Segment a model with memory at rest points, rebuilding its memory

[yichao-liang]

Summary

• A model with hidden state (MODEL_STATE_INIT) skipped settled-tail truncation and rest-point segmentation ("a rest point does not reset a hidden process"), so every candidate parameter point replayed each whole recording from its start.
  Its recordings are now truncated and segmented like any other model's.
• A segment that starts mid-recording carries the actions and observations before it in place of memory (model_state.memory_from_history).
  Restoring that state (restored_model_state, reached from _set_state after the replay pins its parameters) runs the model's own update_model_state over them under the replay's parameters.
  Each candidate point gets the memory it implies, computed as the execution-time tracker computes it from the same observations, with no physics.
• The physical state is re-anchored on the observation at each segment start, as for any segment.

Why

On the from-assets round fixes_r4, fits of hidden-state models took 15 to 56 minutes (Balloons seed 0, Boil seed 0, Bridge seed 0), and the fit gate makes the agent wait for each one after every model edit.
A recording too chaotic as a whole also could not be trimmed phase by phase: Bridge seed 0's whole 1690-step recording scored RMS 0.27 against the 0.1 trimming bar, so its first fits refused to run at all.

Scope

This changes the fit of every hidden-state model, in every arm that fits one (EMPIRIC included), not only EMPIRIC from assets.
How much it saves depends on the scored features: a model that scores the robot, which moves almost every step, finds few rest points.
On the fixes_r4 level-1 recordings:

recording	before	after
Balloons seed 0 (box, balloons)	114 steps	1 segment, 53 steps
Boil seed 0 (robot, jug, faucet)	352 steps	2 segments, 317 + 42
Bridge seed 0 (block z, robot xyz)	1690 steps	1690 steps, no rest point
Bridge seed 0, block z only	1690 steps	9 segments, 917 steps

Test plan

• test_subclass_model_state.py: a state carrying a recorded history restores the memory that history implies under the env's current parameters, through _set_state and simulate; segments of a two-phase recording carry exactly the history before them, leave the recorded states untouched, and rebuild memory per candidate rate.
• End to end on Balloons seed 0's fixes_r4 level 1 (scripted harness, the run's own simulator.py, both fits on one node): fit 450 s before, 221 s after; SSE 8.28 to 1.42 before, 10.66 to 1.25 on 2 segments after; the beliefs agree (for example lift_H 68% [1.05, 1.10] and burst_z [1.11, 1.18] in both).
• Local CI replay: yapf, isort 5.10.1, docformatter, mypy and pylint pass on this commit; the 8 pytest shards pass on its first version, which differs only in one test's import line.

🤖 Generated with Claude Code