节点 n15

I'll analyze the parent node's method and parameters to suggest a tuning change.

The method uses composition reweighting with key parameters:
- `A_TP = -0.55` (type layer weight)
- `A_CM = 1.2` (metabolic maturity weight)
- `A_FATE = 0.6` (pseudotime weight)
- `K_STRAT = 5` (strata per type)

Looking at the performance breakdown, `de_score` on X3 is below the floor (skill 0.445 < 0.5), suggesting the method isn't capturing enough of the right cell states. The `de_direction` on proxy10 is decent (0.706 skill). 

The metabolic maturity coefficient `A_CM = 1.2` is the strongest cell-level signal. Increasing it slightly could sharpen selection of mature cells (which better match the E9.5 target), while reducing `A_FATE` slightly since diffusion pseudotime may be noisy. Let me try a modest adjustment: increase `A_CM` from 1.2 to 1.4 and decrease `A_FATE` from 0.6 to 0.5. This keeps the same mechanism but emphasizes the more robust metabolic signal over the noisier pseudotime signal.

CHANGE: Increase A_CM from 1.2 to 1.4 and decrease A_FATE from 0.6 to 0.5 to emphasize metabolic maturity over noisy diffusion pseudotime, aiming to improve de_score on X3 while maintaining direction signal.

<<<<<<< SEARCH
A_TP = -0.55      # type layer: lower mean proliferation -> larger share
A_CM = 1.2        # cell layer: metabolic maturity (OXPHOS - glycolysis)
A_FATE = 0.6      # cell layer: within-type diffusion pseudotime (more committed cells)
=======
A_TP = -0.55      # type layer: lower mean proliferation -> larger share
A_CM = 1.4        # cell layer: metabolic maturity (OXPHOS - glycolysis)
A_FATE = 0.5      # cell layer: within-type diffusion pseudotime (more committed cells)
>>>>>>> REPLACE