Hrm. I dunno. I think it means that either the bounding box data is incorrect (which seems very very unlikely since you are reindexing right before querying, etc), or that MSVC doesn't compile cpBBSegmentQuery(). Those are the only two possibilities I can think of. The latter seems plausible since it uses a bunch of IEEE infinite (but not NaN) math bits to deal with lines that don't cross and such.
Do you know how to enable strict FP on MSVC? I guess I'd try that before trying to rewrite the function just to be sure. I haven't had issues with it on other platforms, and it's not doing anything particularly fancy. Maybe a compiler bug then? Either way, if strict FP doesn't fix it or ruins performance, then it will have to be rewritten differently I guess.
Another easy check would be to do a simple BB check of the ray segment against the child BBs. Wouldn't be great performance wise though, but would help narrow down the issue to cpBBSegmentQuery().