Analyze a clip range and emit a subtle per-clip correction filter. Returns (filter_string, stats_dict). The filter is bounded to ±8% on any axis

Return the ffmpeg filter string for a preset name. Empty string for 'none'.

Resolve a path that may be absolute or relative to `base`.

Final pass: base → overlays (PTS-shifted) → subtitles LAST → out. If there are no overlays and no subtitles, just copy base to out.

Format a duration as "Ms" or "Mm SSs".

Format a time in seconds as "NNN.NN" with fixed 6-char width for alignment.

Sample N frames from a range and compute brightness/contrast/saturation stats. Uses ffmpeg's `signalstats` filter which gives us YMIN, YMAX, YAVG

Run two-pass loudnorm on input_path, write normalized copy to output_path. Returns True on success, False if measurement failed (caller should fa

Build an output-timeline SRT from per-source transcripts. - 2-word chunks (break on any punctuation in between) - UPPERCASE text - Output

Extract the audio segment and return an RMS envelope of length `samples`. Uses ffmpeg to dump mono 16kHz PCM to a temp wav, then computes a w

Lossless concat via the concat demuxer. No re-encode.

Extract every EDL range into edit_dir/clips_graded/seg_NN.mp4. Returns the ordered list of segment paths. If the EDL `grade` is "auto", analy

Extract N frames evenly spaced across [start, end]. Returns paths in order.

Extract a cut range as its own MP4 with grade + 30ms audio fades baked in. `-ss` before `-i` for fast accurate seeking. Scale to 1080p from 4K.

Find gaps >= threshold seconds inside [start, end] between kept tokens.

Walk a Scribe word list, break into phrases on silence >= threshold OR speaker change. Returns list of {start, end, text, speaker_id}. Scribe

Return True if the source uses a PQ or HLG transfer function.

Return True if the video's height > width (portrait / vertical).

Run ffmpeg loudnorm first pass and parse the JSON measurement. Returns a dict with measured_i, measured_tp, measured_lra, measured_thresh, ta

Return (header_name, duration, phrases) for one transcript file.

The EDL's 'grade' field can be a preset name, a raw ffmpeg filter, or 'auto'. Returns the filter string to embed into the per-segment -vf chain.

Transcribe a single video. Returns path to transcript JSON. Cached: returns existing path immediately if the transcript already exists.