Exact Store replay

MuscleAction.actuator_setpoints is the historical proposed action, captured before actuator post-filters. It remains available for compatibility. MuscleAction.applied_actuator_setpoints is the final post-filter action and therefore the action returned to the SimulationController for environment execution and used by exact Store-backed replay when present.

An ordinary replay definition remains backwards compatible. Exact sequential pretraining opts in to immutable source selection and validation:

replay:
  - agent: teacher
    experiment_run: teacher-run
    instance: immutable-instance-uid
    phase: 1
    mode: train
    strict: true
    repeat: 1

Strict replay requires selected rows to contain an applied action, worker and monotonic decision identity, reward, and explicit terminated/truncated flags consistent with done. It rejects duplicate/non-monotonic decisions, missing fields, empty selections, and more than one worker in a source. Rows are ordered by worker, episode, decision index, then Store ID; each episode is placed in its own Memory shard so successors never cross an episode boundary. All sequential Store replay rejects a source containing more than one worker: interleaving independent worker trajectories would otherwise fabricate adjacency. Strict mode additionally requires exact fields and raises setup errors; non-strict mode retains its best-effort legacy proposed-action fallback.

The columns are nullable by design. Existing databases are upgraded through Alembic revision d21a0b4c6e7; legacy replay continues to use the historical proposed-action field on a best-effort basis.