That's explained in one of the threads linked: "I spawned two Yoshi's by hitting the block with Mario and the p-switch at the same time. I jumped on one Yoshi, got the p-switch in his mouth and let him die, so the second - invisible - Yoshi becomes visible and have a null sprite in his mouth."
Most emulators are not accurate. This makes some game's behavior quite unexpected. Here is an article describing the issues and how so much processing power has to be devoted to achieve full rendering
They have, but note the corollary to 'perfect emulation takes huge amounts of computing power': imperfect emulation can often be done cheaply.
The question for WBE is whether brains are the very rare SNES game which must be perfectly emulated to work at all... or one of the others. The success of machine learning stuff like deep belief networks, while using a fraction of the brain's computing power and minimal biological plausibility, suggests human brains aren't very special snowflakes.
Far from it. Many emulators over the years have had small but important quirks, and although they generally don't affect normal gameplay they _may_ affect this sort of thing, after all this is depending on the precise behaviour of the PRNG and how many times it has been called.
There are techniques out there for running TAS input files against real hardware for most of the common platforms. Some of the TASes on tasvideos.org are specifically flagged as "console verified" [1], indicating that someone's confirmed that they don't depend on emulator quirks.
