according to LTspice, the usage of Q5/6 (compared to resistive load only) does give better current matching through the LTP degen resistors, but at the expense of 2dB less OLG
I think the design wasn't intended to have high open loop gain. Probably, even without the zener diodes in the feedback path, there would be significant THD at moderate power levels already. On the other hand, this should also mean that the amplifier is quite stable (apparently it doesn't oscillate even with a purely capacitive load, and only exhibits some ringing).
That seemed like a better compromise to me at the time.
Yes, the circuit is so stable that I was even able to do without Miller compensation
As I wrote on my site, the amplifier is oversized, precisely because it is not a HiFi amplifier, but a guitar amplifier. Such amplifiers are generally subject to greater loads and should be equipped accordingly. If you really want to, you can get 50 watts from 2 mosfets without any problems. It's all a question of personal philosophy. But I repeat myself, today I would design a different amp.
I fail to understand why, though. It is very stable, it appeared to satisfy your demands, it is maybe overkill but it's of no consequence when it comes to power amps (it's just a handful of components more). Was it noisy or had bad sound? Of course, for a low powered amp, you might as well just do a simple single supply CFB amp with a very simple class A input and a class A driver. Maybe FET based.
Normal use of a current mirrors would reflect the collector currents of Q7 and Q8 to add to the VAS input, but here, instead the VAS itself is a "current mirror", with a limited gain. So, the OLG is reduced dramatically, so there is a not much feedback and therefore very stable. Current mirrors and a symmetric IPS are a troublesome mix, so this approach avoids that problem. The feedback Zeners could use each other instead of extra 1n4148 diodes. I would put a divider ahead of these Zeners so that you could program the clipping level with resistors or a pot. The addition of soft clipping avoids rail sticking and instabilities which are often the source of unpleasant sound. This is a classic current limit circuit that changes the output impedance when it limits. I think it's a better idea to attenuate the input when the current limit is reached, but I've not seen that done in commercial amps. Given the limited OLG, this probably works just fine. ie the LTP will not try to fight the current limit, much.
The fact that impedance changes when clipping (decreasing) could be a good thing per se, because that could mean that the bass and treble is dynamically attenuated, which can lead to a more "forward" and clearer sound. But, of course, it is a quirk. I don't know how well this would play with mixed mode feedback and what would the end result be.