Seconds -> HH:MM:SS for human-readable reports.

Parse an SRT file into Entry objects. drop_long: skip entries longer than MAX_ENTRY_SEC (Gemini chunk errors).

Hearst depth score: how deep each valley is vs its surrounding peaks.

Cohesion at each gap between entry i and i+1 (len == n-1).

Extract one frame at time t. -ss AFTER -i for accurate seek (CLAUDE.md).

Transparent 0-100 pre-rank. Mirrors rubric weighting at a coarse level.

Cheap black-frame check via ffmpeg signalstats avg luma.

Partition entries into runs split wherever a silence gap exceeds SILENCE_GAP. Cutting never crosses a long pause, so no clip carries dead air.

Recursively split a segment that exceeds MAX_SEG_SEC at its weakest gap.

TextTiling within a single silence-bounded run.

Tile frames into one PNG. Robust method: normalize each frame to an identically-sized numbered file, then run ffmpeg's `tile` over the sequence

selection.json holds source timestamps; on a concatenated final video the cut points are the cumulative durations. We reconstruct them from dur.

Main entry: entries -> list of topic Segments respecting duration bounds. Two-stage: (1) hard-split on silence gaps so no segment spans a long pa

Claude's scoring worksheet: pre-ranked by heuristic, full text included.