17 runs x 30 columns -> crashmetrics_features.csv
Columns: run, stiffness, clearance_mm, speed_kph, head_hic15, head_hic15_t1_ms, head_hic15_t2_ms ... head_dominant_hz, head_spectral_centroid_hz, error
Failed  : [('run17', 'ValueError: signal must contain only finite values.')]

run    stiff clear  kph  HIC15 rise ms width ms centroid  colour
run01    0.8    50   35    274    7.63     1.69     7.80  green
run04    0.8    80   40    281    4.47     8.93     6.72  green
run07    1.0    80   35    224    4.01     8.00     6.01  green
run10    1.2    50   40   1067    6.48     4.23     6.25  orange
run13    1.4    50   35    714    6.05     4.60     5.76  yellow
run16    1.4    80   40    856    3.36     6.72     5.08  yellow
Colour counts: {'brown': 1, 'green': 11, 'orange': 1, 'yellow': 3}

Strongest shape features vs HIC15: head_width50_ms -0.31, head_active_ms -0.30, head_end_ms -0.30, head_centroid_ms -0.29
Surrogate log(HIC15) ~ log(stiffness) + clearance + log(speed): exponents 1.90 (stiffness), 5.02 (speed)
Checked on 4 held-out runs: R^2 = 0.99, mean error 16 HIC

flatten(features of run01) has 29 columns; nested scores give keys like 'occupants.driver.regions.head_neck.points'
