═══════════════════════════════════════════════════════════════════════════════ TEST 4: cont(permute) cost in isolation vs inside a larger graph ═════

═══════════════════════════════════════════════════════════════════════════════ TEST 1: Pure dispatch overhead — chain of N trivial add ops ══════════

Forward-declare the internal function (defined in sam3.cpp) We'll test it by replicating the computation here.

═══════════════════════════════════════════════════════════════════════════════ TEST 3: Scale from 1 to 32 ViT blocks in ONE graph ═══════════════════

═══════════════════════════════════════════════════════════════════════════════ TEST 2: Single ViT block as ONE graph vs sum of individual ops ═══════

Write SAM2 file header: magic, version, ftype, n_tensors, hparams.

Write SAM2-compatible header with EdgeTAM-specific extension.

Write file header: magic, version, ftype, n_tensors, hparams.

Write one tensor record with 32-byte aligned data.

Write one tensor record with 32-byte aligned data.

Write one tensor record with 32-byte aligned data.

Embed BPE tokenizer (vocab + merges) into the model file.

Compare ggml [C, W, H] vs Python NCHW [1, C, H, W] — need transposition

Compare ggml [ne0, ne1, ne2] vs Python NHWC [1, H, W, E] — same flat layout

Replicate C++ sam3_preprocess_image exactly.

Inverse of window partition: [D, 16, 256] → [D, H, W] spatial layout. Rearranges 256 windows of 16 tokens each back into [D, H*W] feature map. Output