Update int4 accuracy numbers after PatchEmbed/RefineNet fixes

Add M2 Pro cross-machine validation note to performance section

Regenerate int4 weights after PatchEmbed + RefineNet fixes

Lead with sidecar story, explain int4 speed parity

Fix WiLoR-mini link: abcbdf → warmshao

Integrate Palm Daddy precise numbers: p99 column, stable-window label

Reframe: lead with consistency story, let table speak for itself

Add p95: MLX 3.8x better tail latency than MPS

Update live sidecar table with full MPS percentiles from 102K-row manifest

Update accuracy numbers: 0.024 → 0.006 after PatchEmbed fix

Regenerate weights after PatchEmbed padding=2 fix

Regenerate weights after RefineNet ConvTranspose2d bias=False fix

Fix: near-identical speed wording, zero-arg Quick Start

Sync model card to zero-arg posture — consistent with GitHub README

Update Quick Start to auto-download API

Re-upload float32 weights after convert CLI fix

Fix Quick Start to include mano_path argument

Add float32 weights (2.4 GB, no MANO data)

Add int4 weights (490 MB, no MANO data)

Model card — MANO data not included, separate license

initial commit