Was this the watermarks regression? If so it should be gone on ToT. If this is only on kernel 3.18, it might be worth retesting.

I can figure this out this myself, but it would save me time if someone could point out how 'modetest' determines whether a display is connected and point out the (display driver?) code. Does it rely on the HPD input signal or some other method? 

Yes it uses the HPD pin. But for type-C the HPD signal is generated by the EC because there is some muxing that has to happen first.

A cursory check of our mux regs shows that we're consistently configured to USB + DP when the dock is attached, and our HPD output GPIO is getting asserted.

I updated to a fresh test image (8996.0.0) on my lars unit and things seem to be working consistently now. I'm not yet completely convinced that updating the image was the actual fix, there may be something else going on. I will try a few more tests.

I rolled back to some earlier R55 OS images and DP is still functioning, so I don't think updating my image caused my device to become fixed. Maybe an EC reset fixed things, or maybe the dock itself is in a different state. The HS mux part here is very simple and hard to misconfigure (unless we get polarity wrong, somehow) - there is a single register that configures USB3 / DP / polarity settings, and we're reading it back + decoding it correctly when we do "typec 0" on the EC console. HPD generation is done through a simple GPIO, and I verified it was in the proper state previously (and even tried toggling it by hand, to see if it would help).

I no longer have a unit that doesn't work with this dock, so please ping me if you have a (preferably lars, but I'll take anything) device that doesn't work.