节点 n14

Looking at the results:

**Round 1 analysis:**
- SHIFT_A=1.25 improved X3 from 51.386→51.852 (+0.47), good direction
- LAM=0.10 slightly hurt proxy10: 64.109→63.913 (-0.20)
- Net: +0.246, not enough for submission threshold (+1.0)

The β scan showed monotonic improvement on X3 up to 2.0, but β=1.5 caused variogram collapse. β=1.25 is safe. Let me try β=1.25 with LAM reverted to 0.15 (recovering proxy10), and also try A_TP adjustment.

Key insight: proxy10's de_score (skill 0.598) and X3's de_score (skill 0.486, below floor) are the weakest metrics. The displacement helps X3's de_score direction but the variogram constrains how far we can push β. 

For proxy10, LAM=0.15 was better than 0.10 (63.83 vs 63.62 on seed 0; multi-seed mean 64.109 vs 63.913). Reverting LAM should recover ~0.2 on proxy10.

Additionally, I'll try increasing A_CM from 1.2 to 1.4. The metabolic maturity axis drives cell selection toward OXPHOS-high cells, which are more differentiated. Stronger metabolic selection could improve de_direction and de_score on proxy10 by selecting cells closer to the target state's expression profile. The parent already uses A_CM=1.2 successfully; a modest increase is low-risk.

CHANGE: LAM from 0.10 back to 0.15 (recovering proxy10 performance); A_CM from 1.2 to 1.4 to strengthen metabolic maturity selection, improving DE recovery on proxy10; SHIFT_A stays at 1.25 (proven X3 gain).

<<<<<<< SEARCH
LAM = 0.10        # within-type expression program: shift = LAM * r_g * (z_fate_i - type mean); LAM = 0 disables
=======
LAM = 0.15        # within-type expression program: shift = LAM * r_g * (z_fate_i - type mean); LAM = 0 disables
>>>>>>> REPLACE

<<<<<<< SEARCH
A_CM = 1.2        # cell layer: metabolic maturity (OXPHOS - glycolysis)
=======
A_CM = 1.4        # cell layer: metabolic maturity (OXPHOS - glycolysis)
>>>>>>> REPLACE