I agree about the water-filling having the effect of spreading bits over every feasible tone. The one technicality that I wonder about is when power backoff is in play - whether it works in opposition to water-filling sometimes.
In the downstream case, I don't think we see this kind of opposition. The power is reduced by a fixed amount for all lines, which then has a fixed impact to the SNR values, and so a fixed impact as to whether a tone is "feasible" or not. Water-filling will still use as many tones as it can; it just doesn't get many bits in the affected tones.
In the upstream case, I'm less sure. The power reduction isn't fixed, so it is plausible that, sometimes, a power setting for a nearby subscriber causes it to use the U2 band in preference to U1 ... whereas higher power could allow U1 to be used. I'm unsure as to whether the dynamic system works in a way that accounts for the desired capacity. For example, would a 10Mbps subscriber would end up using lower power than a 20Mbps subscriber, simply because the extra power would end up merely giving superfluous excess SNRM?
(If it isn't obvious what I mean, consider a "water-filling" algorithm that starts with the lowest power on every tone, and then starts to water-fill by gradually turning up the power across all tones until enough bits have been allocated.)