The thing nobody tells you about bed levelling is that the bed is never actually flat. You spend an evening with a sheet of paper getting all four corners to that perfect slight drag, you print one thing, it's beautiful, and three prints later the first layer is squashed on one side and not touching on the other. The bed didn't move. The bed was never the right shape to begin with; it's a slightly warped piece of metal that expands when it heats, and four corner screws can only ever approximate a plane it doesn't have.
I spent far too long treating this as a discipline problem, as if I just wasn't levelling carefully enough. I was levelling fine. The geometry was against me.
What ended the war was admitting the bed has a topology and measuring it instead of pretending it's a plane. A probe (I fitted a BLTouch) takes a grid of readings across the bed before each print and builds a mesh, then the firmware lifts and drops the nozzle a fraction as it moves to follow the actual surface. The middle being 0.1mm high stops mattering because the nozzle just rises over it.
G29 ; probe the mesh
M420 S1 ; enable mesh compensation
Add that to the start G-code and the printer measures its own warped bed every time, then prints around it. I still wing-nut the corners roughly close once in a blue moon so the mesh has a sane starting point, but the nightly paper ritual is gone, and the first layer is right whether or not the bed feels like cooperating. It was never a war I could win by hand. I just had to stop fighting it on its own terms.