SWE-bench Pro network-policy derivative

This repository is materialized from the PR-1593-hardened local SWE-bench Pro tree in LLM360/agentic-evals. Harbor PR 1593 introduced that source materialization.

Each source task.toml receives exactly three policy changes:

1. schema_version = "1.0" becomes schema_version = "1.4".
2. [agent] receives network_mode = "allowlist" and allowed_hosts = [].
3. [environment] replaces allow_internet = true with network_mode = "public".

All other canonical bytes and executable modes are preserved, except directories named __pycache__ and files ending .pyc, which are excluded. Materialize and verify with:

python scripts/materialize.py "$PRO_SOURCE" tasks
python scripts/materialize.py --verify "$PRO_SOURCE" tasks