fix: judge learner answers before tutor moves

[rosshd]

Closes #45

What changed

• Judge pending-check answers before generating the tutor response.
• Generate the next response from the learner's freshly updated state while preserving the prior question's focus and concept attribution.
• Skip grading for navigation and unrelated turns.
• Preserve the pending learner prompt and roll back uncommitted state if tutor generation fails.
• Bound the pre-stream judgment call to 512 output tokens, 20 seconds, and one attempt.
• Add focused CLI and tutor-behavior regression coverage.

Verification

• PYTHONPATH=src make check PYTHON=../../.venv/bin/python OPENLEARN=../../.venv/bin/openlearn
• PYTHONPATH=src make review PYTHON=../../.venv/bin/python OPENLEARN=../../.venv/bin/openlearn
• 429 unittest tests passed, 1 skipped.
• 533 pytest tests passed, 1 skipped, 11 deselected.
• Ruff and smoke checks passed.