Audio is my specialism and I do a lot with discrete designs too, so I appreciate C3's (and C5) function - my initial thinking was that as the rail is 15V, a bit of DC imprecision might not be a problem. But, perhaps I should have asked about the size of the signal at Q5's emitter, and also the relationship between V and mod depth. Ignoring NFB (and finite Ao in each stage), there's nearly 50dB of AF gain there - meaning that 20mVRMS from the pot will make Q5 emitter swing 14Vpk-pk. Looking at the over-mod detect, it looks like the LED lights when Q5's emitter goes lower than around 0.7V. (But the quiescent point is around 8.4 volts, so isn't there a danger of the positive half-cycle might be clipping before over-mod is detected? Presumably it's slightly lower - things don't always follow the theory!). Anyway, yes, if you're encountering such large signals at the output of the two amps, then C3 and C5 save a lot of extra complications, like op-amps or other such heresy
Sounds like playing with the demod NFB was highly educational! I was never much good at control theory, but have got a long way with an experimental approach. By coincidence, I'm having dinner this evening with someone who understands RF and big transmitters, so will try to find out a bit more about this modulation scheme...
Anyway, thanks again. Lots of good ideas in here, and I've got a growing collection of AM-only radios that will need feeding for many years into the future
All the best,
Mark