Past the first 30 s window, whisper.cpp's token offsets can run later than the word's own DTW end, which inverted 37 of 105 words on a 51 s clip and let zero-duration repair smear them into the next pause; such a word now starts at the previous word's end (0 of 105 inverted, end error unchanged). A transcript refused for missing alignment is also removed, because init captions from any transcript.json on disk. The metadata harness now checks the whisper-cli arguments and the refusal.