Inconsistencies like randomly repeated atom indices, maxed out indices at 99999, integer overflowing indices to the negatives (yup, there's a program that has this bug), bad spacing, jumbled order etc. in a PDB are tolerated differently in different programs. Some will carp at missing atoms or non-standard atom names, which are not formatting errors.
In terms of atom indices, these are used in
CONECT entries, but several programs bond by proximity or by known topology —
CONECT entries are absent for canonical amino acids. Repeated residue indices and chain letter repeats cause issues with
LINK and often with peptide chain sequence.
However, in a multimodel PDB the numbers in each model are expected to be reset at every MODEL so if the atom indices reset after the end of a model this is 100% expected.