This is a case of the blind leading the blind, so watch my step. Which I guess translates to "Listen carefully, because I shall say this only once--and probably wrong".

The effect you are describing seems to be associated with the mysteries and magic of TARGs. Those temporary stack items that are created at compile time and reused by different subs/ops and even recursive re-invocations of the same sub. They are related to the well-known optimisation that means that some sub variables don't get cleaned up when they return, but rather persist to be reused next time.

There is some scant mention of them in the sections: "Putting a C value on Perl stack", "Scratchpads" & "Scratchpads and recursion" of Perlguts, but it is less than clear. At least to me when I encountered this a couple of years ago.

I found that if you ensure that subroutine return values are explicitly set sv_2mortal(), many of the traps "went away", but I fear that I may have been creating SVs that would never get cleaned up?

Hopefully, someone who knows what they are talking about will set the record straight, but maybe this mention will trigger understanding in your brain. I've personally got half a dozen ideas for modules that need XS but on whihc I can make no progress for my failure to understand this mechanism (amongst others:().