feat(prebuilt): ffmpeg and OCR stacks for dk.nikse.subtitleedit

[jing2uo]

Eleven releases so dk.nikse.subtitleedit can stop building its dependencies into FlatPark's OSTree repository (flatpark/flatpark#238). That app's ref goes from 12.2 MB to 1.6 MB.

Cut individually rather than as one bundle, so an app can pin just the encoder it needs and a version bump touches one stack:

stack	packed	contents
ffmpeg-full	11.0 MB	ffmpeg n9.0 + ffprobe + libav*
tesseract	1.6 MB	engine + CLI + share/tessdata presets
leptonica	1.1 MB	libleptonica.so.6
x265	1.0 MB	libx265.so.216, 8-bit
x264	0.5 MB	libx264.so.165
sevenzip	0.5 MB	7zr, one dependency-free binary
rubberband	0.2 MB	librubberband.so.3
lame	0.1 MB	libmp3lame.so.0
libass	0.1 MB	libass.so.9 (0.17.4)
uchardet	0.1 MB	libuchardet.so.0

All dev-complete so ffmpeg-full and tesseract can build against them, and both clean their inputs out of their own archive — the consuming app pins those releases too and puts each lib/ on LD_LIBRARY_PATH, the split wemeet-screenshare-hook already uses for opencv-imgproc.

libass is deliberately its own release. mpv-stack builds its own copy, so an app shipping both ends up with two libass.so.9 contending for one soname and the winner decided by module order — the problem flatpark/flatpark#238 had to work around with post-install: rm -f libass.so.9.3.1. One release consumed from both ends settles it by construction. Follow-up: mpv-stack should consume this rather than build its own.

Decided by building, not by assumption

• rubberband picks its built-in FFT and resampler even though fftw3, libsamplerate and speexdsp are all in the SDK. Pinned explicitly with -Dfft=builtin -Dresampler=builtin so a change in upstream's defaults cannot silently add a DT_NEEDED. The result links nothing beyond libstdc++/libm/libgcc/libc.

• leptonica's dependency closure resolves completely inside org.gnome.Platform//50 — libtiff, libgif, libopenjp2, libwebp, libzstd all present.

• tesseract verified against leptonica-v1: the archive carries no libleptonica, and with only that release's lib/ added, bin/tesseract has zero unresolved sonames.

• QSV and AMF are not enabled in ffmpeg-full, and libvpl is deleted. Both are two-layer: ffmpeg links a dispatcher or dlopens a vendor library, and the layer that encodes ships with Intel's media driver or AMD's proprietary stack, neither of which any Flatpak extension provides. Measured in a sandbox on an Intel Arc 140V with --device=dri:

  h264_vaapi   encode succeeded, exit 0
  h264_qsv     Error creating a MFX session: -9
  h264_amf     DLL libamfrt64.so.1 failed to open
  

  VAAPI encoded through the same /dev/dri/renderD128, so the GPU was never the issue. libvpl was cut for QSV and then deleted, manifest and release both: no app in this ecosystem can consume it usefully, and keeping it with a warning would be a trap with a sign on it.

• --enable-libharfbuzz and --enable-libfontconfig: freetype alone stopped being enough for drawtext in FFmpeg 7+, and without fontconfig it only takes fontfile=/abs/path. Verified: drawtext with font=sans renders and tesseract reads the text back exactly.

• ffplay is disabled, and the cp of each input archive skips its manifest.json — the first cut shipped ffplay (putting an unwanted command on the consuming app's PATH) and libvpl's build record renamed to a misleading manifest-base-1.json.

Also here

wemeet-screenshare-hook now pins opencv-imgproc-v2 for its build — v1 is the unstripped 140 MB tree. The released hook artifact is not re-cut: re-pinning it would change com.tencent.wemeet's commit and make every user re-download a 197 MB payload for no functional gain. opencv-imgproc-v1 stays until the hook is next cut for a reason of its own.

README: the ten new stacks in the table, the extra-data list brought up to date, and the strip rule recorded.

🤖 Generated with Claude Code