03 Code Fixes

Date: 2026-06-19 UTC

Installation Fixes

Moved pyarrow from required dependencies to optional parquet extra in pyproject.toml.
Reason: this cluster resolves pyarrow to a dummy wheel that requires an external Arrow module and breaks core pip install -e ..
Core JSONL dataset path remains fully supported. Parquet remains available with pip install -e ".[parquet]" in an environment with Arrow.

New Reproducible Shell Commands

Added:

scripts/smoke_test.sh
scripts/run_train_debug.sh
scripts/run_inference.sh
scripts/run_eval.sh

These scripts run tiny toy-backend jobs with OPENCLAUDE_MOCK=1 by default.

New Inference Command

Added:

scripts/infer_toy_policy.py

It loads a CIL dataset and checkpoint, runs model policy inference when torch/model weights are available, and otherwise emits a clearly labeled fallback action selection. Model-policy toy outputs are bound to actual group targets for readable/actionable JSON.

Evaluation Fixes

Fixed CausalStress policy rollout crashes caused by decoded actions containing predicted_target.
Added task-aware binding before toy simulator execution.

Runtime Stability Fixes

Added DOVLA_TORCH_THREADS handling in trainer/eval with default 1 for stable CPU smoke runs on shared clusters.
Full test suite dropped from a long/hanging run to 126 passed, 1 skipped in 16.21s.

README Updates

Added the four paper-audit shell commands to README quickstart.
Documented that no-torch fallback paths are explicit and should not be interpreted as learned model results.

Verification

outputs/audit_venv/bin/python -m pip install -e . succeeded.
DOVLA_TORCH_THREADS=1 outputs/audit_venv/bin/python -m pytest -q passed: 126 passed, 1 skipped.
Smoke, debug train, inference, eval, and expert-only baseline scripts all ran successfully.