The 2N3904/6 (in array form) and the HFA3096 were exactly the two ends of the spectrum that I was thinking of. I was playing (in simulation only) with the HFAxxxx arrays for some active differential probe designs a while back - must get back to it at some point.
They are rather nice little ICs. However quite fragile as I found
Well I won. The issue was decoupling (tggzzz will be laughing at this and the prototype layout) and missing zobel network which was causing it to squeg. After a little motorboating due to missing compensation capacitor I got the turd polished. Also found out even at way too loud a volume it'll quite happily drive an 8 ohm speaker with 3904/3906 output pair without getting dangerously warm so I swapped that in.
So schematic, a fairly standard discrete opamp. Feedback path / gain setting through R10/R9. C1 decreases bandwidth in the loop. C7 adjusts low end response. C6 adjusts high end response. Phase flips have no gain
Prototype:
Reasonable power gain from 600 into 8 ohms! It'll push out quite a bit more than this but this is loud enough.
Pretty good eyeballed linearity:
Fairly decent step response at the cone!
Now to carefully measure it across the entire response frequency span. Which I can't be arsed to do tonight.
Well that's the first thing I've built from scratch for a long time and I'm enjoying it so far.
I am missing that SDS1202X-E now though if I'm honest
Edit: ran some Queen through it and made a shit speaker box
Actually sounds ok through that feeper. I will invest in this idea further. It's a shit load better than an LM386 anyway.