As quite a separate point from above, I may have found a bug: I put a (?callstack :: CallStack) constraint on TcEvidence.mkTcTransCo and then put the same constraint on TcCanonical.rewriteEqEvidence. GHC complained about a redundant constraint on rewriteEqEvidence, and indeed its call information wasn't propagated. rewriteEqEvidence uses pattern guards and do-notation, but that shouldn't muck with CallStack, should it? I've not tried to reproduce this in a smaller test case.
After staring at your bug confusedly for a few minutes, wondering why I couldn't simplify it, I realized that the actual warning I was getting was in *mkTcTransCo*, not rewriteEqEvidence. mkTcTransCo does not in fact use a CallStack (by default) so the redundant constraint warning there would be correct. If I add a call to error, everything seems to check out. Can you confirm? Thanks! Eric