pick a SAM device the GPU can actually run in objects_3d

[salmanmkc]

on a GPU that reports the WebGPU default maxStorageBufferBindingSize of 128 MB, Detect spams uncaught validation errors during "fitting boxes…" and takes a very long time to produce anything:

Binding size (402653184) of [Buffer (unlabeled)] is larger than the maximum
storage buffer binding size (134217728)
... While calling [Device].CreateBindGroup([BindGroupDescriptor "MatMul"])

that 402653184 is SlimSAM's attention buffer: 4096 tokens over 12 heads at fp16 is exactly 384 MB, 3x over the cap. onnxruntime already requests the adapter maximum, so it can't be raised — the device genuinely won't go higher. machines with a big limit (mine reports 4 GB) never see it, which is why it reproduces for some people and not others.

the device pick was 'gpu' in navigator && navigator.gpu ? 'webgpu' : 'wasm' with fp16 whenever webgpu was chosen, and the comment claimed transformers.js downgrades transparently. it doesn't when device is passed explicitly — it throws, and separately fp16 needs the shader-f16 feature that plenty of mobile/XR GPUs don't expose.

so probe the adapter and check both things before choosing, then try candidates in order: webgpu/fp16 → webgpu/fp32 → wasm/fp32. same model every time, so masks don't change quality depending on the device — only where it runs. cpu stays last so a driver failure the capability check didn't predict still lands somewhere that works.

also stops caching a rejected _samPromise. getSam() hands back the cached promise, so one failed load used to make every later Detect press fail until reload.

simulating the 128 MB limit locally: before, uncaught WebGPU errors and no clean completion; after, it routes to cpu and fits boxes with an empty console. an unmodified adapter is unaffected and still runs webgpu/fp16.

?mask=segmenter is untouched — still an explicit choice, never automatic.

same one-liner exists in the objects3d addon on #417, fixed separately there.

[salmanmkc]

@dli7319 does this fix the GPU errors in the console for you?

[dli7319]

The WebGPU errors are gone but the whole app completely freezes up now during detection when running on device 😅.

[salmanmkc]

The WebGPU errors are gone but the whole app completely freezes up now during detection when running on device 😅.

ya hard to test since my device doesn't hit this memory limit, seeing as it works but is just slower, I believe it has it's own fallback and forcing CPU blocks the thread, so I think best thing to do is leave the error message for now since it works

[dli7319]

I tested this PR on Quest 3 and the same WebGPU errors appear there