Prashant task dep on v1

[prashantcraju]

Summary

• Re-add the task dependency simulator recipe onto msb_curr_min_v1 (cherry-picked from the previous branch).
• Align stats aggregation with the refactored metta/rl/stats.py (sum per-label counters).
• Add a module-level docstring documenting the simulator, Tool wrappers, and sweep utilities.
• Re-ran the simulator and a shortened z-score sweep (both locally and on a SkyPilot GPU sandbox) to confirm behavior.

Testing

Local smoke test

uv run ./tools/run.py recipes.experiment.curriculum_test.task_dependency_simulator.train
num_epochs=50 samples_per_epoch=5 run=local_smoke

Short z-score sweep (local)

uv run ./tools/run.py recipes.experiment.curriculum_test.task_dependency_simulator.sweep_zscore_amplification
num_epochs=100 samples_per_epoch=5 num_envs=32 num_sweep_points=5 min_zscore=1 max_zscore=30
run_prefix=refactor_zscore_short

SkyPilot GPU smoke + sweep (20 epochs)

uv run ./tools/run.py recipes.experiment.curriculum_test.task_dependency_simulator.train
num_epochs=20 samples_per_epoch=5 run=smoke_gpu

uv run ./tools/run.py recipes.experiment.curriculum_test.task_dependency_simulator.sweep_zscore_amplification
num_epochs=20 samples_per_epoch=5 num_envs=32 num_sweep_points=5 min_zscore=1 max_zscore=30
run_prefix=skypilot_zscore_short- WandB link: SkyPilot sweep

Notes

• Diff limited to recipes/experiment/curriculum_test/task_dependency_simulator.py (plus integration_checklist.md if included).
• Verified docstring + stats aggregation match the refactor.
• Ready for review once msb_curr_min_v1 CI is green.

Asana Task