Beautify time codes: snap to the video's real frame times

[niksedk]

Fixes #10235.

The problem

"Beautify time codes" snaps cues to frame boundaries, but every SE5 call site passed TimeCodesBeautifier an empty time-code list, so it always fell back to nominal n/fps arithmetic. That grid is exact for genuinely constant-frame-rate material and wrong for everything else — VFR, remuxes, timestamp discontinuities.

SE4 did have the feature, extracting frame times with ffprobe -show_entries frame=pkt_dts_time. That was the original complaint in the issue: DTS is a decode/packet-domain timestamp, so with B-frames it is not the order frames are shown in, and it keeps the container's start offset (the reporter's own dump shows a ~1.9 s DTS→PTS gap). The port was dropped in SE5 rather than fixed, which took the feature with it and left ExtractExactTimeCodes as a dead setting.

The fix

Extract with ffmpeg's showinfo filter rather than ffprobe:

ffmpeg -i <video> -map 0:v:0 -an -sn -dn -vf showinfo -f null -

showinfo reports frames as they leave the decoder, so times are presentation-ordered, and without -copyts they are normalized to a zero-based timeline — the domain subtitle cues live in. It also needs no new binary: SE5 locates ffmpeg but never ffprobe.

Implementation notes

• Sorted and de-duplicated before use. TimeCodesBeautifier binary searches the list (ClosestIndexTo) and indexes it by frame number, and validates neither — an out-of-order list silently yields wrong frame numbers rather than an error.
• A partial list is rejected, not used. Because index maps to frame number, cues past the end of a truncated/cancelled extraction would snap to arbitrary frames instead of degrading to n/fps.
• Progress uses the shared FfmpegProgressTracker on stdout while showinfo writes to stderr, so the two channels never interleave.
• Extraction is async and cancellable; closing the window kills ffmpeg.
• Results are cached per video under Se.TimeCodesFolder — the full decode happens once (~1 s per 2 min of 720p on this machine).
• Batch convert uses a cached list when present; it never extracts on its own (that would decode every video in the batch).

UI

A checkbox, status and extract button in the Beautify window's top bar. Verified headlessly in all three states:

state
nothing loaded, extraction available	[ ] Use exact time codes from video No time codes loaded [Extract time codes]
extracting	[x] Use exact time codes from video Extracting time codes... [====43%----]
loaded	[x] Use exact time codes from video 72,014 time codes loaded

Verification

Against real ffmpeg on purpose-built fixtures:

• VFR clip — 24 frames at the true irregular 0 / .067 / .167 / .233 / .333 …, exactly matching ffmpeg. A nominal grid cannot express these.
• B-frame clip — 100/100 frames, strictly ascending (presentation order confirmed: frame n:1 is type:B at pts_time:0.04).
• Container with a 10 s start offset — extracted first=0, i.e. the offset that broke the SE4 DTS path is normalized away.

Unit tests cover showinfo parsing (including pts_time:N/A and progress lines), the sort/dedup guard, the incomplete-list guard, and an end-to-end beautify showing exact time codes land a cue on 233/900 ms (real frames) where the nominal 12.5 fps grid gives 240/880 ms (multiples of 80).

Full UI suite: 3493 passed.

English.json is regenerated from the language classes, so it also picks up already-merged speech-to-text strings that had not been regenerated yet.

🤖 Generated with Claude Code