fix(delegation): update live-transcript manifest on async completion

[plcunha]

Summary

Keep live-transcript manifests synchronized when background delegate_task workers finish. Both single-child and batch finalization now write terminal per-task statuses through update_manifest_statuses(), so Desktop/status consumers no longer see completed children as perpetually running.

The manifest update is best-effort and intentionally occurs after completion delivery; a filesystem/logging failure cannot suppress the result returning to the parent conversation.

RED proof

The regression tests were copied onto unfixed origin/main (cff972858) and run there:

$ HERMES_TEST_FILE_RETRIES=0 scripts/run_tests.sh \
    tests/tools/test_async_delegation.py -q \
    -k 'single_async_finalization_updates_live_manifest or batch_async_finalization_updates_live_manifest'

FAILED ...::test_single_async_finalization_updates_live_manifest
  AssertionError: assert [] == [('deleg_...', [result])]
FAILED ...::test_batch_async_finalization_updates_live_manifest
  AssertionError: assert [] == [('deleg_...', results)]
2 failed, 40 deselected

After the fix:

2 passed, 40 deselected

Code path trace

1. Symptom: background children finish and deliver their summaries, but cache/delegation/live/<id>/manifest.json can remain status=running indefinitely.
2. Intermediate: the synchronous aggregation path calls update_manifest_statuses(), while _finalize() and _finalize_batch() only publish completion events and durable state.
3. Root cause: the async finalization rail never bridged its terminal result back into the live-transcript manifest owned by delegation_live_log.

Widening audit

• Single background delegation: affected → updates the manifest with [result].
• Background fan-out batch: affected → updates the manifest with combined.results.
• Synchronous delegation: already updates manifests; unchanged.
• Empty batch results: no update, preserving the manifest rather than fabricating terminal entries.
• Completion delivery: remains authoritative and occurs before the best-effort manifest write.
• Manifest filesystem/import failure: logged at debug and cannot lose the user-facing completion event.
• Stalled/interrupted async finalization: shares the same finalizers and therefore gets terminal manifest propagation.

Verification

$ HERMES_TEST_FILE_RETRIES=0 scripts/run_tests.sh \
    tests/tools/test_async_delegation.py \
    tests/tools/test_delegation_live_log.py -q
75 passed

$ ruff check tools/async_delegation.py tests/tools/test_async_delegation.py
All checks passed!

$ python3 -m py_compile tools/async_delegation.py tests/tools/test_async_delegation.py
$ git diff --check

Validated on PR head d3f89eb0c. GitHub reports the branch mergeable with all required checks passing against the current base.

[alt-glitch]

This was generated by AI during triage.

Related to #67479: that merged live-transcript infrastructure updates manifest statuses on synchronous aggregation, while this focused follow-up adds the missing async-finalization updates.

[teknium1]

Thanks for the focused investigation and regression coverage.

Automated hermes-sweeper review found this behavior is already implemented on current main:

• tools/delegate_tool.py:3146 calls update_manifest_statuses(live_deleg_id, results) after finalizing child results.
• The production background path uses that same aggregation function through _batch_runner at tools/delegate_tool.py:3270-3273, then dispatches it with the matching live-transcript ID at tools/delegate_tool.py:3318-3336.
• The existing implementation arrived with 299e409f15aa5615a8a64be488580be92cda351e (feat(delegation): live-viewable subagent transcripts — tail your subagents while they work (#67479)).

The proposed tests invoke dispatch_async_delegation* directly, whereas live manifests are created and owned by the delegate_task aggregation path. Closing as already implemented on main.