Fix: enable WASM proxy for memory, add retry logic for chunk fetches

Fix: free download buffers after session creation, add GC pause between loads

Fix: reshape duration embedding from [1,1,768] to [1,768] for global_embed rank 2

Fix: remove cross_attn_padding_mask from DiT inputs (not accepted by model)

Fix: load decoder before text encoder to reduce memory pressure, pad tokens to 256

Fix: pad token IDs to 256 length with proper attention mask for text encoder

Fix: int64 tensor types, correct ONNX input/output names (global_embed, latents, out)

Fix: use path instead of name for externalData entries (onnxruntime-web API)

Add detailed error logging for model loading steps

Fix: use relative paths for same-origin model file access (CORS fix)

Fix: implement BPE tokenizer from tokenizer.json directly, remove broken @huggingface/transformers dependency

Fix: use regular script tags for onnxruntime-web, visible status UI, proper error handling

Add Allegretto Mini demo app — full inference pipeline in browser

initial commit