I agree with you and slowing the attack is the only way that I have
been able to approach stable operation with a fast release. But 20ms or
longer attack runs counter to what I understand to be the objective -
an attack speed of less than 13 ms and ideally about 1 ms. So, unless I
have this wrong, how do other receivers accomplish similar speeds
without self-oscillation?

The way my circuit operates (I think) is as follows (I'd be happy to
send a schematic to anyone who is interested) :
a) assume an impulse of signal of duration very much longer than the
attack time
b) the rectified signal is filtered of RF by a series-parallel R-C
attack network whose adjustable output feeds an emitter follower
c) the emitter follower pumps current as a low resistance source into
the release R-C network so the attack is not greatly slowed - its
output feeds the AGC driver amp
d) at some point, equilibrium should be reached - the current flow
through the release resistor and AGC driver base should equal the flow
though the emitter follower - but maybe the emitter follower pinches
off and that could be a cause of instability?
e) the signal drops, the attack network discharges at attack speed and
shuts off the emitter follower, so the release capacitor discharges
through its parallel R at release speed, the voltage to the AGC driver
falls so the AGC bias rises at roughly release speed to increase RF/IF
gain.

Having written that out, I have an idea or two I will try.

Tom