I tried to resist replying to this, but I've hit my head with my palm too much whilst reading this, so here goes.
16 bit audio files store integer values between -32768 and 32767.
24 bit audio files store integer values between -8388607 and 8388607.
Neither of them are "fixed point" - they have no fractional point at all.
The dynamic range of 24 bit is better than you ears could ever cope with, and actually most A/D devices can't practically deal with greater than a 20 bit resolution in any case.
For all practical purposes, storing your files at 24 bit as as good as you'll get.
As has been mentioned before, D/A converters deal with INTEGER values - either at 16 bit (-32768 to 32767) or 24 bit (-8388607 and 8388607), so even if you have everything stored as 64fp it has to be converted back to 24 bit or 16 bit integer values for playback in the end.
The 32 bit fp and 64 bit fp part has nothing to do with the audio recording. It is the number representation used when doing calculations. Because they are floating point representations, they can deal with fractions. They are useful when either mixing several tracks together, processing sound (i.e. plugins), and volume changes, as they go some way to limit rounding errors during these calculations.
Note however, that you're still dealing with values between -32768 and 32768 (for 16 bit), and -8388607 and 8388607 (for 24 bit). The dynamic range does not increase by converting to 64 bit fp, as everything gets rounded to an integer when it's played back. What you do get, is a decrease in cumulative rounding errors during mixing and processing, as during the calculations fractional results are preserved until the final conversion back to 24 bit integers, and obviously 64 bit has more precision (and hence less rounding errors) than 32 bit.
For these range of numbers, floating point numbers perform pretty well. What they're not good at is accurately representing very large or very small numbers. For those extremes, you start to lose accuracy with floating point representations.
So for a single track the signal flow goes 24 bit integer --> process the audio using 64bit fp to avoid rounding errors --> convert back to 24 bit integer for playback.
For multiple tracks, you take each 24 bit integer from each track --> convert to 64 bit fp and process / mix them all together (this part involves adding them all together plus some division to get back to your -8388607 to 8388607 range) --> finally convert the stereo mix to 24 bit integer for playback.
If you're only ever mixing in the box (i.e. the audio files are never processed outside of Sonar/Cakewalk) then there's absolutely no reason to store things as 64 bit fp. Using 24 bit files, with 64 bit processing is absolutely fine.
The ONLY reasons I can see an advantage to using 64 bit fp files are:
1. You've drastically reduced the volume of a track, and bounced the result. In a 16 bit or 24 bit file, there's a theoretical loss in quality (I say theoretical, because I seriously doubt your ears would tell the difference except in extreme cases)... and in any case, why bounce the result of a volume decrease in the first place?
2. You want to process a half-processed file in another program, and load it back into Sonar/Cakewalk. Again, unless your volume is pretty low, I don't see how you would get much of an advantage here. -8388607 to 8388607 is still a pretty big range.
So as long as your tracks are at a decent volume (to get the most out of the -8388607 to 8388607 range), and your using 64 bit fp for processing, I can't see any practical reason to actually store your audio files in 64 bit fp format.
M.