Pdf To Chinese Latex

name: CI

# Runs on push to main and on every PR. Catches regressions in the
# pipeline scripts BEFORE they land — anyone proposing a change to
# render_figures / extract_bibliography / check_tex / extract_tables /
# extract_text / compile_pdf will see whether the canonical tests
# still pass.

on:
  push:
    branches: [main]
  pull_request:
    branches: [main]

jobs:
  self-check:
    name: self_check + stress fixtures
    runs-on: ubuntu-latest
    strategy:
      matrix:
        python-version: ["3.9", "3.11"]
    steps:
      - uses: actions/checkout@v4

      - name: Set up Python ${{ matrix.python-version }}
        uses: actions/setup-python@v5
        with:
          python-version: ${{ matrix.python-version }}

      - name: Install Python dependencies
        run: |
          python -m pip install --upgrade pip
          pip install pdfplumber pypdf PyMuPDF Pillow

      - name: Install XeLaTeX (TeX Live minimal)
        run: |
          sudo apt-get update
          sudo apt-get install -y --no-install-recommends \
              texlive-xetex \
              texlive-lang-chinese \
              texlive-fonts-recommended \
              texlive-latex-recommended \
              texlive-latex-extra \
              fonts-noto-cjk

      - name: self_check.py (end-to-end pipeline + xelatex compile)
        run: python scripts/self_check.py

      - name: Synthetic stress fixtures (regenerate)
        run: python evals/_gen_stress_pdfs.py

      - name: render_figures on stress-2col (must extract right-col fig)
        run: |
          python scripts/render_figures.py \
              evals/test_inputs/stress-2col.pdf \
              --out /tmp/stress-2col-out --auto
          test -f /tmp/stress-2col-out/fig1.png

      - name: render_figures on stress-subfigs (must produce one fig1.png)
        run: |
          python scripts/render_figures.py \
              evals/test_inputs/stress-subfigs.pdf \
              --out /tmp/stress-subfigs-out --auto
          test -f /tmp/stress-subfigs-out/fig1.png
          # there must NOT be a fig2.png — subfigures must stay merged by default
          ! test -f /tmp/stress-subfigs-out/fig2.png

      - name: extract_bibliography on stress-gbt7714 (5 entries)
        run: |
          python scripts/extract_text.py \
              evals/test_inputs/stress-gbt7714.pdf --out /tmp/gbt7714-raw.txt
          python scripts/extract_bibliography.py \
              /tmp/gbt7714-raw.txt --out /tmp/gbt7714-bib.tex
          # Sanity: file must contain 5 \bibitem entries
          test $(grep -c '\\bibitem{' /tmp/gbt7714-bib.tex) -eq 5

      - name: render_figures on stress-caption-sublabels (one fig1.png)
        run: |
          python scripts/render_figures.py \
              evals/test_inputs/stress-caption-sublabels.pdf \
              --out /tmp/stress-cap-out --auto
          test -f /tmp/stress-cap-out/fig1.png

      - name: extract_text loud-failure on a 0-text 6-page PDF
        run: |
          python -c "
          import fitz
          d = fitz.open()
          for _ in range(6): d.new_page(width=595, height=842)
          d.save('/tmp/blank.pdf'); d.close()
          "
          # MUST exit 2 with loud failure
          ! python scripts/extract_text.py /tmp/blank.pdf --out /tmp/blank.txt
          # check the exit was specifically 2 (not just any nonzero)
          python scripts/extract_text.py /tmp/blank.pdf --out /tmp/blank.txt; rc=$?; \
              test $rc -eq 2

      - name: extract_tables on stress-longtable (≥1 table)
        run: |
          python scripts/extract_tables.py \
              evals/test_inputs/stress-longtable.pdf \
              --out /tmp/longtable.tex
          # at least one \begin{table} should have been emitted
          test $(grep -c '\\begin{table}' /tmp/longtable.tex) -ge 1

      - name: check_tex follows \input{} (no false MISSING bibitems)
        run: |
          mkdir -p /tmp/split
          cat > /tmp/split/main.tex <<'EOF'
          \documentclass{article}\begin{document}
          See \cite{key1}. \input{bib.tex}
          \end{document}
          EOF
          cat > /tmp/split/bib.tex <<'EOF'
          \begin{thebibliography}{99}
          \bibitem{key1} A paper. 2024.
          \end{thebibliography}
          EOF
          python scripts/check_tex.py /tmp/split/main.tex

