if these were interpolated, there would be a loss of quality in between each keyframe.

yes, that is why I switched to exponential map: https://mathr.co.uk/blog/2020-06-18_optimizing_zoom_animations_again.html much more efficient, so the same effort gives better quality

using the 4 corners as probes but this overskips quite often.

KF uses the 4 corners plus the 4 midpoints of the view sides.  It compares series with regular (perturbed) iterations iirc, and has two thresholds (the stricter one is enabled with "approx low tolerance" checkbox)

I've read that using some NR iterations to place a probe point near a minibrot may work well

I think this might be what knighty's SuperMB does, not sure

Is the reference orbit the same after each of these near-zero approaches, and could this be used to speed up reference calculations?

there is a unique periodic point (nucleus) at the center of each minibrot cardioid (and disc, but maybe that's not so useful?) that returns exactly to 0 and repeats with a set period, which is a good reference to pick - you still need to update the series approximation coefficients (but if you use a series in 2 variables you only need to do it for 1 period - this is knighty's NanoMB technique).  kf doesn't use periodicity of references outside of nanomb code (for power 2 mandelbrot only).  mandelbrot-perturbator uses periodicity of references and properties of perturbation to not need to restart iterating pixels when they are glitched, too.

Also, is there any program currently that can color and assemble a zoom video using the OpenEXR output data file format?

https://mathr.co.uk/zoomasm can, some months after you asked (it uses EXR input in exponential map layout (e.g. output from KF), does not (yet) output EXR video frames but that could be interesting for HDR).