The open tennis-ball feeder works out its own launcher with no radar gun. Each ball slows the two wheels a little, and the firmware fits each wheel's drop using the commands its speed loop sent, treating the wheel as a single lag of those commands. Earlier today another run found the plants that break that assumption. Tach jitter is the worst, and the board now flags it well. An ESC with a lag of its own is the awkward one. It left the calibration drill's spin reading 20-24% off at worst, and some of those drills flagged nothing at all.
Two lags look like one. If the ESC ramps its output, there are two lags between the command and the wheel. A single-lag fit absorbs the extra one into the wheel's lag and the moment the ball arrived. The fit still looks good, but the drop reads 1-5% low, and the spin estimate is built on the difference between the two wheels' drops, so it suffers most.
Give the model the second lag. The fix is simple once you have the lag: run the logged commands through it before fitting. With the true lag, the drill's worst spin error falls from 23.8% to 1.4% (30 ms ESC) and from 19.9% to 1.8% (80 ms).
But you have to know it, and guessing is dangerous. Telling a build with no ESC lag that it has 20 or 50 ms made the spin reading 41-44% off, roughly twice as bad as the problem being fixed. The one comfort is that the board's existing "unsure" flags fire in that case: about a third to half the wheels were marked.
Let the balls decide. A single ball can't separate the two lags reliably. Picking per ball fixed the bias but blurred the first-order builds. The lag belongs to the build, so the new tool sums each wheel's leftover error across a drill for six candidate lags and takes the smallest. On the simulator it chose 0 for every first-order build, so nothing changed there. It chose 20 ms for a 30 ms ESC and 100 ms for an 80 ms one. Through the whole drill that gives a worst spin error of 6.5% and 3.0%, against 23.8% and 19.9% before. The coarse grid is now the limit (20 ms against the true 30 ms), so a finer search near the winner is next.
What's missing. On a real feeder, the pick needs each ball's raw timing, and the board doesn't send that to the Pi yet, so the new ESC_TAU_US setting is manual for now. It is all simulation: nobody has logged a real ESC's step response to see which shape it has. The code, the tables and the sweeps are in the repository, under "An ESC lag picked from the build's own balls".