> I believe PRR could nail it down.
Sore elbow.... maybe I can use the electric screwdriver?
> What is the purpose of this capacitor?
Reduce power buzz. You have rectifier, 20uFd, unknown choke, 20uFd, +300VDC. This point will be clean enough for plates but (maybe) not for screens. It is customary to take one more filter-stage between plate tap and screen tap.
> resulted in lower gain indeed.
Higher G2 voltage will require a higher G1 voltage swing. For -gain- you want G2 as low as will make your required Power. For Power you need G2 high enough to slam your load. Here they split the difference: good gain for effective NFB with ample power (+24dBm is less than 1/10th what 2x6V6 can make).
> resulted in lower gain indeed.
Then something is wrong. The R20 R6 network(s) constrain the gain of the 12AT7-6V6 stages with about 4:1 NFB margin. A 32V or 14% change of G2 voltage should not give audible gain change.
Since this stage is push-pull, the pentode-triode effects are nil, canceled from side to side. Removing C7C mostly raises the ripple-buzz. (It may or may not have good or bad effects on distortion, but this is very secondary.)
> make the 6V6 stage clip later.
This seems reckless. Two 6V6 will make Ten Watts, which can SMOKE any solid-state input behind it.
And I strongly suspect it is the 12AT7 clipping. The 6386 plates idle at 60V and go to 137V when limiting. The R35 C10 rise-time is faster than the C1 R14 coupling-network time constant. So 75V of that +77V jolt tries to pull 12AT7 grids positive. In fact the C1 R14 time constant now has the ~~1K forward grid resistance, oh and we should now account R11 and 6386 resistances.
Fixing the 12AT7 stage is trivial, but the fix will slam the 6V6es to "off", clipping there. While this too might be fixed (might be a full re-design), that just defers the problem to some later point down the chain.
It is NOT meant to be "slammed". It assumes a polite responsible professional operator. But accidents happen. And something always clips. This unit might feed a 25,000 Watt audio amp, the modulator for an AM radio broadcast transmitter. Clipping these things is illegal: it "splatters" hash into the air-space of other broadcasters. So illegal, that 25,000 Watt modulators are not built to stand heavy clipping. And when they fail it is exciting and costly.
With the Sta-Level, an "accident" gives <25mS of clipping in the 12AT7 (which is not harmed) and then it settles to non-clipping sorta-clean. As long as such accidents are rare, everything is fine. The 12AT7 clipping causes much less splatter than modulator-whacking, and the short duration of clipping/splatter may go un-noticed by other stations and audiences.
I'm not sure what you want a slow-attack limiter to do with a severe overdrive? It is designed to ramp-up over 20mS-30mS. Those transients are supposed to pass un-changed. Unless you work far below reference level, they MUST overload some later stage. If you need NO overload, you need to accept the over-control of 50uS (0.05mS) attack times and use a far beefier sidechain driver than a couple 0.1u caps and a hollow diode.