Mix until you're done and happy with the sound. Leave nothing for the mastering engineer to do. If you don't trust yourself and are leaving things for the ME to fix or finish, you're doing everybody a disservice.
That's fine, but the question was about what you want to
receive as a mastering engineer, where you
don't have control over the mix. I've received files that had clipping, files that were rendered at 16 bits without dithering, and files that registered as 0 but had significant intersample distortion (like +3 dB!), so whoever mixed the thing didn't realize that even though the file itself wasn't distorting, they were mixing while monitoring a distorted signal through the D/A converter. If you trust the people doing the mix to deliver you something that's incredible, great. My experience is that isn't always the case. I ask clients to send me a version with master bus processing bypassed and another with their processing included. Often, I'll find the processed one has a stereo limiter. Just replacing it with a multiband one often makes the client
much happier.
But if you're mixing a relatively diverse collection of songs one by one, maybe in separate sessions, it probably is best to leave the final decisions about absolute levels and dynamic range to the mastering stage.
Spot on. The most recent album I mastered was Trigger, by Bryan Ferry's guitarist. It had been recorded in multiple studios, by multiple engineers. I didn't ask that there be no master bus processing, but the engineers delivered that anyway. The mixes were all good, but they didn't hang together in a unified way at all. Mastering solved that issue. I've had the same issues with compilation and live albums. I suppose purists could argue that if the cuts were wildly different, then that's the way it should be. I just don't happen to agree, but of course if that's what the artist wants, that's what they'll get.
Then render to floating point and don't worry at all (!!!) about the levels. Absolute meter readings - peak or RMS - mean absolutely nothing as long as they fall somewhere between like -300 and +300. If the ME needs to turn it up or down, they will. As long as it's not actually clipped, anybody who claims it's "too loud" or doesn't have enough "headroom" is actually just incompetent.
Again, I'm not delivering the mix, but I wish the people doing mixes followed your advice. I see clipping a fair amount, if not baked into the file, then levels that foster intersample distortion and therefore has the potential to influence the mix because people aren't hearing an accurate representation of the music. So I advise people to leave some headroom, which just about guarantees they're not going to run out of headroom when the signal hits the analog reconstruction filter in their audio interface (not to mention that the most non-linear part of D/A converters is at the very lowest and very highest levels - I know a lot of engineers, with platinum records on the wall to back up their opinions, who never peak signals much about -6 or even less because they swear it sounds better).
Still, use floating point just in case something does go over 0dbFS. Then the ME can decide how to smash that peak down rather than letting the truncation just chop it off.
That was actually one of the main points I was curious about. I've always been happy to receive 24-bit files, but I wondered if MEs were starting to ask for 32-bit or 64-bit FP. I don't know the answer to that one.