fix(gateway): persist async task origins across restart

[zhexuany]

Summary

Detached producers currently reconstruct their return route from mutable session/cache state. That breaks after /new, compression, named-profile routing, or a gateway restart. This PR persists an immutable core origin for background delegation and terminal tasks.

The durable origin contains the source snapshot, message anchor, owning profile, session key, and parent physical session. Completion routing validates profile and compression lineage before delivery.

For terminal(background=true, notify_on_complete=true):

• periodic output is throttled/coalesced direct status delivery (no synthetic user turn);
• only final completion re-enters Hermes;
• local notified processes write output and exit status through a child-owned profile-scoped sink, so a gateway restart can recover new output and the real exit code;
• undelivered finals remain checkpointed until explicit adapter acceptance or user consumption;
• fast-exit registration and prune races are covered.

This PR is adapter-neutral. Feishu live-thread anchor behavior remains in #72024; Feishu conversation/session identity remains in #75104.

Verification

• Core async/terminal focused set: 110 passed
• Additional CLI/delegation compatibility: 28 passed
• Ruff, Python compile, and diff checks: passed

Known boundaries

• Non-host sandbox backends keep their existing reconnect/lost boundary after gateway restart.
• PTY + durable notification redirects stdout to the child-owned log, which may affect rare CLIs that require stdout TTY detection.
• Transport-owning relay adapter provenance is not added in this PR.
• A recovered periodic watcher may repeat one old output tail after restart; final delivery remains deduplicated/durable.

[teknium1]

Thanks for tackling a real race: current tools/terminal_tool.py:2552-2561 starts the process before it stamps watcher/notification state at tools/terminal_tool.py:2689-2772, and current routing still falls back to mutable session/cache state in gateway/run.py:20748-20763.

Problems

• tools/process_registry.py:1532 acknowledges a terminal completion while merely draining/formating it. mark_notification_delivered() deletes its output/exit artifacts and checkpoint state (tools/process_registry.py:1375-1389) before a consumer accepts the synthetic completion. In particular, the TUI can requeue drained events when already busy (tui_gateway/server.py:9655-9660), so the event is retained only in memory after its durable recovery state has been removed.

Suggested changes

• Move terminal acknowledgement out of drain_notifications() and perform it only after the CLI/TUI successfully accepts the synthetic completion. Retain/requeue the event and artifacts on busy or failure paths, and add coverage for that TUI deferral path.

Automated hermes-sweeper review.

[teknium1]

drain_notifications() has only formatted the event here; it has not established consumer acceptance. mark_notification_delivered() deletes the checkpoint and durable output artifacts, but the TUI can still requeue this drained event when its session is busy. Please acknowledge only after the CLI/TUI has successfully accepted the synthetic completion, preserving artifacts on deferral or failure.