fix: suppress stale process watch notifications

[Redcircle031]

Summary

• Skip all queued notification types for consumed process sessions, not just completion events
• Mark killed process sessions as consumed before moving them to finished
• Add regression coverage for stale watch events and kill suppression

Test Plan

• uv run --extra dev --extra pty pytest -q tests/tools/test_process_registry.py -q

[teknium1]

Thanks for the focused regression coverage. The stale-watch-event premise is still present on current main, but this patch needs updating for current delivery semantics.

Problems

• tools/process_registry.py:890 must preserve main's completion-specific _drain_should_skip() path at tools/process_registry.py:1184; it carries the _poll_observed CLI dedup contract documented at tools/process_registry.py:1138-1149 and covered by tests/tools/test_notify_on_complete.py:381-405.
• The TUI poller bypasses drain_notifications() and only suppresses consumed completion events at tui_gateway/server.py:8745 and :8810. Gateway watch delivery also takes all watch events at gateway/run.py:2569-2570 before injecting them at gateway/run.py:11873-11879. Those paths need the same consumed-session guard.
• The new kill test assigns _is_host_pid_alive at tests/tools/test_process_registry.py:1040, but current main uses _host_pid_is_ours() at tools/process_registry.py:1485, so it does not control the current detached-host branch.

Suggested changes

• Keep _drain_should_skip() for completions, add a separate watch-event consumed guard, update the TUI/gateway consumers, and test each path.
• Rework the kill test around a mocked local session.process and _terminate_host_pid.

This is an automated hermes-sweeper review.

[teknium1]

Current main uses _drain_should_skip() for completion events so _poll_observed still suppresses the CLI's duplicate post-poll completion. Preserve that completion branch and add a separate consumed-session check for watch events instead of replacing it with is_completion_consumed() for every event.

[teknium1]

This does not control current main's detached-host kill path: kill_process() calls _host_pid_is_ours() before _terminate_host_pid. Use a mocked local session.process plus _terminate_host_pid, or mock the current identity-check seam, so the test reliably exercises the killed path.

[cirwel]

Independent confirmation from a current CLI maintenance run: stale completions from superseded watchers arrived after later authoritative results. This queue-consumption fix complements #41903, which removes process completion notices from synthetic user input. Suggested integration regression: consume or kill an older watched session, complete a newer session, drain notifications, and assert the old session emits neither a user-input message nor terminal output while the current completion is displayed once.