feat(plugins): extend public subagent lifecycle contract

[akivavh]

Summary

• version the public plugin subagent lifecycle contract to v3
• preserve opaque child identity, parent/tool-use identity, bounded lifecycle timestamps and active-task counts without persisting prompts or raw payloads
• relay bounded subagent.progress milestones through the existing host surface
• add generic durable origin notifications and terminal completion registration using Hermes' existing async-delivery queue
• retain in-process terminal relaunch while explicitly treating process-restart loss as unknown, not recovered
• accept currently registered plugin toolsets while preserving parent permission narrowing

Safety boundaries

• no plugin-specific or vendor-specific core tool is added
• plugins receive no live parent or child agent object
• opaque context is cleared before terminal callback delivery
• callback/rate-limit failures leave delivery retryable and cannot rewrite terminal state
• origin acceptance controls final delivery acknowledgment; restored completions deduplicate
• no process-restart execution takeover is claimed

Validation

• python -m pytest -q tests/agent/test_subagent_lifecycle.py tests/tools/test_async_delegation.py — 66 passed
• Ruff check and format-check on the changed Python/tests — passed
• exact consumer composite against contract v3 — durable restore=1, terminal deduplicated, zero active supervisors/async residue, secret scan clean
• merge-tree preflight against current origin/main — clean

The consumer connector and any provider/live acceptance remain outside this public host PR.

[alt-glitch]

This was generated by AI during triage.

Related: #72501 merged the v1 lifecycle API; this PR extends that API to contract v2 with distinct process-local context, progress, and relaunch behavior.

[teknium1]

Thanks for extending the lifecycle boundary while keeping child objects opaque and reusing the existing durable delivery rail.

Problems

• tools/async_delegation.py:679-685 has a race for deterministic notification ids. Two concurrent publishers can both pass the lookup at :732-734; the loser then hits the durable uniqueness error and unconditionally removes _records[delegation_id] at :685, which may be the winner's record. The winner subsequently cannot finalize because _begin_finalization() requires that in-memory record (:969-972), leaving its durable row running without a completion event.
• tests/agent/test_subagent_lifecycle.py:397-415 exercises duplicate publication sequentially, so it does not cover this interleaving.

Suggested changes

• Make registration/reservation ownership-aware and atomic; only remove the exact record created by the failed attempt.
• Add a barrier-based concurrent duplicate-publication regression test asserting one terminal row and one delivered event.

Current main remains at lifecycle contract v1 (agent/subagent_lifecycle.py:25), so the extension is not superseded. GitHub currently marks this branch dirty against main, so salvage also requires reconciling current production-file changes.

Automated hermes-sweeper review.

[teknium1]

This assignment is not ownership-safe against concurrent deterministic notification registration. Two callers can both pass the lookup in publish_external_notification; if the later durable INSERT raises IntegrityError, its cleanup at line 685 unconditionally removes this id and can delete the winning caller's in-memory record. Reserve atomically or only remove the exact record created by this attempt.