HaplotypeCaller PRIORITY_LOW bug
At some positions (such as 20:10001430 in src/test/resources/org/broadinstitute/hellbender/tools/haplotypecaller/expected.testGVCFMode.gatk4.g.vcf) reads with mismatches or indels in the BWA-aligned pileup get realigned during local assembly (where they will get evaluated as variants in the active region). However, the flanking regions adjacent to the active region are evaluated using the original alignments. This can lead to het calls outside the active region that get capped to GQ0 homRef calls.
The solution is not entirely trivial since we hard clip reads after we trim the active region and compute the likelihoods used for their realignments based on this clipped sequence. One relatively easy hack that would be a partial improvement would be to omit reads that had been realigned from the pileup in the flanking regions.