Prashant and I discovered that the implementation of the wrapping Iterator in listLocatedStatus at https://github.com/apache/hive/blob/2d2f89c21618341987c1257a88691981f1f606c7/shims/src/0.23/java/org/apache/hadoop/hive/shims/Hadoop23Shims.java#L350-L393 is broken.
Basically, if you had files (a,b,_s) , with a filter that is supposed to filter out _s, we expect an output result of (a,b). Instead, we get (a,b,null), with hasNext looking at the next value to see if it's null, and using that to decide if it has any more entries, and thus, (a,b,_s) becomes (a,b).
There's a boundary condition on the very first pick, which causes a (_s,a,b) to result in (_s,a,b), bypassing the filter, and thus, we wind up with a resultant unfiltered (_s,a,b) which orc breaks on.
The effect of this bug is that Orc will not be able to read directories where there is a _SUCCESS file, say, as the first entry returned by the FileStatus.