The open tennis-ball feeder has a known fault on one simulated build. With the first-order wheel model, a 0.8 s wheel lag and a 0.95 / 0.35 launcher, the slow-wheel drill always stopped at its third shot with "release not done in 2.0 s". Every other build finished. This run traced it frame by frame.
What the trace showed. Shots 1 and 2 both plan the wheels at 6,065 and 526 rpm. After shot 1 the slow wheel braked to 507 rpm and then crept back at about 10 rpm a second. That is 3.6% under target. The board feeds a ball only once both wheels have held within 3% for six revolutions, which is 684 ms at 526 rpm. At 25.2 s one reading of 511 rpm completed that hold. The board's "settling" flag cleared and the Pi sent FEED. The next reading was 509, so the gate reset and the ball stayed queued.
The wheel was back in the band at 25.9 s and held it. The stroke started at 26.6 s and takes 0.9 s, so the ball would have gone at about 27.5 s. The Pi gave up at 27.2 s, exactly 2.0 s after the FEED, and latched the e-stop.
Who was wrong. The firmware wasn't. It waits because a ball fed on a slipping wheel lands short. The 2 s timeout was sized for the stroke plus a margin. Nobody had sized it for a gate that re-checks a slow wheel after the Pi has already fed, and at 526 rpm that re-check alone takes 0.68 s.
The fix. The Pi now starts its 2 s count from the last telemetry frame that shows the gate shut, so it waits as long as the board is still checking. A shut gate can stretch the wait by 10 s at most, so a wheel that never settles still stops the drill. A lost FEED frame on an open gate still aborts after 2 s, as before. Two new tests hold the gate shut for 1.5 s and for 30 s. On the old code the 1.5 s test aborts. On the new code it finishes, and the 30 s test aborts at the cap.
On the simulator, the build that always aborted now runs all 30 shots. Shot 2 goes 2.3 s after its FEED.
What is left. The undershoot is still there. A 0.8 s wheel braking to 526 rpm under the speed-scaled gains settles 3 to 4% low and takes about 2 s to recover. The gate waits it out, so the cost is time, not accuracy. It is the next item on the list. All of this runs on the simulator's first-order wheel, and a real wheel may undershoot differently.
The trace, the script and the change are in ventures/open-tennis-robots (firmware/esp32/README.md, "The slow drill's release timeout").