Exactly. Roll mode has no trigger...
Interestingly, triggered roll mode (as a additional option) is coming into fashion and also Siglent have introduced it at one point. The important difference to the normal triggered acquisition is that it only works at slow timebases (which applies to roll mode in general), where normal mode leaves you with just an activity bar and the message "slow acquisition" until you finally see a graph (which might take several minutes, depending on timebase settings). By contrast, triggered roll mode gives you the joy of watching the trace developing as soon as the trigger fires, but then it also stops once a full screen width has been recorded, just like a standard acquisition would. This works in normal as well as single trigger mode.
It might sound like a good idea, because watching the trace develop is less boring than staring at an activity bar, but on the other hand this variant still has all the usual drawbacks of roll mode, i.e. lower samplerate and maybe also limited memory - which is the compromise required for continuous recording without blind time
and continuous updating of the screen in parallel.
And roll should restart if I change timebase or input att. What is use for a graph with random axes ?
This. Roll mode originally was meant as a substitute for the classic stripchart recorder - probably you were able to change the transport speed during operation back then - even though I very much doubt anyone serious about their work would have actuall done that.
Anyway, Roll mode on a scope as a substitute for the chart recorder is just a crouch at best, because max. recording time is limited and you get no long physical chart for documentation purposes. Unless your scope happens to have a thermal printer built in, you need to dump the memory and post-process the data after the recording has finished - and for that you could just as well (or even better) use standard acquisition with long memory. Apart from that, I don't think a standard DSO with its limited accuracy and resolution (even if its 12 bit) could be a substitute for the real strip chart recorder applications.
I think there is a reason why logging multimeters as well as dedicated data loggers exist.
So the real use for roll mode on a scope is just the permanent visual observation of a slow signal and for this it is certainly not important what direction the trace initially runs until it starts rolling or that it clears the screen when the recording speed (timebase) is altered.
I can think of many reasons why it would not be a good idea to try to continue roll mode on a different speed. If it did not restart, we get a wrong timescale for the existing data on the screen. If we try to rescale and redraw them, this would not only create a time jump, because the additional data manipulation and transfer would interrupt the acquisition.
The rescaling would also be lossy if the speed is increased, because this usually means a higher sample rate (as the memory remains constant at it maximum supported length) and samples that have not been taken initially cannot magically emerge all of a sudden.