CrystalNNFingerprint: No Voronoi neighbors found for site - try increasing cutoff. Should I treat as NaNs or average the OK sites?

An example of a structure that fails (only on site 6):

Structure Summary
Lattice
    abc : 12.350696 14.104949458808422 14.527473876232783
 angles : 63.92566223904422 79.2990282889014 65.62253736027132
 volume : 2070.508763122277
      A : 12.350696 0.0 0.0
      B : 5.821764 12.847438 0.0
      C : 2.697509 5.787997 13.048756
    pbc : True True True
PeriodicSite: Hg (5.8268, 12.8480, 0.0037) [0.0004, 0.9999, 0.0003]
PeriodicSite: H (9.7351, 5.6024, 8.8724) [0.5786, 0.1297, 0.6799]
PeriodicSite: H (13.8559, 6.0049, 4.5341) [0.8994, 0.3109, 0.3475]
PeriodicSite: H (8.0213, 12.4316, 10.1972) [0.1886, 0.6156, 0.7815]
PeriodicSite: I (10.7995, 9.5062, 6.3019) [0.5227, 0.5223, 0.4830]
PeriodicSite: O (14.0980, 9.7276, 9.5889) [0.7801, 0.4261, 0.7349]
PeriodicSite: O (12.7394, 13.2700, 4.1544) [0.5427, 0.8895, 0.3184]
PeriodicSite: O (8.7643, 5.5144, 8.7059) [0.5033, 0.1286, 0.6672]
PeriodicSite: O (7.3719, 8.8273, 3.1464) [0.2716, 0.5785, 0.2411]
PeriodicSite: O (13.4883, 5.9796, 3.6148) [0.8710, 0.3406, 0.2770]
PeriodicSite: O (7.9270, 12.6247, 9.2341) [0.1743, 0.6638, 0.7077]

Wondering if I should treat this one as a row of NaNs or if I should take the average fingerprint of the sites that computed OK.

Happy to switch over to matminer discussion, but figured it was worth asking here due to the use of CrystalNN.

Related to decisions in Consider use of fingerprint distance instead of `StructureMatcher` for comparison between generated and test · Issue #39 · sparks-baird/matbench-genmetrics · GitHub