Discretization

It throws information away on purpose, so it has to earn its place. A gradient-boosted tree already searches for its own thresholds, per branch, against your actual loss, so binning a column before one can only lose resolution.

Where it genuinely pays: giving a linear model a step it could never fit, handling a relationship that reverses direction, matching a rule already written in bands, and keeping extreme values without deleting the rows.

Equal-width bins have equal widths and unequal counts, so on skewed data most rows land in one or two of them. Equal-frequency bins hold the same count and simply get wider in the tail, which is usually what you want. And supervised binning chooses edges using the target, which makes the edges fitted state that leaks unless computed inside each fold.