I should add this: the above experiments were using the codes from the database. I tried again with the codes learned on the TSU1000, and the effect was the similar, although those codes more reliably jump 2 db at a time rather than 2 or 3 with the database codes.

I tried again learning the code into the 9600. That produces a very long (20 lines or more in the editor) code which jumps 1 db at a time, but won't keep going with the button held down.