How an experiment puts the agent somewhere sensible before the first step.

Mirrors tbp/monty/frameworks/environments/positioning_procedures.py::PositioningProcedure, a Protocol in Monty and a behaviour here, which is how this port spells every abstract class (docs/monty-differences.md §2.6).

A procedure is asked for one step at a time and answers with actions the environment should take, plus whether it is finished. It is not a motor policy: it runs before the motor system does, against the environment directly, because a run has to see something before a policy has anything to decide.

Monty's protocol also declares depth_at_center as a static method; that is a function of PositioningProcedures, not of an implementation, so it is not a callback here.

State

Monty's procedures mutate themselves — the number of turns attempted, whether the extra orientation step for multi-object scenes has been done. Here a step answers {result, procedure}, and the caller keeps the procedure it was given back, as everything else in this port threads state.