Thank for clarification, so I can conclusion hash RMD160 is possible to collision the reason;
-snip-
but this just paper calculation and until full collision found this is only hypothesis, maybe ratio 1:2 between hash sha256 with hash RMD160 is like coordinate in curve secp256k1 which one X coordinate have two Y coordinate
Just don't get your hopes up on this if you're going to use this for something like the puzzle transaction
since you wont likely get a full collision of the HASH160 result even if you procure a high number of GPUs since that possibility of that is too low.
You only demonstrated a partial 48-bit collision which is expectedly common since it's relatively small.