The two lanes are not comparable, and neither is promoted.
No all-zero / zero-norm sample exists.
The *_mag_data.csv / lsm9ds1_mag_data.csv family has spher_az/spher_elswapped (Known Defects): mars_4, mars_8, and all insane_mag_calib_raw/klagenfurt files (18 total). The px4_mag/lsm_mag family is correct-order.
On mars_4 and mars_8, the support-side *_mag_data.csv files are strict untrimmed time-supersets of the primary flight-window magnetometer files, not alternate streams. Exact pairs:
mars_4px4_mag.csv has 14975 rows over [1633592509.973706, 1633592684.6437771] while px4_mag_data.csv has 21227 rows over [1633592509.973706, 1633592757.3021052]
mars_4lsm_mag.csv has 3530 rows over [1633592510.3510547, 1633592684.6514337] while lsm9ds1_mag_data.csv has 11263 rows over [1633592510.3510547, 1633593066.6976933]
mars_8px4_mag.csv has 18361 rows over [1633698151.7142746, 1633698367.3847542] while px4_mag_data.csv has 31848 rows over [1633698151.7142746, 1633698523.5388005]
mars_8lsm_mag.csv has 4354 rows over [1633698152.355344, 1633698367.3977065] while lsm9ds1_mag_data.csv has 7262 rows over [1633698152.355344, 1633698511.0729492]
px4_mag_data_revised.csv (23 runs) carries the primary px4_mag vectors bit-identically, re-keyed onto the time-calibrated ground-truth timebase. Fuse against ground_truth_* using these timestamps.