Add MP4 DASH/CMAF subtitle support (wvtt, stpp, tx3g in fragmented MP4)

[niksedk]

Summary

Subtitles in MP4 DASH/CMAF files (fragmented MP4) mostly failed to load. This PR makes all common DASH subtitle shapes work — WebVTT (wvtt), TTML/IMSC1 (stpp) and MPEG-4 timed text (tx3g) — whether the file is a subtitle-only DASH representation (init + .m4s segments, concatenated or onDemand single-file), a muxed fragmented MP4 with video/audio, a lone .m4s segment without its init, or a progressive MP4.

What was broken

Problem	Effect
Fragmented WebVTT used the movie (mvhd) timescale instead of the track's mdhd timescale	All DASH-WebVTT cue times skewed (5/3 too late with GPAC's 600/1000 layout)
subt handler + stpp samples unsupported	TTML-in-MP4 tracks completely invisible, progressive and fragmented
Moof kept one traf, Traf one trun, tfhd track ids unused	Muxed fMP4 yielded no subtitles at all
IsmtDfxp parsed each mdat as a single XML document	mdats holding several TTML sample docs were silently dropped (partial extraction)
moov/mvex/trex defaults not parsed	Last DASH segment's cue truncated/dropped when trun/tfhd omit sample durations
seconv 10 KB minimum for MP4	Subtitle-only DASH files (a few KB) never reached the MP4 parser
CmafParser divided ticks with integer math and preferred mvhd	.cmft cue times truncated to whole seconds, and in seconds instead of ms

What this PR does

• Mp4Parser: fragmented samples are now sliced precisely per track — all traf/trun boxes are kept, each fragment resolves its track via tfhd track id, sample byte ranges come from data offsets + sizes, and timing comes from tfdt + sample durations in the track timescale (with per-track continuation when tfdt is absent, as in MP4Box -frag output). Missing durations/sizes fall back to mvex/trex defaults.
• Sample decoders: wvtt cue boxes (vttc/payl/sttg, incl. position info), stpp TTML documents (deduplicating cues repeated across segment documents; zero-based Smooth-Streaming-style docs are shifted by the sample start), and tx3g length-prefixed text. Lone .m4s segments without an init are sniffed by content.
• Progressive TTML: subt handler recognized; stpp sample parsing added to Stbl, so TTML tracks appear as normal subtitle tracks (incl. the track picker).
• IsmtDfxp: delegates to the box-aware parser when it finds TTML cues; otherwise splits concatenated TTML documents before XML parsing.
• seconv: small MP4 files (> 100 bytes) are tried against the MP4 parser, falling back to the text loader as before.
• CmafParser: correct timescale preference and millisecond math.

Testing

• 10 new unit tests build minimal init/media segments programmatically: track-vs-movie timescale, lone segment, muxed multi-traf, TTML document times + dedup + zero-based shift, progressive subt/stpp, trex defaults, and TTML document splitting (incl. ttm: metadata red herrings and namespaced roots).
• Verified end-to-end against GPAC (MP4Box) DASH packages — live profile (init + 8 .m4s), onDemand profile, -frag muxed files, and progressive muxes, for all three codecs: every variant now extracts all 6 test cues with timings identical to the progressive reference (previously: wrong times, 3/6 cues, or nothing).
• All existing suites pass: libse 831, seconv 290, libuilogic 140, UI 1270.

🤖 Generated with Claude Code