Files
feuerwerk/tools
adminandClaude Opus 5.5 30b618d9a9 Show game view recordings in their true colours in other players
A game view recording looked darker and its colours off in VLC, and
each round of recording a recording made it worse. Chrome's encoder,
fed RGB frames, wrote full-range H.264 and left the colour out of the
stream. Only our Matroska head said "full range", and VLC reads only
the stream. Chrome itself plays the converted MP4, which was right.

- The GPU now turns the game view into I420, limited range BT.709,
  before it is read back (PictureReader). The encoder gets the planes
  with that colour, as it does for a recording of the whole editor.
- Where the planes lie in the read-back picture is a plain function,
  i420Layout in recording.ts.
- Only 1.5 instead of 4 bytes a pixel come back from the GPU.
- avcColour.ts writes the colour into the SPS of the encoder's avcC as
  well, so a player that reads only the stream knows it. An SPS with a
  VUI only gets its colour block replaced. One without gets a VUI that
  holds the colour. An SPS it cannot read stays as the encoder made it.
- A first try sent the frames through a 2D canvas. That gave limited
  range too, but only 18 to 20 of 33 frames at 1080p30.

Found on the way:

- With the colour in the stream alone, VLC 3.0.8 still showed the
  full-range picture too dark. It ignores the range.
- The one old full-range original on the server got the same SPS fix
  without loss: the clusters are byte-equal, the decoded frames are
  equal (framemd5), backup in /root/config-backup-20261013/videos/.
- Older game view recordings stay full range. Fixing that in VLC would
  mean encoding them again.

Tests:
- tsc, eslint, Prettier clean.
- Gate on melektaus-atx: CHECK EXIT 0, 308 files, 4772 tests.
- avc-colour.test.ts new (9), expectations read back with ffmpeg
  trace_headers. Counter-check: the old bug in the no-VUI path turns
  it red.
- e2e 67: ALL PASSED, 12 checks. 67d-ii reads tv/bt709 from the file
  and from the bare H.264 stream, red with the old code.
- 1080p30 34 frames, 640x360@60 68 frames, none dropped (was 33/47).
- Checked by the user: a new recording looks right in VLC 3.0.8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-13 19:48:44 +02:00
..
Plugins belong to the project, and they load before it: the project file gained the field plugins: ["demo-plugin.js"] (default empty, no formatVersion bump), the Plugins tab's switch writes into it as an ordinary undo step (SetPluginActiveCommand) instead of into a server-wide plugins.json, and the start order is turned around - startPlugins only fetches the list of files, then editor.boot() runs and ProjectStore.load(id, prepare) hands the raw JSON through before it is deserialized, where pluginsOf(raw) reads the list and PluginManager.syncTo() activates exactly those; that is the point of the issue, because what a plugin registers has to stand there before normalizeGraph would throw it away as unknown (#126, #127), a project switch first stops what the new project does not name and then starts what it does, a file the server does not have is passed over silently, and after signing out the manager is dead so a late event cannot start anything in a torn-down session; on the server plugins.json, PluginStore::setActive, isActive, the route PUT ?r=plugins/{file} and the 403 for inactive modules are gone and GET ?r=plugins only answers {file,bytes,updatedAt}, which makes every plugin file fetchable - the consequence accepted in #117, written down in the wiki for #115 and #116; publishPlugins(root) only copies files now, the seeder switches the demo plugin on in the fixture project, and the Plugins tab carries a puzzle piece as its icon while only a plugin's own tab still goes without one (the knob and the socket are cubic curves, not arcs, which Pixi's parser draws as a spike after an L); 3 new e2e checks in module 38 and 5 new unit tests cover the project as the owner of the switch, its undo, the order, pluginsOf against crude input and an unknown file name (#119)
2026-09-20 14:24:27 +02:00