OUTPUT_SCALE < OUTPUT_MAX_LIMIT?

This does sound like a problem with the PM-voltage conversion stage.
You can probably linearise it with the "Lincurve" HAL component.

Any idea why the high range changes with PWM frequency?

Can you point me in the right direction with lincurve?

If I’m using a quadrature encoder for feedback, does it even matter to linearize the voltage output? Would the closed-loop setup adjust according to the RPM feedback? 

Thanks in advance!

I would guess that the filter is not using a push-pull opto and so the off-state decay time is slow. But that is just a guess.

If you are closing the spindle speed control loop. then no, the linearity isn't all that important.

Do you think the linearity would affect the PID as far as performance? Would fixing the linearity ultimately make PID tuning easier?

You should get faster settling to the commanded speed with a better linearity.

For whatever reason I thought you were using a Mesa card with an analog spindle output. I get those numbers confused.

If you're using one of those Chinese $8 BOBs with the DB25 then Andy is probably right about the Optocoupler output's rise/fall time but the in that case I'd lower the PWM frequency by alot. No idea what kind of Opto they're using on that thing but optos can be very slow. You should probably reduce the PWM frequency to 1khz or less, maybe even less than 100hz. Then try to get the output voltage correct by adjusting the hm2 PWM scale which has a range of -1.0 to 1.0, The scale is supposed to adjust the duty cycle to the input value. I would think that if the optocoupler was slow the duty cycle would have a better effect than the frequency.