The comments described the machine the rig was written on -- "this
machine has no display", "there is a real GPU here", an Android
emulator's GLX quirk -- which says nothing to anyone reading it from a
different checkout. What the reader needs is what the script supplies
and why, which is now all they get.
`sway`, `swaymsg` and, when `--shot` is passed, `grim` are checked up
front and named in the failure, rather than surfacing as a compositor
that would not start.
The `# shellcheck disable=SC2086 -- prose` directive did not parse, so
the suppression was not in effect; the prose moves to its own line.
Clean under shellcheck now.
A rendering claim about iris was verified by hand from another checkout,
because the compositor script and the input replay lived in ai-app's
submodule and not here.
`scripts/run-headless.sh` starts a headless sway on its own socket, runs
an example against it and screenshots the result. `rig-input`'s
`replay-touch` drives a recorded gesture in through Wayland's virtual
pointer, since a headless compositor has no input device to move.
`iris::harness` gains the `.touch` parser, so a recording means the same
thing replayed into a harness as into a window rather than being read
twice by two parsers.
The ai-app copy's `--phone` became `--mode`, since which phone is not
iris's business; its `IRIS_SCALE` has nothing to hand a density to here,
so it waits for one.
iris-ai
iris