controller dataset builder and scheduler interface

[pranay5255]

Summary

Build the controller dataset and the minimal adaptive scheduling interfaces needed for Exp 4. This issue owns turning forest traces into controller supervision and defining the scheduler contract, even if the full adaptive runtime lands incrementally.

Scope

• Derive controller state features from canonical trace rows.
• Define controller action labels and rationale fields.
• Build the dataset exporter for controller rows.
• Add a small training-ready feature format and baseline heuristic policy.
• Define the runtime interface needed later for STOP_AND_SUBMIT, SPAWN_MORE_WORKERS, DEEPEN_TOP_BRANCH, DIVERSIFY_PROMPT, RUN_VERIFIER, and SWITCH_TO_PATCH_MODE.
• Wire only the scheduler protocol / hooks now; full adaptive execution can follow once the dataset is stable.

Modules to build

• project/evmbench/evmbench/experiments/build_controller_dataset.py
• project/evmbench/evmbench/experiments/controller_features.py
• project/evmbench/evmbench/experiments/controller_labels.py
• project/evmbench/evmbench/experiments/controller_policy.py
• project/evmbench/evmbench/experiments/scheduler_protocol.py

Important context

• Current modal_forest.py is a fixed stage pipeline, not a stepwise adaptive scheduler.
• The controller work should not assume a live reranking loop already exists.

Acceptance criteria

• Can export controller rows from an extracted trace dataset.
• Feature vectors are reproducible from saved run artifacts alone.
• Scheduler protocol is explicit enough that evaluate_phase6.py / modal_forest.py can integrate it later without redefining the state/action contract.