Dedup duplicate join notifies + reject joins on dead broadcaster

[queueeee]

Symptom

Live deploy at 1080p60 / 6 Mbit / viewer_limit=5: bot connects fine, stream allocates, then on the very first viewer click the bot goes into a tailspin. The log told the story:

22:53:01.780  notifyjoinstreamrequest clid=72   ← first
22:53:01.785  notifyjoinstreamrequest clid=72   ← +5ms
22:53:01.787  notifyjoinstreamrequest clid=72
... 9 retransmits within 16 ms
22:53:01.888  video_broadcaster.subscribed total_subscribers=1
22:53:02.276  video_broadcaster.subscribed total_subscribers=9   ← !!!
22:53:02.302  rss_mb=1343                       ← 1.3 GB RSS for one viewer
22:53:26.554  video_broadcaster.source_ended    ← ffmpeg/x11grab OOM-killed

The TS6 server retransmitted the same notifyjoinstreamrequest 9 times inside 16 ms. Our existing dedup checked self._viewers.pop(...) before the slow path (createOffer + ICE gathering = tens of ms), but the viewer slot only lands in _viewers after that slow path completes. All 9 retransmits raced past the dedup, each spawning a fresh peer connection + broadcaster subscription. The 9× concurrent ICE gatherings + encoder subscriptions blew RSS to 1.3 GB and OOM-killed the capture pipeline. After that, the broadcaster's pump task exited but subscribe() kept handing out queues nothing would ever fill — so subsequent viewers' UI hung on "connecting" forever (= "kommt nicht wieder").

Fixes

StreamPublisher._joining — a lock-protected set[int] populated before any expensive work (offer creation, ICE, broadcaster subscribe) and discarded in a finally block. Duplicate notifies for the same clid log stream_publisher.duplicate_join_dropped and bail out cheaply without touching aiortc or the broadcaster.

VideoBroadcaster.is_alive — False after the pump task exits for any reason (source MediaStreamError, encoder crash). When false, _handle_viewer_join sends respondjoinstreamrequest decision=0 and exits — the TS6 client UI fails the join cleanly instead of hanging on "connecting".

Test plan

• pytest — 250 passed, 1 skipped.
  • New: test_duplicate_join_for_same_clid_only_creates_one_pc bursts 9 duplicate notifies and asserts exactly one PC, one broadcaster subscription, one audio relay subscription.
  • New: test_join_refused_when_broadcaster_source_dead flips is_alive=False and asserts decision=0 in the response with no PC, no subscriptions.
  • New: test_is_alive_* covers the broadcaster's lifecycle flag.
  • The test PC mock for createOffer / setLocalDescription now await asyncio.sleep(0)s — without it the in-flight dedup looked effective only because the mocked path ran to completion synchronously, but real aiortc's ICE-gathering yield is what lets parallel tasks observe each other's _joining state.
• mypy --strict clean.
• ruff check clean.
• Live: at 1080p60 click should now produce a single peer with stream_publisher.duplicate_join_dropped logs for the 8 retransmits, no RSS spike, no OOM.

Operator notes

Even with this fix, 1080p60 is still asking a lot from libvpx-realtime on a small VPS. If the bot stays alive after this PR but the encoder still can't keep up (frames pile up, RSS climbs), drop to 1080p30:

SCREEN_FPS=30

Or 720p60:

SCREEN_WIDTH=1280
SCREEN_HEIGHT=720
SCREEN_FPS=60
STREAM_BITRATE=4500

cd /opt/ts6-stream-bot
git pull
docker compose -f docker-compose.yml -f docker-compose.host.yml down
docker compose -f docker-compose.yml -f docker-compose.host.yml up -d --build
docker compose -f docker-compose.yml -f docker-compose.host.yml logs -f

https://claude.ai/code/session_016DuCjRJK995Tj9aDhhB9at

---

Generated by Claude Code