节点 n14

I'll analyze the situation carefully.

**Current state:** Parent node baseline is 57.116. I need to gain >1 point. Round 0 (reg=0.1) failed (-0.375). Round 1 (winsorization) failed due to SEARCH block mismatch.

**What I know from the parent's notes:**
- κ=0.25 is on a flat plateau (0.2–0.3 → 56.6–56.9)
- covariation is the weakest component (dropped from ~47 to ~44 with OT)
- The winsorization idea was suggested but never tried
- reg=0.1 hurts slightly

**The key insight:** The per-cell correction `κ·scale·C_j` is added *after* the type-level displacement `df` is computed, but both go through `max(0,·)` clipping together. The issue is that `scale = α·r` multiplies the per-cell correction too. For X3, r=0.25 (target is 0.25 days after last input), so scale=1.75*0.25=0.4375. The per-cell correction magnitude is κ·scale·||C|| ≈ 0.25·0.4375·12.5 ≈ 1.37 per cell on average.

**A more promising approach:** Instead of winsorizing (which failed to match), let me try reducing `--ot-min` from 10 to 5. This would include more cell types in the OT correction (aSHF has 7 source cells currently marked "too-few"), giving more types per-cell diversity. But aSHF only has 9 output cells, so the impact is minimal.

**Better idea:** Try `--alpha` from 1.75 to 2.0. The parent node 8 tuned α=1.75, but that was *before* the OT correction was added. With the OT per-cell correction, the type-level displacement might benefit from being slightly stronger to maintain the covariation structure. But this is risky.

**Actually, the most promising untried change:** Lower `--ot-reg` to 0.02 or 0.03. Sharper OT coupling → more concentrated barycentric displacements → δ_j becomes more informative (less diffuse noise). The parent tried reg=0.1 (worse) but never tried *lower* reg. Lower reg makes the coupling more peaked, meaning each stage-2 cell is matched to fewer stage-1 cells, making the displacement more specific. This could improve both cell_state and potentially direction.

Let me try `--ot-reg` from 0.05 to 0.02.

CHANGE: Decreased --ot-reg from 0.05 to 0.02 (sharper entropic OT coupling produces more concentrated barycentric displacements, making per-cell corrections more specific and less noisy; reg=0.1 was worse but lower reg was never tried).

<<<<<<< SEARCH
    parser.add_argument("--ot-reg", type=float,
                        default=float(os.environ.get("VEC_OT_REG", "0.05")))
=======
    parser.add_argument("--ot-reg", type=float,
                        default=float(os.environ.get("VEC_OT_REG", "0.02")))
>>>>>>> REPLACE