Calibration feedback 04 -- 2026-09-26, from evidence/shift-calibration-holdout-04 (4 oracle + 3
oracle_bad, one process of 7) and evidence/shift-calibration-dev-04.

shift_holdout_v1 (sha256 03b8c8a5...)
  oracle: passed every check (35/35) in 4 of 4 runs.
  oracle_bad did NOT fail: 5:reached, 10:prop_still (3 of 3 runs)

shift_dev_v1 (sha256 f9c9a8aa...)
  oracle: 35/35.
  oracle_bad did NOT fail: 0:reached, 6:reached, 11:prop_still

Why, and it is not the author's doing: the robot's bridge now plans drive_to routes around
keep-out zones and learned obstacles, REFUSES a drive_to whose target lies inside a zone, and
stops any drive the moment it starts pushing something. A wrong behaviour scripted with drive_to
is therefore corrected by the bridge before it can happen. A calibration_bad that must push the
pallet or enter a zone has to use raw motion the bridge does not plan: turn toward the target,
then set_velocity {"v": 0.5, "w": 0}, _wait the seconds needed, then stop_robot.
