Repair downstream turn progress and recovery

[witbrock]

What changed

• keep the current live-progress entry observable after a long quiet interval without resetting its original request clock or phase history
• recursively hydrate nested diagnostic blob references, with cycle protection
• expose pending durable workflow recovery as a non-waiting inspection instead of prescribing another long wait
• preserve an honest model-authored partial-workflow answer when it names every exact durable instance handle

Why

Recent real-path replay work exposed downstream reliability defects after durable workflow submission: active progress could expire, nested diagnostics could remain compacted, recovery guidance could consume another long observation interval, and the useful pending answer could be replaced by a generic finality fallback.

The repair is generic support code. It adds no Gmail-, arXiv-, user-, workflow-ID-, or prompt-specific route.

Impact

Long-running authorised work remains observable and can return an honest pending result with a precise recovery handle. Diagnostic readers receive the full nested payload, while effect finality and canonical read-back requirements remain intact.

Validation

• 197 affected backend tests passed
• fatal Ruff checks passed
• changed Python modules compiled
• git diff --check passed
• the current server is healthy while running this dirty candidate