The reason these tests are funny is partly because conditions must be generate from operands first, then a conditional branch is made from those conditions. The branch target decides how the future operands are calculated, and so propagates any errors onwards. This was to make the test as sensitive as possible without much code.
However it makes the recent in/out tracking more complicated and not a good fit vs the original code.
The opcode reporting currently dumps the opcode sequence starting with the op which generates the conditions (ftst/fcmp) and operand tracking reports the input to those. The actual branch is a few ops later in the sequence.
I've been busy today to do much with it but will see if I can make it a bit simpler for diagnostics.
D.