16 bit certainly suffices as a delivery format. Many listening environments provide far less dynamic range (e.g. cars) and environments where the lowest level details aren't buried in the ambient noise floor are going to be few and far between. Even though it is possible to encode information below the noise floor in practice information that far down isn't going to be perceptible. (A practical definition of "perceptible" might be "able to be distinguished in an ABX test by a trained listener".)

As for how much safety margin is "enough" during initial capture... We're looking at a stark choice between either 2 or 3 bytes per sample. 2 bytes seems to be borderline, while 3 bytes seems to be plenty. It would be hard to quantify a safety margin more precisely than that.