=== step 1/2: experiment.py (axis gate + PCA/low-rank reconstruction, 9-fold x 3-seed) ===
[    0.0s] === axis calibration gate (ds15, subject A) ===
[    0.0s] axis0(ap): ref-span=120.13mm all-span=124.58mm max|disp|=10.56mm
[    0.0s] axis1(si): ref-span=150.18mm all-span=161.80mm max|disp|=20.17mm
[    0.0s] axis2(lr): ref-span=75.13mm all-span=83.77mm max|disp|=10.54mm
[    0.0s] cyc1-own-t0 vs cyc3-ref-t0: mean_3d_offset=3.81mm max_3d_offset=9.46mm (plan anchor: mean 3.81mm, max 9.46mm)
[    0.0s] PASS: axis calibration gate ok

[    0.0s] loading ds5 subjects A and B ...
[    0.8s]   subject A: n_points=8219 cycles=[1, 2, 3, 4]
[    0.8s]   subject B: n_points=10125 cycles=[1, 2, 3, 4, 5]
[    0.8s]   subject A ROI 'liver_dome': 197 points within 20.0mm
[    0.8s]   subject A ROI 'liver_inferior': 252 points within 20.0mm
[    0.8s]   subject A ROI 'liver_lateral': 230 points within 20.0mm
[    0.8s]   subject B ROI 'liver_dome': 220 points within 20.0mm
[    0.8s]   subject B ROI 'liver_inferior': 254 points within 20.0mm
[    0.8s]   subject B ROI 'liver_lateral': 233 points within 20.0mm
[    0.8s] total folds (leave-one-cycle-out): 9
[    0.8s] 
=== PASS 1: spacing selection sweep (default K/sigma/alpha) ===
[    1.0s]   fold A/cyc1 done
[    1.1s]   fold A/cyc2 done
[    1.3s]   fold A/cyc3 done
[    1.5s]   fold A/cyc4 done
[    1.6s]   fold B/cyc1 done
[    1.8s]   fold B/cyc2 done
[    2.0s]   fold B/cyc3 done
[    2.2s]   fold B/cyc4 done
[    2.4s]   fold B/cyc5 done
[    2.4s]   parallel-2sag spacing sweep pooled RMSE(mm): d=10:0.759, d=20:0.748, d=30:0.735, d=50:0.719  -> best d=50
[    2.4s]   parallel-2cor spacing sweep pooled RMSE(mm): d=10:0.744, d=20:0.738, d=30:0.733, d=50:0.726  -> best d=50
[    2.4s] 
=== PASS 2: canonical 5 arms, full metrics, 3 seeds x 9 folds ===
[    2.7s]   fold A/cyc1 full-metrics done
[    2.8s]   fold A/cyc2 full-metrics done
[    3.0s]   fold A/cyc3 full-metrics done
[    3.2s]   fold A/cyc4 full-metrics done
[    3.3s]   fold B/cyc1 full-metrics done
[    3.5s]   fold B/cyc2 full-metrics done
[    3.6s]   fold B/cyc3 full-metrics done
[    3.6s]   fold B/cyc4 full-metrics done
[    3.8s]   fold B/cyc5 full-metrics done
[    3.9s] 
saved run_raw.pkl, done with core computation.
=== step 2/2: analyze.py (results.json + figs) ===
=== PREDICTION TEST (27 paired fold x seed samples) ===
orthogonal dvf_rmse3d_mm per-sample mean=0.6465  parallel-2sag(best d=50) mean=0.6373
delta (parallel - orthogonal), positive=orthogonal better: mean=-0.0092mm (-1.09% rel)
sign test: 11/27 samples favor orthogonal (threshold <=19 for 'no strong pos direction')
paired bootstrap 95% CI on mean delta: [-0.0168, -0.0030]
rmse_lr_mm: parallel-2sag=0.5648  orthogonal=0.5792  ratio=0.975 (need >=1.5 for mechanism)
LR squared-error share of orthogonal arm total: 61.93% (need <15%)
cond_a(orth clearly better)=False  cond_b(parallel reverses)=False  cond_c(LR ratio<1.5, mechanism wrong)=True
=> prediction_outcome = refuted

wrote results.json
done: figs/arms_rmse3d.png, figs/paired_diff_27.png, figs/axis_breakdown.png, figs/spacing_sweep.png
done. results.json + figs/*.png written.