The mapping quality of the read you highlighted is 60 (bottom of screenshot) and I don't see visual differences between the reads. So the MQ filter is "failing" because, despite the many mismatches, those are high-quality alignments. My best guess is that, since you've aligned to a closely-related but not identical reference, that the multiple-mismatching reads are from sequence not present in the reference to which you aligned. Also, to my eye, the non-reference allele appears to be occurring on both reads with no other mismatches and reads with lots of mismatches, so that specific variant is "good".

You could perform a pre-pass eliminating reads with edit distance (NM) > 10 (or 8 or whatever). Or, with 118x coverage, you could pass your reads through an assembler and align whole contigs, identify those that seem not to be present in the reference, and include them as "decoys" to mitigate these mappings.