节点 n9

Looking at the results from round 0:
- SHIFT_A 1.0→1.5: X3 improved from 51.386 to 52.242 (+0.86), confirming the monotonic β trend continues
- LAM 0.15→0.20: proxy10 improved from 64.109 to 64.300 (+0.19)
- Overall: +0.634, still below the +1.0 threshold needed

The X3 variogram at β=1.5 is 0.001482 (skill 0.518) — still above floor but declining. At β=2.0 it was 9.91/10 (below floor). So β=1.75 might be the sweet spot: more X3 gain while keeping variogram just above floor.

For proxy10, the λ improvement is small. Let me try λ=0.25 to see if the trend continues.

The weighted score formula is (proxy10 + 2×X3)/3. To gain +1.0 over parent (55.627), I need ≥56.627. Currently at 56.261, need +0.366 more.

If β=1.75 gives X3 ~52.7 (interpolating between 52.24 at 1.5 and ~52.7 at 2.0 based on the original scan pattern), and λ=0.25 gives proxy10 ~64.4, then: (64.4 + 2×52.7)/3 ≈ 56.6. That's marginal.

Let me try β=1.75 and keep λ=0.20 (since the λ gain is small and uncertain, don't over-tune two things at once on the proxy side).

Actually, the variogram concern is real. At β=2.0, variogram was 9.91/10 (floor is 10.0, skill 0.5). At β=1.5 it's 10.37/20 wait no — it's 10.37/20.0 for variogram score out of 20. The skill is 0.518. The floor is skill 0.5. So there's only 0.018 skill margin. At β=1.75, it might drop to ~0.505 or below floor.

Let me be more conservative: try β=1.6 instead. This gives a smaller step but safer on variogram.

Alternatively, I could try a different approach: keep β=1.5 but tune other parameters that affect the overall score. Looking at the breakdown:
- cell_state went from 58.06 to 60.00 (big gain from β increase, likely mmd_u improvement)
- covariation dropped from 54.01 to 53.33 (variogram getting worse)
- de_recovery improved from 52.05 to 52.64

The mmd_u on X3 improved significantly (0.03099→0.0289, skill 0.533→0.559). This is the main driver.

Let me try β=1.75. Even if variogram drops slightly below the current level, as long as it stays above floor (skill 0.5), the mmd_u and de_score gains should compensate. The variogram at β=1.5 has skill 0.518, and at β=2.0 it was below floor. Linear interpolation suggests floor crossing around β≈1.85-1.9. So β=1.75 should be safe.

CHANGE: SHIFT_A from 1.5 to 1.75 (continuing the monotonic X3 improvement trend; variogram at β=1.5 has skill 0.518 with floor at 0.5, so 1.75 should remain above floor while gaining more on mmd_u and de_score).

<<<<<<< SEARCH
SHIFT_A = 1.5     # two-stage per-type displacement: delta_g = SHIFT_A * (dt_out/dt_in) * (pb_last - pb_prev)_g; 0 disables
=======
SHIFT_A = 1.75    # two-stage per-type displacement: delta_g = SHIFT_A * (dt_out/dt_in) * (pb_last - pb_prev)_g; 0 disables
>>>>>>> REPLACE