Replace Haali with libmatroska

[CoffeeFlux]

Replaces the bundled Haali MatroskaParser.c with a C++ demuxer, agi::matroska::Demuxer (src/matroska.{h,cpp}), which MatroskaWrapper now uses for subtitle import and HasSubtitles.

Design

• libebml/libmatroska (system, or WrapDB fallbacks) parse the EBML header, Info and Tracks. Everything else (the top-level walk, SeekHead, Attachments, clusters and blocks) uses a small bounded EBML reader, so only the bytes that are needed get read.
• Opening a file stops at the first cluster and follows the SeekHead for metadata stored after the clusters, like the Haali parser did. It only walks the whole file if Info or Tracks still haven't been found. HasSubtitles runs on the UI thread whenever video is opened, so this matters on slow storage.
• Clusters are indexed lazily and read one block at a time, so memory use doesn't grow with cluster size.
• The demuxer returns plain values rather than pointers into parser state. It has typed errors (IoError, InvalidDataError, LimitError, ...), size limits, and cancellation.
• Requires libebml >= 1.4.4 and libmatroska >= 1.7.1, the bundled versions, so a system libebml is always new enough for the bundled libmatroska. Older distributions such as Ubuntu 22.04 (1.4.2/1.6.3) use the bundled copies of both.

Behavior changes from master

Subtitle timing. The demuxer reports timestamps as stored in the file. StartTime() is the earliest audio or video timestamp: what FFmpeg (and so LAV/MPC-HC) reports as the start time, and what mpv rebases to for Matroska in practice. Import makes that time zero.

• For nearly all files this is the same as master. The Haali parser subtracted the first block in the file, which is normally audio or video.
• Files with only subtitles are no longer shifted to start at zero. Fixes Error in loading the timing (and perhaps not only that) of the subtitles in an MKS. #91.
• A subtitle line before the first audio/video block no longer moves every other line earlier.
• Lines starting before the start of the file are clamped to zero, and lines ending before it are dropped.
• Aegisub's video timeline still makes the first video frame time zero (Lift assumption that video frame 0 occurs at time 0 #21). In files where video starts after audio (e.g. the 83 ms Crunchyroll delay), imported lines therefore look late by that delay inside Aegisub, as they do on master. They are exported with their original timing, and this becomes consistent if the video timeline moves to the file start (e.g. Fix timestamps #508).

Other changes

• Attachments are only located when opening a file. Their data is read on request.
• Tracks with encryption or an unsupported content encoding are reported as unsupported, instead of making the whole file fail to open.
• More than 32 tracks and track numbers above 255 are supported.
• Partially downloaded files: the complete blocks in a cut-off last cluster are read.
• Unknown-size (live) clusters are supported.
• ContentEncodingScope is honored, including a zlib-compressed CodecPrivate.
• Repeated Tracks elements are tolerated.
• The deprecated TrackTimecodeScale is applied as RFC 9559 §11.2 specifies.
• The decompression and total-size limits from 5a53f19 still apply: 16 MiB per packet, 64 MiB of subtitle data in total.

Import UI fixes

• Cancelling the subtitle-reading progress dialog works again, and no longer uses a dangling progress sink.
• An error while reading track information is reported instead of crashing.
• A malformed ASS packet gives an error instead of an uncaught bad_lexical_cast.

Testing

• tests/matroska/fixtures holds 14 small files. Most are generated by generate.py, which builds the edge cases directly as EBML. A gtest dumps everything the demuxer exposes for each file and compares it with the hand-written .behavior file next to it.
• The subtitle timing of the fixtures was compared against master's MatroskaParser.c. The remaining differences are the intended ones listed above.
• On a real 3.9 GB BD release (two ASS tracks, nine fonts), the output is identical to the earlier libmatroska-backed version of this PR. Every packet's start, end and size also matches ffprobe.
• Start times were compared with ffprobe and mpv, including MKVs remuxed from m2ts with large timestamps.
• Opening a file for HasSubtitles on a cold cache takes about 1 ms, compared with about 250 ms when walking every cluster.

Not tested: importing through the GUI (progress dialogs, cancelling, the track chooser).

Known limitations

• libebml allocates whatever size an element inside Info/Tracks declares, so a crafted file can make it allocate up to about 2 GB. This is libebml's behavior and is left as is.

🤖 Generated with Claude Code