I know I said last year that it was debugged, but it wasn't. Now it is. But it's now checked out and verified from top to bottom including deriving the algorithm, the pseudo-code implementation, and the actual Propeller code. Backed up with spreadsheet comparisons between the algorithm outputs at each stage versus the base transform.

And it's still fast! 256 samples takes 645,520 cycles (31.7 KHz) and 1024 samples takes 3,534,656 cycles (23.1KHz) with full 32 bit precision. (Good enough for filtering.) I need to code up the 16 bit precision version, which shouldn't be too difficult, then put in some sample usage code.