fix(app-server): stop skills watcher on shutdown

[fcoury-oai]

Why

Pressing Ctrl+C in the TUI could make Codex take about 5 seconds to exit.

Shutdown tracing showed that the delay was the in-process app-server waiting for its outbound task to finish until the shutdown timeout elapsed. The remaining outbound sender owner was consistently skills_watcher.event_loop.

The app-server-owned watcher task captured Arc<OutgoingMessageSender> and waited indefinitely on file-watch events, so shutdown could not close the outbound channel promptly.

What Changed

• Add an explicit shutdown signal to SkillsWatcher.
• Make the watcher event loop exit when either a file-watch event arrives or shutdown is requested.
• Call skills_watcher.shutdown() from MessageProcessor::clear_runtime_references() during app-server teardown.

This lets the watcher task drop its OutgoingMessageSender clone during shutdown, so the outbound channel closes without waiting for the 5 second timeout.

How to Test

1. Start Codex TUI.
2. Press Ctrl+C.
3. Confirm Codex exits promptly instead of pausing for about 5 seconds.

Focused regression check from the shutdown investigation:

• Before this fix, 3/3 sampled shutdowns timed out in app_server.runtime.wait_outbound_handle, app_server_runtime.wait_shutdown_ack, and app_server_client.wait_shutdown_response.
• With the same fix applied on the investigation branch, 2/2 sampled shutdowns had no shutdown timeouts and app_server.runtime.wait_outbound_handle completed in 0ms.

Local validation:

• just fmt
• git diff --check

Local lint note:

• just argument-comment-lint did not reach this diff because local Bazel analysis failed in external llvm / BuildBuddy setup before the linter ran.

[fcoury-oai]

Superseded by #23578.