# Python
__pycache__/
*.py[cod]
*$py.class
*.so
.Python
.venv/
venv/
env/
.pytest_cache/

# Editor / OS
.vscode/
.idea/
.DS_Store
Thumbs.db
*.swp
*.swo

# LaTeX build artifacts (in case anyone compiles locally inside the repo)
*.aux
*.log
*.out
*.toc
*.synctex.gz
*.fls
*.fdb_latexmk

# User output (translations go in user-chosen folders, not the skill repo)
output/
*_中文版/
*_中文版.zip

# -*- coding: utf-8 -*-
"""Generate synthetic layout stress-test PDFs for evals/test_inputs/.

These PDFs are NOT realistic papers — they are minimal fixtures
designed to stress-test specific layout failure modes that the
heuristics in `render_figures.py` and `extract_bibliography.py`
need to handle. Real-publisher coverage (actual Elsevier / IEEE
quirks) still requires fetching real papers separately.

Usage:
    python evals/_gen_stress_pdfs.py

Generated:
    evals/test_inputs/stress-2col.pdf       — two-column body, fig in right column
    evals/test_inputs/stress-subfigs.pdf    — three subfigures under one caption
    evals/test_inputs/stress-gbt7714.pdf    — Chinese title + 参考文献 in GB/T 7714
"""
import os
import sys
from pathlib import Path

import fitz  # PyMuPDF

OUT = Path(__file__).resolve().parent / "test_inputs"
OUT.mkdir(parents=True, exist_ok=True)


# Resolve a CJK-capable font. Windows ships SimSun / Microsoft YaHei in
# Fonts/. On macOS / Linux we fall back to Noto Sans CJK. fitz needs the
# file path; insert_text/insert_textbox then uses fontname="F0" etc.
def cjk_font_path():
    candidates = [
        r"C:\Windows\Fonts\simsun.ttc",
        r"C:\Windows\Fonts\msyh.ttc",
        "/System/Library/Fonts/PingFang.ttc",
        "/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc",
    ]
    for p in candidates:
        if os.path.isfile(p):
            return p
    return None


def gen_two_column(path):
    """Right-column figure under a body text page that uses two columns."""
    doc = fitz.open()
    page = doc.new_page(width=595, height=842)
    pw = 595
    # Title spans full width
    page.insert_text((50, 60), "Two-column layout stress test",
                     fontsize=14, fontname="helv")

    left = fitz.Rect(50, 100, 285, 720)
    right = fitz.Rect(310, 100, 545, 720)
    body = ("Left column body text body text body text body text body "
            "text body text body text body text body text body text body "
            "text body text body text body text body text body text body "
            "text body text body text body text body text body text body "
            "text body text body text body text body text. ") * 4
    page.insert_textbox(left, body, fontsize=10, fontname="helv")

    # Right column: short body + figure cluster + caption
    page.insert_textbox(
        fitz.Rect(310, 100, 545, 230),
        ("Right column lead-in paragraph. The figure below appears only "
         "in this column, not spanning both columns."),
        fontsize=10, fontname="helv",
    )
    shape = page.new_shape()
    shape.draw_rect(fitz.Rect(330, 250, 520, 380))
    shape.draw_circle(fitz.Point(425, 315), 35)
    shape.finish(color=(0, 0, 0), width=1.2)
    shape.commit()
    page.insert_text((330, 410),
                     "Fig. 1. Single-column figure (right column only).",
                     fontsize=9, fontname="helv")
    # Discussion text below the caption (right col continues)
    page.insert_textbox(
        fitz.Rect(310, 430, 545, 720),
        ("Discussion. " + ("Discussion text in the right column. " * 12)),
        fontsize=10, fontname="helv",
    )

    # Page 2: references
    p2 = doc.new_page(width=595, height=842)
    p2.insert_text((50, 60), "References", fontsize=12, fontname="helv")
    p2.insert_textbox(
        fitz.Rect(50, 80, pw - 50, 200),
        "[1] A. One, B. Two. A paper. J. Tests, 2024, 1(1): 1-10.\n"
        "[2] C. Three. Another. J. Tests, 2024, 1(2): 11-20.\n"
        "[3] D. Four. Third. Test Press, 2023.",
        fontsize=10, fontname="helv",
    )

    doc.save(str(path))
    doc.close()


def gen_subfigs(path):
    """Three subfigures (a)/(b)/(c) sharing a single 'Fig. 1.' caption."""
    doc = fitz.open()
    page = doc.new_page(width=595, height=842)
    page.insert_text((50, 60), "Subfigure stress test",
                     fontsize=14, fontname="helv")
    page.insert_textbox(
        fitz.Rect(50, 90, 545, 180),
        ("This page tests the layout where three sub-figures share one "
         "Fig. 1. caption — the script should extract the whole row as a "
         "single fig1.png by default, not split it."),
        fontsize=10, fontname="helv",
    )

