agent: put the action bounds in the system prompt and give a stall/at-bound flag in move results

[jeqcho]

Today the only place the model learns the Cartesian bounds is the move_to tool description ("Per-dimension bounds: left_x: [0.15, 0.48], ..."), and a clamped or unreachable target comes back as the same "executing move_to over N steps" as a normal move. Agents then infer contact or reach limits from stalls and get it wrong (2026-08-26, rig-2: 18 calls attributing a left_y bounds wall to arm reach; rig-3: a correct dead-band computation, but only after 17 calls).

Proposal:

1. Repeat the bounds (and which axes are pinned) in the system prompt's embodiment section, next to the frame conventions, so they are part of the model's initial world model rather than a tool footnote.
2. Have the move result carry a short status: clipped: left_y 0.30 -> 0.25, stalled after 12 steps (reached x=0.148), or ok, computed from the approver record and the readback vs. target, so the model does not have to diff eef_state itself.
3. Make the give_up hint say that operators can change limits between trials (the rig-3 operator offered exactly that mid-run and the model did not know it was an option).

Related: robocurve/inspect-robots-yam#135 and the yam issues on default bounds and bounds_clamped reporting filed today.