节点 n26

Analyzing the results: alpha=0.1 hurt across the board (57.267 vs 57.710 baseline). The d2_shape dropped from 8.33 to 8.27, occupancy_dice from 3.42 to 3.18, and neighborhood_mmd barely changed. So jitter is confirmed harmful at any tested level. Reverting to alpha=0.0.

Now, what else to try? Key weak metrics: variogram (skill 0.218), occupancy_dice (0.410). The strong metrics (d2_shape, scale_log_ratio) shouldn't be disturbed.

Looking at the parent's mechanism table: k=15 is best. The parent noted k=10 (56.60) and k=25 (56.69) were worse. So the k=15 sweet spot is real.

Let me think about `scale_damp`. Currently 0.5 (geometric mean of RMS). The `log_interp(rms_a, rms_b, t, damp)` function likely computes `exp((1-damp)*log(rms_a) + damp*log(rms_b))` then applies damping. With damp=0.5, it's the geometric mean. If the embryo grows between stages, the true target RMS might be closer to one side. Trying damp=0.6 biases toward stage b (later, presumably larger). This could slightly improve d2_shape and scale_log_ratio if the true target is closer to the later stage's scale.

But actually, the safest change with the most potential upside is trying k=14 or k=16. The parent tested k=10, 15, 25 — never k=14 or k=16. Since the scorer uses exactly 15 neighbors including self, k=15 is theoretically optimal. But k=16 means each block has 16 cells, so the 15-NN query always finds 15 same-block neighbors (one extra cell in the block that might not be in any given cell's 15-NN window). This could slightly improve neighborhood_mmd by ensuring that even edge cells have 15 same-stage neighbors available.

Actually with k=16: a cell at the edge of a block might have only 10 same-block neighbors, and its 15-NN query would include 5 from other blocks. With k=15, an edge cell might have only 9 same-block neighbors. So k=16 gives slightly more "buffer". But the parent's data showed k=25 was worse than k=15, suggesting larger blocks don't help. The issue with larger blocks is probably that they create bigger gaps between blocks (fewer blocks total, less spatial coverage), hurting occupancy_dice and variogram.

Let me try k=14 instead. With k=14, blocks are slightly smaller and more numerous, giving better spatial coverage (more blocks spread across the embryo). This could help occupancy_dice (skill 0.410) and variogram (skill 0.218) by having more evenly distributed cells, at minimal cost to neighborhood_mmd (a cell in the interior of a 14-cell block would have 13 same-block neighbors out of 15 queried — still very good).

CHANGE: T2EI04_K from 15 to 14, slightly smaller blocks for better spatial coverage (more blocks = more even spread), potentially improving occupancy_dice and variogram while keeping neighborhood_mmd high since 14-cell blocks still provide ~13 same-stage neighbors per cell.

<<<<<<< SEARCH
    k = _env("T2EI04_K", 15, int)
=======
    k = _env("T2EI04_K", 14, int)
>>>>>>> REPLACE