a) Scrapped the way in which we constructed candidate haplotypes because it wasnt fully correct and yielded corner conditions with incorrect genotyping and likelihood computation. Ideally, a haplotype should "cover" the read and the most likely alignments should be such that the ends of the read are inside the ends of the haplotype. This wasn't happening, and if you have a "dangling read off a haplotype" the probabilistic alignment model may prefer to shift a read instead of scoring it correctly - this is especially bad with tandem repeat insertions.
So now, we build haplotypes based on the reference context and adaptively change them based on read alignment positions, plus some padding and uncertainty in the alignment.
b) Changed the way soft clipped based are dealt with. Instead of either ignoring them or using them, we only use them if the read start or end position (after soft clipping) are within eventDistance of the current location. This is done because it's very common that BWA's strictly local SW implementation will soft clip every single read at an insertion position because it couldn't place that end of the read without too many mismatches, but the read is legit and the bases are good quality. If we don't take these bases into consideration, reads which are informative of an insertion event are essentially discarded because the informative part is clipped away.
c) Several cleanups and fixes to the context-dependent gap penalty model based on length of HRun.
git-svn-id: file:///humgen/gsa-scr1/gsa-engineering/svn_contents/trunk@5464 348d0f76-0448-11de-a6fe-93d51630548a
b45afe5ba8