[adri0]

Follow-up to #138. A game stopped by Season 4's two-hour limit (Adef vs Vswed, video 100) needed its game_end added by hand. Now extraction adds it.

What changes

• extract._detect_time_limit: finds the first sample where the game clock reads 2:00:00 or more (_TIME_LIMIT_S). The clock has to tick up to it: if the previous reading is more than _CLOCK_STEP_TOLERANCE_S below, it's a misread and doesn't count.
• extract._game_events: the per-game event steps (start, claims, winner, end), moved out of extract_video so they can be tested without a video. When the limit is reached:
  • Only claims seen by then count, plus _MARK_DEBOUNCE_SAMPLES, so a square already showing at the limit is still confirmed. The overlay freezes the clock, but a square handed in late can still turn on the board (Season 4 match 107).
  • If no line or unbeatable majority settled the game first, a game_end is added as the last event. determine_winner(..., time_up=True) then decides on the count: majority for the leader, a tie when level.
• The limit applies to every season. On scadustats-site, no game's clock goes past 2:00:01, and only Adef vs Vswed reaches it. The next longest game ends at 1:58:52.
• Docs: EventType docstring and CLAUDE.md.

Testing

• 6 new tests in tests/pipeline/test_extract.py: a win on the count, a tie, a claim after the limit dropped, a claim showing at the limit counted, a misread jump to two hours ignored, and a game settled before the limit keeping its own game_end.
• Full suite passes (533), Ruff clean.
• Not run against a real VOD: video 100's VOD was deleted.

Known limits

• An unmark under way at the limit (it needs 4 samples to confirm) is lost.
• Two misread samples in a row at two hours or more would still stop the game.

🤖 Generated with Claude Code