Driven-run start: commanderTo checks only the flat distance after .go xyz, never z (a wrong z puts the commander under the terrain and the bots follow into packs) #292

Found in the 2026-10-22 placement investigation (r10 Zul'Farrak, wiped party before any run started; evidence in ~/TortoiseBots-audit/tests/r10-19z29/slot2 and slot1, method file TARGETED-TEST-METHOD.md).

Root cause of the 'fake z': there is no teleport offset. .go xyz X Y Z uses the z you type and never snaps to the ground (core Commands.cpp HandleGoXYZ). At ZF 1895 1075 and 1915 1080 the ground is a raised shelf at about z 25 and 28, so a commander sent to z 9.5 stood 15-19 yd UNDER it; the agent only drops to a navmesh floor below it (internal/session/fall.go, actions.go) so with the ground above it (or no tile loaded) it keeps the typed z and sets z_unknown=true. .groupgo then placed the bots on the real ground (z 25-28), and they walked down toward the buried commander into a roaming patrol. .bot summon copies the commander's z exactly (PlayerConvenience.cpp:64-133), so it puts a bot under the ground or in the air whenever the commander's z is wrong (measured in the Barrens: 15 yd under -> bot stuck under the ground 21 s later).

Gap in the agent: internal/dungeon/start.go commanderTo (around lines 644-681) verifies only the flat distance after .go xyz, never z.

Fix: after the .go xyz, read .gps (server heights: GroundZ = top surface, FloorZ = floor under the commander's current z), re-send .go xyz X Y FloorZ+0.3, and require status z within 0.5 of FloorZ (and FloorZ equal to the leg's level when FloorZ and GroundZ differ); refuse the start with the numbers otherwise. Add a test with a fake .gps reply (z under the shelf) and a mutation proof.