METHODOLOGY: rotation A/B as prose explanation, drop the table

METHODOLOGY.md

@@ -14,16 +14,7 @@ We also deliberately ship it as **one self-contained file** (not a pack + extern
 
 Per DiT block linear (`qkv_proj`, `out_proj`, `fc1`, `fc2`):
 
-1. **QuaRot-style block-256 Hadamard rotation.** Multiplying the weight (and, at runtime, the activation) by a block-256 Hadamard matrix spreads per-channel outliers so a 4-bit grid captures the signal instead of clipping the spikes.
-
-   **Measured — same-weights A/B** (identical weights, shell, AdaLN, scene, seed, sampler and step count; the *only* difference is rotation on vs off; 832×480, 61f, `euler` 4-step, seed `757358688076805`, dense):
-
-   | block mode | lapvar (sharpness) | CLIP-IQA ↑ | MUSIQ ↑ | s/step |
-   |---|---|---|---|---|
-   | **rotated (shipped)** | **377.5** | **0.423** | **44.6** | 2.16 |
-   | un-rotated (plain NVFP4) | 220.9 | 0.386 | 39.4 | 1.83 |
-
-   Rotation wins **every** quality metric — lapvar **+71%**, CLIP-IQA +0.037, MUSIQ +5.2 — and visibly: the un-rotated render carries a milky rain-bloom haze that softens mid-ground detail and neon glyphs. Both variants are bit-deterministic across repeats (SSIM 1.0), so the gap is real, not seed noise. **Cost:** the runtime Hadamard adds **~15%/step** and gives a *different* draw for a given seed (it reorders which weights land on the fp4 grid). **Honest nuance:** un-rotated NVFP4 is *not* broken — at enough steps on easy/low-contrast scenes it looks fine (many community "stock nvfp4" H3 files do) — but in a like-for-like isolation rotation is the sharper half, which is why we ship it. And despite the rotation cost, our full path is at **parity** with a community stock-nvfp4 file on the same card (10.51 vs 10.37 s/step @768p/61f, +1.4% noise — nunchaku's W4A4 kernel is efficient and high-res is attention-bound), while running **~1.9× faster wall-clock** via the 4-step distill.
+1. **QuaRot-style block-256 Hadamard rotation.** Multiplying the weight (and, at runtime, the activation) by a block-256 Hadamard matrix spreads per-channel outliers so a 4-bit grid captures the signal instead of clipping the spikes. We confirmed it in a clean same-weights A/B (identical weights and settings, only rotation toggled): rotation wins every no-reference IQA metric and lifts a Laplacian-variance sharpness proxy ~70%, and it shows — the un-rotated render carries a milky bloom that softens mid-ground detail and neon glyphs, while rotated keeps crisp edges. Both runs are bit-deterministic, so the gap is real. The cost is ~15%/step and a different per-seed draw. Honest nuance: un-rotated NVFP4 is *not* broken — at enough steps on easy scenes it looks fine (many community "stock nvfp4" H3 files do) — but like-for-like, rotation is the sharper half, which is why we ship it. And despite the rotation cost, our full path is at parity with a community stock-nvfp4 file on the same card (+1.4%/step, noise) while running ~1.9× faster wall-clock via the 4-step distill.
 2. **NVFP4** — fp4 (e2m1) packed 2-per-byte + a **group-16 fp8-e4m3 microscale** (336 group-scales per channel on the big linears). The fine per-group scale is what preserves detail; a coarse 1-scale-per-channel int4 (~2.3× the quant error on outlier-heavy weights) does not.
 3. **bf16 islands** kept unquantized: patch/time embed, AdaLN/modulation, token_refiner, proj_in/out. Small, sensitive, not worth quantizing.