code/test_serve.py

"""
Regression test for prompt parsing.

Worth having as a standing test rather than a one-off check, because a wrong
parse is invisible: it still yields a syntactically valid 54-character state, so
the model is simply asked to solve a different cube and the damage shows up only
as an unexplained low benchmark score. The fixture is generated by the real
renderer (scripts/print-prompt-pairs.ts), so it also fails if the prompt layout
ever changes.

    python -m pytest test_serve.py      # or: python test_serve.py
"""
import json
from pathlib import Path

from serve import extract_state

FIXTURE = Path(__file__).parent / "fixtures" / "prompt_pairs.json"


def load():
    return json.loads(FIXTURE.read_text())


def test_extracts_state_from_real_prompts():
    for case in load():
        got = extract_state(case["prompt"])
        assert got == case["facelets"], (
            f"depth {case['depth']} index {case['index']}:\n"
            f"  want {case['facelets']}\n  got  {got}")


def test_rejects_prose_without_a_net():
    """The prompt's prose alone contains plenty of face letters -- naming the
    colours up front and the move vocabulary at the end. Parsing must fail loudly
    on text with no diagram rather than assembling a state out of that prose."""
    prose = "Use U, D, L, R, F, or B. U=White (Up), R=Red (Right), F=Green (Front)."
    try:
        extract_state(prose)
    except ValueError:
        return
    raise AssertionError("expected ValueError on prompt with no net diagram")


if __name__ == "__main__":
    test_extracts_state_from_real_prompts()
    test_rejects_prose_without_a_net()
    print(f"ok: parsed all {len(load())} fixture prompts; rejected prose-only input")