[DavidTbilisi]

Summary

An MCP speak left a flatpak mpv running for 8 hours, and only SIGKILL stopped it. The investigation found five problems, all fixed here.

1. Stopping mpv didn't reach a flatpak mpv. The PID Popen holds is bwrap, not mpv, and SIGKILL to it leaves mpv-bin running. mpv is now started with start_new_session=True and stopped with the new mpv_ipc.kill_process_group. That covers MpvIpc.close() and StreamingAudioPlayer.stop().
2. No timeout when mpv never starts playing. MpvIpc.wait_done now raises if no appended entry starts within START_TIMEOUT (10 s), instead of waiting forever.
3. Headless mpv loaded all user and system scripts. Flatpak mpv's bundled mpris plugin blocks forever without DBUS_SESSION_BUS_ADDRESS, for example when an MCP server is started by stdio_client without env=. mpv then answers IPC but never plays and ignores SIGTERM and quit. Headless mpv (both the IPC and the stdin-playlist launch) now gets --load-scripts=no. Windowed mpv keeps the user's scripts.
4. MpvIpc.start() could leak mpv. If mpv connected but didn't answer observe_property, start() raised without closing it. It now closes mpv first.
5. False success. Chunk errors are now recorded in StreamingAudioPlayer.playback_error. A blocking speak raises ToolError with the reason, instead of answering "played in 0.7s" after mpv reported "audio output initialization failed".

Test plan

• 11 new test cases plus one extended existing test (the stdin-playlist mpv now gets its own session). On the old code, 10 of the new ones and the extended one fail. The other new one checks that windowed mpv still loads scripts, which this PR leaves unchanged. One test kills a real sh wrapper's process group and checks that its child dies too.
• Full pytest suite passes; no new mypy or flake8 findings
• Manual: MCP speak (blocking) through a plain flatpak mpv, started by stdio_client with its default minimal env:
  • As shipped: fails in 2.4 s with "playback failed: audio output initialization failed", and no mpv is left. Before: hung for over 120 s and left mpv running.
  • With --load-scripts=no stripped out to force the mpris hang: fails after the 10 s start timeout with "mpv did not start playing within 10s", and the stuck mpv is killed through its process group.

🤖 Generated with Claude Code