I would suggest that it is best to start using devices and topologies that have the best "inherent" linearity and then to seek to improve the linearity first.

Ideally we have the output stage the dominant source of distortion and apply just enough loop feedback around it to reduce this distortion to what we find acceptable.

My definition of "acceptable" is at least 10dB below the distortion of Speakers and similar in nature, if we do not actively exploit distortion cancellation between amplifier and speaker.

Now I know no dynamic, electrostatic or planar speaker that has much less than 0.1% THD (either 2nd or 3rd HD dominant) at 1 Watt input and non that have much below 1% at rated power.

So I would suggest that "acceptable performance" is one that produces no more than 0.3% 2nd HD and 0.1% 3rd HD at rated power below clipping and less than 0.03% 2nd HD and 0.01% 3rd HD at 1 Watt, with the distortion below 1 Watt behaving monotonic (distortion falls with level) and preferably also above 1 Watt.

It should be noted that well designed Class A output stages easily match the above figures without NFB, while optimised Class AB output stages only exceed the above by less than 10dB or so.

---

I think global feedback is perfectly acceptable, where care has been taken for the amplifying circuit to not suffer bandwidth modulation with signal levels (degeneration and local feedback can be used to produce such results) and the necessary stability requirements etc. have been fulfilled and where thermal memory distortion inside the loop is reliably attenuated by the feedback loop (and not adds increased problems).

Degeneration has limits, it is my experience that beyond a certain point degeneration is less effective in linearising circuits than looped feedback, plus it invariably raises the output impedance, which can impact open loop bandwidth (though not slew rate).

Global looped feedback has more potential problems, but it is probably much easier to design (it's design effort approaches "brainless" these days) and at least in measured terms it can hide a multitude of sin's, including excessive VAS distortion etc., so people feel they do not need to address them locally. It is an easy option to good specs. The results can sound good, however IME the results of carefully optimising the whole circuit and making do with less global lopped NFB sound better and measure reasonably similar.