There is an entire bunch of not very well known things going on with this circuit.
The K596 FET is from a capacitor microphone and has current of about 300 uA at gate voltage=0. If you use an ordinary FET you will need some negative bias, you can experiment with various LEDs between the source and ground.
Q3 acts as an upper cascode transistor, Q2 acts as a folded cascode transistor. An increase in current through the FET causes Q3 to conduct more, a decrease in the current through the FET causes Q2 to conduct more.
Q2 directs some of the FET signal to provide Armstrong oscillator type positive feedback.
Q3 directs some of the FET signal to a current mirror based AM detector.
The audio output may need some additional RF filtering before going to an audio amplifier and a supply rail decoupling capacitor would be good.
I'm just experimenting with this circuit in the AM band with good results, how it holds up at different frequencies I don't know. Anyway technically it seems good in terms of sustaining amplifier phase lag or lead, impedance matching etc. I don't have LTSpice at the moment to fully check. The built behavior seems fine.
Not to detract from your regen experiments, but have you ever considered designing a regenerative superhet? I think that would offer much better performance without much added complexity, but still there are a lot of interesting design problems to solve, like how to prevent the varying RF front-end tuning from affecting the loading on the regenerative stage and the LO tuning, how much IF gain is needed, what IF amplifier configuration to use, avoiding oscillation in the IF strip, choice of mixer and LO drive level, avoiding or minimizing spurs, etc. I built a regenerative superhet circuit (posted elsewhere here on the forum) that works well, but requires a lot of gain stages due to running off of a very low 1.2 volt supply voltage. I think that with a higher supply voltage like 3, 6, or 9 volts, it is probably possible to make a simpler but still very good regenerative superhet that should be able to easily exceed the performance of a straight regen.
Maybe this simpler circuit could work but you would need a lot of audio ampification.
The 1H audio inductor could be a cheap EI14 600 ohm audio transformer with the 2 winding connected in series with the correct phasing.
Each winding is about 130 ohms DC and 290mH, 800 turns.
The saturation current is only 1 or 2 mA so you have to be careful.
If you can tolerate a pintrest pin I wrote a little more about the circuit:
If you want to simpify the AM detection you can get rid of Q4 C2 and connect the base of Q3 to positive and the top of the RFC to positive and then connect your own AM detection to the collector of Q3. Or even you could make the RFC an audio frequency inductor by passed with say a 22n capacitor. I think you would need a lot of audio amplifier gain in that case.
There is a lot going on in that circuit and there are many variations possible. I also have "noisy regen" variations to explore.
I'll be a month of Sundays sorting all that out.
Anyway I want to see how general the negative resistance obtained from a cascode - folded cascode combination is. I presume that connecting the collector of Q2 directly to the resonant tank circuit would work. However then R2 should really be a large value, maybe 20meg+.