Clean up debug diagnostics now that WebGPU works end-to-end

Use INT4 encoder (MatMulNBits — WebGPU-supported); ORT 1.23 keeps webgpu fp16 fix for decoders

Per-step diagnostics: pinpoint which ORT call crashes

Surface full error info on transcribe failure

Upgrade onnxruntime-web 1.20 → 1.23 (webgpu bundle); revert WASM forcing — 1.23 fixes fp16 transformer NaNs

Force decoders to WASM (WebGPU fp16 in ort-web 1.20 returns all-NaN on this transformer)

Diagnostics: logit type/len/health/first-8/top-5 with Number() coercion

Await tensor.getData() for WebGPU outputs (audio_embeds + logits) so data is actually copied back to CPU

fp16: use canonical u16 bit-pattern viewed as Float16Array; diagnostic top-5 dump

Encoder: prefer static INT8 (QLinearConv); INT4 fallback. Recovers 92.7%-class quality in browser

Float16Array for fp16 tensors; per-session WebGPU fallback (decoders stay on webgpu)

Encoder: try INT8 first, auto-fallback to INT4 if ConvInteger unsupported in browser

Use ort.webgpu bundle + auto-fallback to wasm if webgpu init fails

Bump cache key to invalidate RTN weights (GPTQ ship)

Use INT8 encoder + INT4 decoder (91.9% accuracy); force-English prompt default

Switch Space to static SDK: pure browser inference via onnxruntime-web