Some of things you say doesn't really make sense to me (I'm not an expert though):
- Shouldn't accuracy determine bandwidth? Because what's the point of having an analog input bandwidth say 10 Ghz when your timebase has only 1ns resolution? With that step size you won't be able to reconstruct Ghz range signals (except in some very special and limited cases).
Usually you would have time resolution and accuracy commensurate with your sampling bandwidth however there are exceptions. To give an extreme example, a sampling voltmeter like a Racal-Dana 9301 or HP 3406 has no timebase or triggering at all and a bandwidth which only depends on construction and sampling gate pulse width. These instruments can make RMS voltage measurements beyond 1 GHz. If I connect one of my RMS voltmeters to the direct output of my sampling oscilloscope, I can do the same thing to 10+ GHz with no triggering.
Another place where timebase accuracy and resolution do not matter is X-Y displays using sampling inputs.
- In sequential ETS you can set the sampling delay to a large (predetermined) value and then measure its time interval. You can use this to calibrate the timebase. Because you measure long intervals and then divide it down to the step size your measurement can be more precise. I don't think you can do similar calibration in random ETS because you also need to measure the strobe to trigger edge delay as you've mentioned.
While this does not apply to my Tektronix 7T11A, another reason that sequential ETS can be more accurate is that assuming low jitter in the sampling strobe, multiple measurements of the delay can be taken and averaged. Every random equivalent time sampling measurement is unque and has to be made in a single shot. On the Tektronix 7T11A, all time measurements are single shot measurements so random ETS is just slightly less accurate than sequential ETS. Both get down to the 10 picosecond range.
- I agree that maximum frequency of the delay lines in those old instruments are a limiting factor. I'm not really sure why this was the case though. You can easily buy 18 Ghz semi-rigid coax delay lines nowadays. Insertion and returns loss of the coax might still be a concern.
Dispersion in a transmission line limits rise time. So for instance a Tektronix 113 delay line which uses 50 feet of 7/8" Spir-o-line (like smooth Heliax), provides a 60 nanosecond delay but limits rise time to about 100 picoseconds. It also weighs 50 pounds and is the size of a suitcase.
For an extreme example of this, consider old 100 MHz HP analog oscilloscopes where do to age, the shield of the delay line does not make good contact between strands limiting bandwidth to below even 100 MHz. Flexing the cable solves this.
- You should be able to measure a repetitive periodic signal's leading edge by phase locking a low phase noise oscillator to the signal and trigger from that. So delay lines are not always necessary.
That is how random equivalent time sampling works except that there is no requirement for low jitter in the source or oscillator. If a sequential sampling timebase can support enough delay, then it can trigger on one pulse and measure the next but this depends on low jitter in the source.