code/tests/test_acl_main_figures.py

from pathlib import Path
import subprocess
import sys
from xml.etree import ElementTree

from scripts import build_acl_main_figures as figures


main = figures.main


REPO = Path(__file__).resolve().parents[1]


def test_main_renderer_writes_exactly_three_paper_figures(tmp_path):
    main(["--repo", str(REPO), "--output-dir", str(tmp_path)])

    assert {path.name for path in tmp_path.iterdir()} == {
        "fig1_method.svg",
        "fig2_accuracy_gain.pdf",
        "fig2_accuracy_gain.png",
        "fig3_threshold_sensitivity.pdf",
        "fig3_threshold_sensitivity.png",
        "captions.tex",
    }

    svg = ElementTree.parse(tmp_path / "fig1_method.svg").getroot()
    assert svg.tag.endswith("svg")
    svg_text = "".join(svg.itertext())
    for label in (
        "Same step-0 function",
        "PEFT tangent space",
        "Fisher-informed tangent space",
        "Ineffective region",
        "0.04 → 0.92",
        "28 → 7",
        "26×",
    ):
        assert label in svg_text

    captions = (tmp_path / "captions.tex").read_text()
    assert captions.count(r"\begin{figure*}") == 1
    assert captions.count(r"\end{figure*}") == 1
    assert captions.count(r"\begin{figure}") == 2
    assert captions.count(r"\end{figure}") == 2
    assert r"\includesvg[width=\textwidth]{paper_figures/main/fig1_method}" in captions
    assert r"\includegraphics[width=\columnwidth]{paper_figures/main/fig2_accuracy_gain.pdf}" in captions
    assert r"\includegraphics[width=\columnwidth]{paper_figures/main/fig3_threshold_sensitivity.pdf}" in captions
    assert "fig3_boundaries" not in captions


def test_main_renderer_runs_as_a_script(tmp_path):
    result = subprocess.run(
        [
            sys.executable,
            str(REPO / "scripts/build_acl_main_figures.py"),
            "--repo",
            str(REPO),
            "--output-dir",
            str(tmp_path),
        ],
        cwd=REPO,
        capture_output=True,
        text=True,
    )

    assert result.returncode == 0, result.stderr
    assert "wrote 3 main figures" in result.stdout


def test_accuracy_gain_plot_overlays_absolute_methods_and_marks_delta(tmp_path, monkeypatch):
    captured = {}

    def capture(fig, _output_dir, _stem):
        captured["figure"] = fig
        return tmp_path / "fig2_accuracy_gain.pdf"

    monkeypatch.setattr(figures, "_save", capture)
    data = figures.load_data(REPO)
    figures.validate(data)
    figures.plot_accuracy_gain(data, tmp_path)

    ax = captured["figure"].axes[0]
    patches = {
        gid: [patch for patch in ax.patches if patch.get_gid() == gid]
        for gid in ("peft", "fisher", "gain")
    }
    assert {gid: len(items) for gid, items in patches.items()} == {
        "peft": 5,
        "fisher": 5,
        "gain": 5,
    }
    assert {bar.get_width() for bar in patches["peft"]} == {
        bar.get_width() for bar in patches["fisher"]
    }
    assert round(patches["peft"][0].get_y() + patches["peft"][0].get_height(), 2) == 80.39
    assert round(patches["fisher"][0].get_y() + patches["fisher"][0].get_height(), 2) == 91.22
    assert round(patches["gain"][0].get_y(), 2) == 80.39
    assert round(patches["gain"][0].get_height(), 2) == 10.83
    assert sum(text.arrow_patch is not None for text in ax.texts if hasattr(text, "arrow_patch")) == 5
    texts = {text.get_text() for text in ax.texts}
    assert "+10.83 pp" in texts


def test_threshold_curve_reports_low_accuracy_rate_across_cutoffs(tmp_path, monkeypatch):
    captured = {}

    def capture(fig, _output_dir, _stem):
        captured["figure"] = fig
        return tmp_path / "fig3_threshold_sensitivity.pdf"

    monkeypatch.setattr(figures, "_save", capture)
    data = figures.load_data(REPO)
    figures.validate(data)
    figures.plot_threshold_sensitivity(data, tmp_path)

    ax = captured["figure"].axes[0]
    lines = {line.get_gid(): line for line in ax.lines if line.get_gid()}
    peft = lines["peft-threshold"]
    fisher = lines["fisher-threshold"]
    peft_at_70 = peft.get_ydata()[list(peft.get_xdata()).index(70)]
    fisher_at_70 = fisher.get_ydata()[list(fisher.get_xdata()).index(70)]

    assert round(peft_at_70, 2) == 37.33
    assert round(fisher_at_70, 2) == 9.33
    texts = {text.get_text() for text in ax.texts}
    assert "28/75 (37%)" in texts
    assert "7/75 (9%)" in texts
    assert {text.get_fontsize() for text in ax.get_legend().get_texts()} == {6}