Fix hidden PDDL grader dependencies

[parkavenue9639]

Summary

• Declare T136's unified-planning and up-pyperplan dependencies in the top-level grader requirements.
• Pin planner versions so Docker-produced plan pickles and host-side grading use compatible implementations.
• Check installed dependencies before attempting the T136 compatibility install.
• Add regression coverage for embedded grader dependencies and a real pyperplan solve using the bundled T136 fixture.

Problem

T136 is marked as a closed task with external_dependency: none, but its embedded grader previously depended on a runtime PyPI installation:

• unified-planning
• up-pyperplan

These packages were missing from the top-level requirements.txt. If the dynamic installation failed because of network, permissions, or environment restrictions, the checker returned all-zero automated scores. This conflated grader setup failures with agent performance.

The supplied solve.py also uses these packages inside the agent runtime. All PawBench Dockerfiles install the top-level requirements, so declaring them there keeps the runtime and host grader environments aligned.

Changes

• Add:
  • unified-planning==1.3.0
  • up-pyperplan==1.1.0
• Change T136 grading to import-check dependencies first and only use dynamic installation as a compatibility fallback.
• Add a test that:
  • Ensures packages explicitly installed by embedded graders are declared in requirements.txt.
  • Verifies pyperplan can solve the bundled task01.pddl fixture.

Validation

ruff check tests/test_grader_dependencies.py
PYTHONPATH=. python -m pytest tests -q