Fix dropped async delegation completion after context compression

[heyparth1]

Summary

This PR fixes an issue where async delegation completions dispatched before context compression were dropped by the #55578 fail-closed guard if they completed after the parent session rotated to its compression-continuation child.

Changes

• Extracted the pinned session resolution logic into a new _resolve_pinned_delegation_session method.
• Added a carve-out to the fail-closed guard: if the pinned session was ended due to compression, the completion is rerouted to the live compression continuation tip (get_compression_tip) instead of being dropped.
• Maintained the existing fail-closed behavior for explicitly ended (/new reset) or unknown sessions to prevent illicit conversation revival.

Closes #65779

[tonydwb]

Code Review Summary

Verdict: Approved

Fixes dropped async delegation completion after context compression. Clean, well-scoped fix. No security concerns.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the focused fix. The premise is confirmed on current main: gateway/run.py:11241-11266 drops a completion when its pinned parent has ended_at, while hermes_state.py:3568-3629 already provides a lineage-safe compression continuation resolver that excludes branch, delegation, and tool children.

Problems

• tests/gateway/test_async_delegation_session_binding.py:94-100 uses inspect.getsource() and a symbol-string assertion to prove wiring. This does not execute the routing path, and AGENTS.md prohibits source-reading tests.

Suggested changes

• Replace that test with an executable handler-path regression: a compression-ended parent should route through switch_session to a live tip, while session_reset must still make no switch.

Automated hermes-sweeper review.

[teknium1]

Please replace this source-text assertion with an executable routing test. It can pass even if the handler no longer reaches the resolver at runtime, and AGENTS.md explicitly prohibits source-reading tests.

[teknium1]

Fixed via #69312 (merged) — salvage of #64530 (first submitter) with the ack-semantics layer from #65838 folded in. Your PR independently identified and fixed the same inner-guard drop; credited in the merged PR body. Thanks for the fix!