Is the bounce really that bad? I always got the impression that was normal, and these things weren't supposed to be agile or anything, just set it and wait for it to stabilize.
If you wanted an agile source in that era (60s tech), you might go for a dirtier oscillator followed by a filter. Switched-cap filters are pretty cheap and effective, and the clock frequency is readily available here so it's no big deal to make a variable filter in this way. (Whereas using OTAs for a variable continuous-time filter is a PITA. Also, "just make it good in the first place" would be the HP approach, endless optimization of a circuit to balance performance at all corners. That's the final-boss PITA that you're up against, here, I think.)
So, speaking of OTAs, I found this to be pretty effective:
OTAs are certainly contemporary, if not in IC form then they're easy enough to make discrete as well.
If you've taken higher level EE courses, you may recognize this as an implementation of the differential equation of an oscillator. IC1A and IC2A are integrators, R10 introduces a damping term, and IC1C provides a controlled negative damping term. The differential equation equates the second integral of the signal, to itself, with a first integral term having a coefficient of zero for no damping (stable amplitude), or positive or negative for decay or growth respectively.
The components are nonideal, so there will be some poorly-defined amount of damping in the system, varying with operating point; effectively, this makes damping biased towards decay by default, with a variable amount of growth on top of that. Exactly how much growth is needed to reach stable (constant amplitude) operation, that depends, but we can solve that with a feedback circuit.
My intent with this circuit was to get reasonably low distortion, and stable (or even controllable) amplitude, with frequency variable over a modest range. It seems the amplitude is quite controllable indeed, which is at odds with the theoretical basis -- the trick is the OTA's tanh(x) transfer curve, only linear for very small inputs (~10mV). As amplitude rises, average gain drops, gradually turning the integrator output towards a more triangular shape (it's nonlinear, it introduces distortion of course), but also limiting the amplitude. So it happens that I could measure the circuit as shown, with just a trimpot for amplitude control -- I didn't have to implement AGC. (I did add that later, using a simple precision rectifier and error amp; alas I don't have the schematic handy.)
The best part about this circuit is it produces quadrature (or, very nearly so, again given that the damping terms cause a little phase error). A downside is, because there's a fixed integrator in the loop, the amplitude of one channel varies against the other, while varying frequency. So it's not a constant amplitude quadrature oscillator. That could be fixed by changing IC2A to another OTA (hmm, and R10 might have to be variable as well?), or adding an amplitude regulator to the quadrature output.
Also, the noise performance probably isn't great (the impedances are quite high, and the voltage dividers can't be helping), but I'm not trying to do precision audio testing with this. YMMV.
So, this is commonly called a quadrature oscillator, or state variable oscillator (state variables meaning, the integrals of the signal).
This is rather beside the OP question (Wien bridge), but it sounds like you're more interested in any good oscillators, than in that type specifically, so that's okay.
And, as for that -- you can implement the gain control element in much the same way, i.e. using an OTA as a variable resistor or gain stage. Then you can control the dynamics however you like.
The problem with the lamp is, while its thermal time constant is largely diffusion based (which is good, that should leave about 45 degrees phase margin at any frequency in the asymptotic range), it's limited in size (the asymptote doesn't extend to infinite time, it has a longest time constant somewhere in there), so will naturally go unstable at low enough frequencies. To solve this, simply use a bigger bulb, of course! But then settling will be obnoxiously slow at high frequencies, taking thousands of cycles say.
There can be no simple solution to this -- you need the amplitude control function itself, to be controlled proportionally alongside the oscillator.
So we might solve that, by using both halves of a dual OTA, one for the variable resistor in the oscillator loop, one for a variable term in the control loop. A third (normal op-amp) does error amp duty. Varying the second OTA, in step with the oscillator setting, the control loop response can be made to track the oscillator frequency.
Also, you don't want the control to respond too quickly, because it will convert some of the detector's ripple into amplitude ripple -- distortion. In general, disturbance in the control loop is mixed into the oscillator response, so for example if the control has some 60Hz injected into it, you'll get 60Hz sidebands on your output. Same for any other noise, including input noise of the error amp itself, etc. This sets a limit on possible settling time versus distortion, and on the noise floor, phase noise, whatever.
A possible solution is an ideal detector, i.e., given a waveform A sin(w t), it returns the amplitude A regardless of phase t. The easiest way to implement this is using the Pythagorean identity, sin^2 t + cos^2 t = 1. For which you need an ideal quadrature circuit -- which is a point in favor of the quadrature oscillator. You then need a pair of squarers, then the sum is the amplitude --
with zero ripple, no filtering required -- which can then control the amplitude exactly over time. (Which, if this is starting to sound like more of a PITA, you're right. Now you understand why analog computers went out of favor...)
Otherwise, the problem with tracking the control response is, it has to track with the oscillator rate. Presumably you'd use a double ganged potentiometer to set the Wien bridge resistors; a third gang I guess could be added to get a control voltage, which can be offset as needed, to at least match the required response at two points. Hopefully a linear response matches nicely -- it's not obvious to me offhand if it would be linear, or follow a curve. If it actually needs a curved response (like, a hyperbolic section often shows up in situations like this, i.e., the gain needs to vary inversely with control level), then it can be tuned to match at two crossing points, and it'll settle worse away from those points. (Hopefully not so much worse that it oscillates.) Maybe it's a narrow enough range, and the tolerance is loose enough (say +/-20% of ideal settling), that it's fine. Anyway, the point I want to get at is, wouldn't it be great if the Wien bridge itself were variable, by that same control voltage, and proportional to it? So you could transform the resistors or capacitors with OTAs, or some other kind of analog multiplier.
At this point, it's just a little rearrangement, really, to make a state variable oscillator with OTAs -- so, one might argue it's still semantically correct, assuming one allows the above transformations. (Which is just a really technical way of saying, "I know you said Wien; well, teeeechnically...")
Anyway, despite the exponentially greater semantic complexity... this is really making a DDS look attractive, isn't it?
Tim