Local media nodes DO run headlessly — say what each one needs

[mikkel]

What breaks today

references/graph-format.md marks 6 node types LOCAL, browser-only and then says:

the rows marked LOCAL/browser-only (resize, vframes, combine, soundtrack, trim, extractaudio) are browser-only media processing — the executors load such graphs with a warning and fail fast at run with UnsupportedNodeError, before any network call.

That is false. SKILL.md line 248 in the same skill already said the opposite, so the skill contradicted itself. An agent that reads the reference file drops a whole class of workflow it could actually build and run.

The truth, read off the executors

• nanoodle-js/src/local-media.mjs implements the ops headlessly. It runs a pure-JS path first (PNG resize via its own PNG codec, PCM-WAV trim, MP4CAT lossless mp4 concat) and shells out to ffmpeg/ffprobe on PATH for everything else.
• nanoodle-py/src/nanoodle/local_media.py has no pure path. Every one of the 6 node types shells out to ffmpeg.
• Neither library marks any node type unsupported. nanoodle-py's graph.UNSUPPORTED_TYPES evaluates to the empty tuple; nanoodle-js's graph.mjs has no equivalent flag and lists all 6 as local: true.
• UnsupportedNodeError is raised only for an unknown node type (nanoodle-js/src/workflow.mjs:183-188, nanoodle-py/src/nanoodle/workflow.py:208-215). The retired draw type is the live example.

What the change does

references/graph-format.md gets a per-node, per-language table:

node	Node CLI	Python CLI
resize	pure JS for a PNG source, else ffmpeg	ffmpeg
trim	pure JS for a PCM WAV source, else ffmpeg	ffmpeg
combine	pure JS lossless mp4 remux when every clip is mp4 with matching stream parameters, else ffmpeg	ffmpeg
vframes	ffmpeg	ffmpeg
soundtrack	ffmpeg	ffmpeg
extractaudio	ffmpeg	ffmpeg

The "refused up front, before any network call" sentence stays, narrowed to the case where it is true: an unknown node type.

The table cells lose the browser-only half of their (LOCAL, browser-only) tag and read (LOCAL media).

SKILL.md loses one imprecise clause. A missing-ffmpeg error does not arrive "before any paid call" — it arrives when that node's turn comes, so paid nodes upstream have already spent. The page now says so and tells the reader to prove the setup with a media-only graph first.

No node type was added or removed. The draw retirement is already complete in this repo.