[pjerra]

Fixes #292.

A driven run's start moved the commander with .go xyz and then checked only the flat distance, so a wrong z (under a shelf or the terrain) went unnoticed and the bots followed the commander into the floor.

What changed:

• After .go xyz the start selects the commander by GUID, then sends .gps and reads the reply's FloorZ.
• A reply about another unit (X/Y more than the start range from the commander) is ignored.
• If the commander stands off the floor, the start re-sends .go xyz once at the floor height; if the floor is far from the target (another layer) the start is refused with a plain reason.
• No reply within the timeout: the start goes on and notes once per run that the height could not be checked.

Tests: go test ./... passes. New tests cover under-the-ground, the single re-send, a floor far from the target, a reply about another unit, the select-before-.gps order, and booking each GPS line. Mutations of each check turn a test red.

Reviews: Opus cold review SHIP, Codex review found no actionable issue.

Known limits (minor, follow-up):

• The off-floor test compares FloorZ with the agent's own view of the commander's z; the server's Z: from the same reply would be more robust.
• FloorZ of -200000 (no height data) is treated as "cannot check".
• README still says the two GM lines are the only GM commands of a start; .gps and the re-send are now booked too.