First of all, those two values, $4243 and $75, only apply to NTSC with the display window set at standard position. For PAL, you'd use $5686 and $82, respectively.
It should be clear, that the number of cycles between VIA T1 interrupts must be the same as number of cycles per frame, so the interrupt is issued at exactly the same position for each frame. Actually, the timer underflows to 0 and -1 before it is reloaded from the latch value, which is the reason you have to load the number of cycles
minus 2 into the latch.
NTSC has 261 lines with 65 cycles each, PAL has 312 lines with 71 cycles each. And 261*65-2 = 16963 = $4243, 312*71-2 = 22150 = $5686. A constructive way to verify those values would be to use a slightly different one, then you'll see that the interrupt position begins to drift: upwards for smaller values (as it takes less time for subsequent interrupts) or downwards for larger values (as it then takes more time for the next interrupt to occur).
The values $75 (NTSC) and $82 (PAL) were more or less derived by experiment. They have been chosen so the interrupt happens in the first raster line below the display area, to maximize the time until the display area of the next frame begins. Other applications might require to sync the VIA interrupt to a raster line shortly before a certain effect is supposed to happen.
The exact begin of interrupt processing can be delayed by up to 7 cycles, depending on when the current instruction ends. Furthermore, the initial sync to the raster line can introduce another (permanent!) inaccuracy of another 7 cycles because of the CMP/BNE loop. And actually, the initial sync should first check for the raster value being *not* in the register, so it doesn't trigger already first time the loop has been entered and so puts the interrupt position somewhere in the (double) line up to its end (but not near the beginning anymore).
However, these considerations in the last paragraph are only important for cycle exact/stable interrupts. You'd need these only for much more demanding applications than syncing an animation to the frame. For these more involved applications, methods exist to sync the interrupt position with single cycle accuracy (i.e., better than the 7 cycle accuracy of the simple CMP/BNE loop) and, when the interrrupt has happened, introduce an own variable delay to compensate the interrupt delay, so the own interrupt routine begins processing at the exact same position (not only raster!) on screen each frame.
For one example, take a look at the
VFLI panning slide show. Here, the interrupt is synced to a position slightly above the display window. For another example, Victragic's version of
Pitfall! draws the background entirely by changing the background register as the raster beams races along the screen, in a *big* cycle exact interrupt routine.