Upon further testing, it appears the channel swizzling was backwards during CSP lut importing.
I used Nuke as the baseline in terms of lut channel order, so it's possible that I've actually broken things if Nuke's CSP Vector node is broken. But if Nuke's works, OCIO's now agrees with the results. Can anyone confirm Nuke's Vectorfield implementation of .csp files works?
I am including the file I used for testing. This gains down the red channel by 0.5 - green and blue are unchanged.