Problem:
-Downsamplers were treating reduced reads the same as normal reads,
with occasionally catastrophic results on variant calling when an
entire reduced read happened to get eliminated.
Solution:
-Since reduced reads lack the information we need to do position-based
downsampling on them, best available option for now is to simply
exempt all reduced reads from elimination during downsampling.
Details:
-Add generic capability of exempting items from elimination to
the Downsampler interface via new doNotDiscardItem() method.
Default inherited version of this method exempts all reduced reads
(or objects encapsulating reduced reads) from elimination.
-Switch from interfaces to abstract classes to facilitate this change,
and do some minor refactoring of the Downsampler interface (push
implementation of some methods into the abstract classes, improve
names of the confusing clear() and reset() methods).
-Rewrite TAROrderedReadCache. This class was incorrectly relying
on the ReservoirDownsampler to preserve the relative ordering of
items in some circumstances, which was behavior not guaranteed by
the API and only happened to work due to implementation details
which no longer apply. Restructured this class around the assumption
that the ReservoirDownsampler will not preserve relative ordering
at all.
-Add disclaimer to description of -dcov argument explaining that
coverage targets are approximate goals that will not always be
precisely met.
-Unit tests for all individual downsamplers to verify that reduced
reads are exempted from elimination
95b5f99feb