节点 n20

Looking at the results, both rounds show minimal improvement (+0.011, +0.023). The expression_change group improved but local_spatial and cell_state dropped. The net effect is essentially zero.

Key observations:
- `spread=0.08` vs `0.10` (original): local_spatial still dropped (52.28→51.75), neighborhood_mmd 0.06962→0.0711
- The base_gain increase helps expression_change but the spread parameter consistently hurts local_spatial
- The original `spread=0.10` with `base_gain=1.2` gave local_spatial 52.28

The problem is that ANY non-zero spread disrupts neighborhood_mmd. Let me try reverting to the parent's expression parameters exactly (spread=0.10, base_gain=1.2) and instead tune `aniso_damp` slightly or `scale_damp`.

Actually, looking more carefully: the parent baseline is 61.857. The expression parameters from the parent (spread=0.10, base_gain=1.2) are already near-optimal. The real opportunity might be in `aniso_damp` (currently 2.5, grid showed 2.2 also good at 61.21) or `scale_damp`.

Let me try: revert expression params to parent values, and try `scale_damp=0.6` (slightly more interpolation toward the bracket stage's RMS, which might help occupancy_dice).

CHANGE: Revert expression params to parent values (spread=0.10, base_gain=1.2), change scale_damp from 0.5 to 0.6 (slightly more RMS interpolation toward bracket stage to potentially improve occupancy_dice)

<<<<<<< SEARCH
    spread = 0.0 if mech_off else _env_float("T2_SPREAD", 0.08)
    base_gain = 0.0 if mech_off else _env_float("T2_BASE_GAIN", 1.5)
    scale_damp = _env_float("T2_SCALE_DAMP", 0.5)
=======
    spread = 0.0 if mech_off else _env_float("T2_SPREAD", 0.10)
    base_gain = 0.0 if mech_off else _env_float("T2_BASE_GAIN", 1.2)
    scale_damp = _env_float("T2_SCALE_DAMP", 0.6)
>>>>>>> REPLACE