test: structural checks for the WebGPU compositor WGSL, close out the README GIF

[BernardJen]

Covers #62 (as far as is possible without hardware) and closes #49.

#62 — the shader still cannot be executed here

Re-checked rather than assumed. navigator.gpu exists, but requestAdapter() returns null by every route — default, low-power, forceFallbackAdapter: true, and SwiftShader forced on the command line. So the compositor that shipped dormant has still never rendered a frame.

Leaving it wholly unverified isn't great, so the WGSL is now exported as COMPOSITOR_WGSL and checked structurally:

Check	Why it matters
vs / fs entry points exist	the pipeline names them; a mismatch is a runtime pipeline error, not a compile error
all three @bindings declared	the JS builds entries for 0/1/2; a missing one fails bind-group creation
texture_external + textureSampleBaseClampToEdge	the zero-copy import is the entire premise of #62
Layout struct fits 32 bytes	the buffer is allocated at 32; a larger struct truncates silently
quad array holds 6 vertices	pass.draw(6) would index out of bounds otherwise
Y flip present	compiles perfectly and renders the video upside down
balanced braces, no stray backticks	a backtick inside the template literal has broken this build twice

These are the failure modes most likely to be sitting in code nobody has run — each would surface as a pipeline or link error on the first machine that has a GPU.

Confirmed the tests actually bite, by mutation: renaming the vertex entry point and removing the Y flip each fail exactly one test.

This is explicitly not a substitute for running it, and the test file says so in its header. #62 stays open until it can be benchmarked on real hardware — enable gpuCompositing: true in settings.json at the wall.

#49 — closing

The issue asked for "a screenshot/GIF", and its stated rationale was "a single image will communicate the product better than the feature list". The dual-view screenshot added in #131 does that and is live on the README.

The GIF half needs the app running against real capture hardware. I can't produce that, and a fabricated one would misrepresent the product. This removes the commented-out placeholder rather than leaving dead markup in the README — adding a real clip later is a two-line change and doesn't need an open issue to track it.

Verification

150 tests (was 142), lint clean, build clean, npm run shadercheck 60 runs / 0 failures.

🤖 Generated with Claude Code