docs(RESULTS): recipe arms closed — headwise Muon separated-negative, Sinkhorn flat

[ronaldtse]

What

Closes the TODO.impl recipe ledger with measured verdicts (full 1,200-para windowed protocol, teacher 2.2890 reproduced exactly, labels sha e70ce991, paired between-students bootstrap on final_preds):

• headwise Muon (run-014): 4.8164 — NEGATIVE, separated. Δ vs 2.1 = +0.2267pp, CI95 [+0.016, +0.411], p=0.017. The DeepSeek-sec-2.5 lever, positive at 671B, is significantly harmful on a 300M byte-student — a clean scale-boundary result.
• Sinkhorn embeddings (run-015): 4.5547 — FLAT. Δ = −0.0903pp, CI95 [−0.267, +0.069], p=0.858. Indistinguishable from AdamW on the tied tables.

2.1 (4.5701) stands as the frontier student. Every optimizer/architecture lever on the board is now measured; the only frontier mover on record remains data-side (teacher r5→r6→r7).