Silent bug dropped 42% of video shot scores for weeks before alerts flagged it

A routine alert on July 9, 2026 revealed that a vision-scoring function had been silently failing to score roughly 42% of video shots, a figure consistent with a quality audit conducted a week earlier on July 2. The root cause was a token budget of 300 set in shot_vision_qa.score_shot_frame(), too small for the Qwen3-VL model to complete its reasoning trace and return a valid JSON answer — mirroring a previously fixed bug in a related module. A separate defect in wan2_1.py was also found hardcoding a 300-second HTTP timeout, overriding a deliberately set 900-second limit designed for the video model's cold-load scenarios. Both issues were resolved by correcting the token and timeout values, and a new AST-based test was added to catch similar misconfigurations at boot time. Developers noted that all downstream quality signals built on the scorer had been operating on degraded data until the fix was applied.