I saw something similar when looking for a good fit in my box joints. When test fitting with parts laying flat, it would bind in the bottom (near spoilboard) while the tops of the fingers still had clearance, so the fingers were evidently a bit trapezoidal, even with a finishing pass.
I am not sure of the true cause but my best guess is that the spindle axis of rotation is not exactly parallel to the shank of the tool. This makes the effective tool radius (about axis of rotation, not shank center) slightly off and dependent on “azimuth” relative to the collet.
With a single flute bit, the cutting edge is at a different azimuth at different z height, so depending on the orientation of the bit in the collet, you can get a swept area thats wider or narrower at the bottom.
It seemed like a long shot but I did two tests: one, I inserted a plain rod and tried to measure if the collet was coaxial to the rotation. It was off a little but hard to measure. Two, I marked the bit and the collet and tried different orientations. This did make a difference in the fit of the box joints, and I got the fit to be pretty similar across the 12 mm or so of z.
I also examined the bit itself but it looked perfect as near as I could tell, and I observed it on two different bits, so I doubt the bit is the problem. And anyway if the bit were tapered and the collet were coaxial to rotation, changing orientation of the bit in the collet wouldn’t have made a difference.
Anyway, try rotating the bit in the collet and see what happens.