0dBFS in floating format is a bit of a misnomer. It's not "full scale" in floating point, instead it's actually defined as the value
1. This is the value that when converted to fixed point matches 0dBFS in the fixed point format, by design. It's the highest level any audio converter will input or output.
In other words, going over 0dBFS in floating point does not imply clipping. Instead, it means "this would clip if sent to the audio interface."
When converting audio from fixed to floating point, the highest value you will ever get is 0dBFS. For 16-bit converters, this corresponds to 32767, which is the highest value for a signed 16-bit integer. (Note that you can also see 0dBFS at the bottom of a wave, which corresponds to the value -32767, in 16-bit fixed format.)
While you won't normally see real clipping inside your DAW (or any floating point audio processing), it's still best to keep internal signal values below 0dBFS. This is because some audio algorithms are very complicated and have to be adjusted from the theoretical equations to avoid infinites and infinitesimals. (That is, some matrix math can work perfectly "on paper" but fail when coded because intermediate values exceed the range of double-precision floating point. So, the algorithm is optimized or adjusted to avoid these cases.) When doing this adjustment, it can help to optimize the adjustments to work best when the signals are in a given range, and that range is usually in the normal audio range, which is say -60 to 0dBFS.
You won't see this issue for simple audio algorithms like delays and convolution reverbs, but you could see it for more complex algorithms.