节点 n16

1. 在 X3（qiu_heart_early，T_raw > 1.5）上把组成外推的 T clip 从 2.0 降到 1.5、平滑 s 从 5.0 降到 3.0，mmd_u 降 8%（0.0316->0.0291）、variogram 降 14%（0.00171->0.00148），两项合计 +1.75 分；说明更保守的组成外推 + 更锐的频率趋势在真实两输入尺子上是同向收益。
2. 同一改动在 proxy10 上逐位零变化（四项原始值完全相同）：该尺子 T_raw 未触及 1.5 的上限，参数扫描只对真正走 two_input_path 外推分支的尺子有效，选 sweep 目标前应先确认参数在该尺子上是否 binding。
3. 总榜分 +1.30 小于 T1 约 2 分噪声，单靠一次总分不能判定 tune 有效；应看 score_parts 里目标项（此处 mmd_u、variogram）是否同向移动，本例两项都动且幅度超过各自地板分的几个百分点。
4. X3 的 de_score 原始值 -0.013 -> 0，是负值被指标定义截到 0 的结果，不代表 DE 排序变好；de 两项各 +0.11/+0.08 分应视为噪声内。
5. 耗时 54.8s -> 71.8s（+31%）而内存峰值不变 5.73GB：纯常数调整不应增加这么多时间，可能是重跑波动或 s 变小后重采样权重更极端导致更多迭代，tune 节点需留意耗时是否可复现。

1. 针对 X3：单独扫 T clip ∈ {1.2, 1.5, 1.8} 与 s ∈ {1.0, 3.0, 5.0} 的 3x3 网格（每次只动一个，避免本次两参同改无法归因），指标以 mmd_u、variogram 的原始值为主判据而非总分。
2. 针对 proxy10：先打印该尺子的 T_raw 与两输入时间间隔，确认走的是 two_input_path 还是单输入退路；若 T_raw < 1.5，则 proxy10 需要另一族参数（型内选择阈值、λ=0.15 增殖程序）才能改善，不要继续在这条分支上调参。
3. 针对 covariation/variogram（两把尺子上最弱项，proxy10 skill 恰为 0.500 地板、X3 0.513）：在组成重加权之外，加入保持基因间共变的型内细胞选择（按 ODE 速度范数选细胞而非位移表达），而不是逐基因独立偏移或加独立噪声——后者按评分规则会直接损害 mmd_u 与 variogram。
4. 针对 mechanism 判定：在 run.py 的 mech_off 分支下跑一次对照（关闭频率趋势重加权），确认 X3 的输出与得分确实随机制开关而变，以便后续节点能把机制收益与常数调参收益分开。

ANALYSIS.json written

REVIEW.json written

Looking at the results from round 0:
- The nonzero-only mask had zero effect on proxy10 (identical scores) — meaning the program wasn't creating new nonzeros where there were none (the shift was already mostly on nonzero elements, or the effect was negligible).
- γ=0.7 slightly hurt X3 (52.10 vs 52.21 baseline), particularly de_direction dropped (0.0696 vs 0.074).

So I should revert γ back to 0.5 and try a different direction. 

Key observations:
- X3 variogram skill is 0.487 (below floor 0.5) — the weakest metric
- X3 de_score skill 0.525, de_direction 0.530 — modest
- The smoothing parameter s=5.0 in the frequency trend extrapolation hasn't been swept
- T clip at 2.0 hasn't been swept

For X3 (qiu_heart_early), T = (t_target - t_last)/(t_last - t_prev). If T is large, the trend extrapolation is aggressive. Reducing the clip from 2.0 to 1.5 would make composition more conservative, potentially improving variogram (less distortion of cell-type composition).

Also, reducing s from 5.0 to 3.0 makes the ratio more responsive to actual frequency changes (less smoothing), which could better capture which types are expanding.

Let me try: reduce T clip from 2.0 to 1.5 (more conservative extrapolation), and reduce smoothing s from 5.0 to 3.0 (sharper frequency trend signal). Both changes target the between-type composition which affects variogram and mmd.

CHANGE: T clip from 2.0 to 1.5 (more conservative composition extrapolation), smoothing s from 5.0 to 3.0 (sharper frequency trend), to improve variogram and mmd on X3.

<<<<<<< SEARCH
    T = min(T_raw, 2.0) * float(os.environ.get("TSCALE", "1"))
=======
    T = min(T_raw, 1.5) * float(os.environ.get("TSCALE", "1"))
>>>>>>> REPLACE

<<<<<<< SEARCH
    s = 5.0
=======
    s = 3.0
>>>>>>> REPLACE