I'm afraid you are doing things we don't recommend; also, we don't use FSL (ever) so we can't comment on trying to make AFNI match what FSL (or SPM) does -- I wouldn't even try. Both of these situations make it difficult to advise you.

I will mention that the combination of high-passing and using polynomial baseline regressors in AFNI is not a reasonable idea -- the non-constant polynomial baseline components are there to remove the slow fluctuations in the data, which is what the high-passing is for. So if you high-passed the data and then added a constant baseline back in, using '-polort 0' would make sense. Also, I would now recommend 3dTproject for bandpassing rather than 3dBandpass (but there will probably be little difference in the end).

I added the mean signal back in and globally normalized the signal to 10,000 because FSL's FILM (prewhitening) function appears to expect that scale of intensities. Does 3dDeconvolve/REMLfit have similar assumptions about the input? Do the data need to be scaled or saved as a certain datatype (i.e. float, int, double, etc.)?

The processing analysis in AFNI is somewhat different than FSL, in that the "high-passing" (removal of polynomial drifts in our case) and the pre-whitening (in 3dREMLfit) and task GLM are all done at the same time. I've never tried removing the baseline drift, except for a constant, then running 3dREMLfit -- the idea is antithetical to the "fit it all in one model" approach I take in that program, and so I can't really predict the results. The closest analogy to FSL that I would recommend in AFNI would be to eschew the polynomial regressors in AFNI and instead put in trigonometric regressors to pick up the low frequencies you want filtered out, and let the linear regression + pre-whitening model do it all at once.

I'm afraid I can't comment further on your attempt to force AFNI into the FSL model (or vice-versa).