*Bump*
Improvements:
- As built, it was drifting at, ehh, a few Hz per minute or something like that. Not bad enough to be unplayable, but needs to be reset every song or two.
- Maybe the "gain" transistors (the top right PNP, and the NPN directly below it -- these drive the feedback capacitor and MOSFET gate) were leaky? I tried a smaller type: MMBTH10/81. Absolutely worse (slewing at 10-100 Hz/sec!). Well damn. Not surprising though, they're only 25V devices. Ok, maybe a high voltage RF part: I next tried BFQ221/241 (100V 100mA 1GHz complements -- pulls from a CRT monitor!). Same as original. Well, it's no improvement, but at least it's not the BJTs.
- Maybe the MOSFET is bad, leaky or something. Seems really unlikely, but whatever. Replaced with a 2N7000 that was handy (the TO-92 of the 2N7002). No effect.
- So... oh, I have a 0.01uF polystyrene cap in my junk box, what the hell, let's give that a try.
Now it's drifting maybe a few Hz over 10 minutes. Leaky freaking polyester cap! Could be dielectric absorption at work, too.
Lesson learned: polyesters may not be particularly low leakage. This S&H application doesn't need low absorption (it's sampling pretty much the same voltage every time), but it does need low leakage, and I didn't realize how sensitive that was! (A few Hz oscillator drift corresponds to mV of varactor drift, so out of a 30V supply with an average 15V bias, it's in that regime of precision stuff.)
Also shows that you can use BJTs (even pedestrian 2N3904/6s) for surprisingly good S&H, with a carefully thought out circuit.
Tim