This rules out all but the tiniest diodes (single dice in a hybrid assembly?), and even rules out a lot of transmission lines, since the width and height of the TL are similarly limited, lest TE/TM modes take over and introduce weird dispersion. Needless to say, as the cross section of the TL goes down, the losses go way up, and so the length must be that much smaller to maintain sharpness.
This is what limited the performance of Tektronix's
early sampling heads. The size of the GR-874 connectors allowed other propagation modes but nobody had an instrument fast enough to measure or confirm it initially.
This requires monolithic construction. There is actually nothing you can do with conventional materials (PCBs and SMT components) or a normally-equipped machine shop that will succeed here.
On a larger scale it requires transmission line construction for the entire switching assembly which suggests placing a mercury wetted reed relay inside a coaxial transmission line or coplanar waveguide such that the impedance is maintained through the entire structure. But the reed relay's glass envelope alters the dielectric constant so that much be taken into account and as you point out, the physical size will allow other propagation modes limiting performance above a certain frequency. I assume that is what limits performance of the
Tektronix 284 pulse generator to 70 picoseconds even with the tunnel diode embedded in the coaxial transmission line.
The S-2 sampling head and Tektronix 284 with GR-874 sized transmission lines are both limited to 75 or 70 picoseconds respectively. I suspect changing the connector on the S-2 would improve its performance to at least 8GHz.
Like you say, that rules out such a fast pulse generator using commonly available construction techniques unless you can build it into the rigid cable used with 3.5mm or smaller RF connectors. The NLTL assemblies I have seen for sharpening an existing pulse used MMICs and hybrid construction.
This is a best effort project. If I can compress the 100 pS pulse from Leo's unit to even 10-20 pS it will be better than what I have now. <3 pS is a goal, not an expectation.
I'd be much more interested in suggestions for diodes. For example, might a microwave transistor have lower junction capacitance than readily available diodes?
Look for a suitable varactor or step-recovery diode which are essentially the same thing intended and specified for different applications. I have been told than some bipolar RF transistor junctions are also very good for this but I think you are going to have to qualify any part you want to use.