope

This is generic runtime dynamic retry/recovery behavior. It needs no provider, network call, or particular target.

Start at the runtime dynamic retry/recovery path and trace exact-result rehydration, archived verified results, and replacement-attempt admission. Confirm how the current verifier handles newly published outputs; done means changed prompts reject exact reuse while bounded prior-generation publications remain immutable, clearly labeled, and available only to the same replacement attempt.