Have a closer look to an 'impulse' in music or a crescendo or sudden loud noise in any editing software.
You will see that such an 'impulse' spans multiple samples.
Well, since all real PCM systems must have an anti-alaising filter that starts to cut off below half the sampling frequency, you can not, ever, get a single "impulse", only a train of them. No real input into a real anti-aliasing filter should be able to create such a signal, because it contains frequencies that were filtered out.
Most "attacks" are much wider than two samples, in fact, because of the impulse response of the anti-aliasing filter. The closer the filter cutoff gets to half the sampling frequency, the longer that impulse response *MUST* be. As they say, it's not just the law, it's mathematics.
With a NOS DAC that impulse becomes 'jagged' and has extremely poor timing characteristics because it rises and drops at the wrong moment and stepped and too steep.
BUT at the output the proper anti-imaging filter will remove those 'edges' completely. So this is not actually a real issue if filtering is properly done. That word "properly" has some connotations, of course. Except, well, continue below, please.
So while some people like it and feel impulse response has improved (because it is said to be so) in reality the actual signal is 'f'ed up, has lots of ultrasonic crap that might become problematic. In most cases it isn't problematic and the hearing is so crappy and bandwidth limited that it isn't even that audible when average roll-off is 'compensated'. Only when 44.1 and 48kHz files are used. For higher sampling frequencies the EQ must be 'off'
Now you're talking about the (sin x)/x rolloff for a step function. Now this gets really complicated. But let's assume that the sinc compensation is literally perfect (HAH!) just for grins, ok?
In an r2r dac, let's assume that the MSB is perfectly accurate (HAH! fat chance), so are all other bits (***HAH***) and that there are no low-level nonlinearities as a result. It's a PERFECT r2r DAC in terms of level out vs. digital in.
That's going to give you a perfect output spectrum, right?
Well NOT necessarily. The dynamic behavior of each of all of the switches in the ladder must ALSO be absolutely identical (if you think level is hard, this is even worse). Furthermore, there must be zero capacative coupling from switch control voltage to the output r2r current. Yeah. Zero. Physics. That again.
So we put in a sample/hold after the DAC, so we can capture the "perfect" level with the SH system. I've spent many nights working with SH systems long ago. Aside from dielectric absorption, switch pedestals, droop, and input-level dependent jitter, all of which are bloody hard problems, yeah, let's assume you whupped all that, and it's PERFECT.
So then you're good, right?
Oops. No. IN ADDITION the waveshape of the "steps" must be a linear function, i.e. exponential approach to each level, NOT slew-rate limited approach.
Yes, I turned this into a discussion of why r2r DAC's are a (*&(*& to build. It wasn't really an accident.
The point is simple, 2x oversampling or 4x oversampling is a no-brainer for a bazillion reasons, and going to delta-sigma with good signal processing likewise, that way there's just no nonsense about low-level linearity, etc.
Such is life. It is possible to get an 18 bit Burr Brown DAC to 95.5dB SNR, including ground isolation, output buffering, etc. Been there, done that. Never EVER want to do that again. Yes, that's 16 bits, but that was the actual spec on the device, so I don't feel too bad about that.