Sadly all that made no difference. I corrected the LF response even further by adding 1uF below the 8k shunt feedback capacitor, to reduce the gain to unity at DC and from about 10Hz down, which is how I think it should really be, but that didn't help either.
So I am presently considering the good old silicon rings around the 6SN7s. Curiously enough there is a brand of these called Cary ...
I also rolled off the HF response by adding 22pf (or possibly 220pF pF, can't remember) in shunt at the input, 10K grid stoppers between the volume pots and the V1 grids (instead of the wires, easy enough), and 15pF across the 100K feedback resistor, which should all help with the AM reception while still keeping it reasonably wideband at 100KHz (was nearly 1MHz). AM break-in only happens at the customer site so I can't verify that at present.
To document the gain arithmetic above: the gain of the grounded-grid stage is given by 100/4.7 ~= 21, divided by 2 because of sharing the cathode resistor with V1A, which halves the current and therefore the plate voltage swing, so about 10.5, or ~20.5dB. This is A in the general formula for gain: A/(1+beta*A), which is used when where A is not large enough to ignore, so taking beta as 8/100 this yields closed-loop gain of ~5.7 or 15.18dB. This is still a little high for the customer's equipment, so I actually increased the 8K resistor to 10K, which yields gain of 5.14 or 14.24dB. All this needs to be corrected further for the two cathode-follower stages, one inside the feedback loop: I will update this post for that.