    # Three sub-figures arranged horizontally
    panels = [
        (90, 220, 200, 350, "(a)"),
        (220, 220, 330, 350, "(b)"),
        (350, 220, 460, 350, "(c)"),
    ]
    for x0, y0, x1, y1, label in panels:
        shape = page.new_shape()
        shape.draw_rect(fitz.Rect(x0, y0, x1, y1))
        shape.finish(color=(0, 0, 0), width=1.2)
        shape.commit()
        cx = (x0 + x1) / 2 - 8
        page.insert_text((cx, y1 + 14), label, fontsize=10, fontname="helv")

    # Single caption for all three
    page.insert_text((130, 390),
                     "Fig. 1. Three sub-panels: (a) box; (b) box; (c) box.",
                     fontsize=10, fontname="helv")
    page.insert_textbox(
        fitz.Rect(50, 420, 545, 700),
        "Discussion. " + ("Discussion text continues. " * 15),
        fontsize=10, fontname="helv",
    )

    p2 = doc.new_page(width=595, height=842)
    p2.insert_text((50, 60), "References", fontsize=12, fontname="helv")
    p2.insert_textbox(
        fitz.Rect(50, 80, 545, 180),
        "[1] X. One. A paper on subfigures. J. Layout, 2024.\n"
        "[2] Y. Two. Another. J. Layout, 2024.",
        fontsize=10, fontname="helv",
    )

    doc.save(str(path))
    doc.close()


def gen_gbt7714(path):
    """Chinese-source paper with GB/T 7714 numbered references.

    Skipped silently if no CJK font is available on this system —
    the eval entry can still be defined; running it just needs a
    machine with a CJK font.
    """
    font_path = cjk_font_path()
    if font_path is None:
        print(f"  [skip] {path.name}: no CJK font found on this system")
        return False
    doc = fitz.open()
    page = doc.new_page(width=595, height=842)
    page.insert_font(fontname="cjk", fontfile=font_path)

    page.insert_text((50, 60), "多车场电动车路径优化研究", fontsize=14, fontname="cjk")
    page.insert_text((50, 90), "（GB/T 7714 stress test）", fontsize=10, fontname="helv")

    body = ("摘要：本文以多车场电动车的实际配送任务为背景，提出一个考虑软时间窗与"
            "电池约束的混合整数规划模型，给出两类情景下的最优解策略。算例结果表明，"
            "允许车辆在多个车场之间灵活返回可显著降低总配送费用。")
    page.insert_textbox(fitz.Rect(50, 120, 545, 250),
                        body, fontsize=10, fontname="cjk")

    page.insert_text((50, 290), "参考文献", fontsize=12, fontname="cjk")
    refs = (
        "[1] 张三, 李四, 王五. 多车场电动汽车路径优化研究[J]. 系统工程理论与实践, "
        "2020, 40(5): 1234-1245.\n"
        "[2] 赵六. 物流网络设计与优化[M]. 北京: 清华大学出版社, 2018: 56-78.\n"
        "[3] 钱七, 孙八. 基于启发式的车辆路径问题求解算法[C]//"
        "中国运筹学会年会论文集. 北京: 中国运筹学会, 2019: 100-110.\n"
        "[4] 周九. 城市配送中心选址问题研究[D]. 上海: 同济大学, 2017.\n"
        "[5] Smith J, Brown K. Vehicle routing problem with time windows[J]. "
        "Operations Research, 2019, 67(4): 890-910."
    )
    page.insert_textbox(fitz.Rect(50, 310, 545, 700),
                        refs, fontsize=10, fontname="cjk")

    # Subset the embedded CJK font so the output PDF stays committable.
    # Without this a SimSun-embedded PDF is ~18 MB; subsetted it drops to ~50 KB.
    try:
        doc.subset_fonts()
    except Exception:
        pass
    doc.save(str(path), garbage=4, deflate=True)
    doc.close()
    return True


def gen_theorem_proof(path):
    """A math-paper-shape PDF whose body uses theorem / proof environments
    rendered as visible blocks. Tests whether the translator's LaTeX
    template needs `amsthm` + custom theorem styles, whether check_tex
    handles `\\begin{theorem}` / `\\end{proof}` correctly, and whether
    extract_text preserves the theorem block boundaries.
    """
    doc = fitz.open()
    page = doc.new_page(width=595, height=842)
    pw = 595
    page.insert_text((50, 60), "Theorem-heavy stress test (math/physics)",
                     fontsize=14, fontname="helv")
    page.insert_textbox(
        fitz.Rect(50, 100, pw - 50, 180),
        ("This document tests whether the skill's LaTeX template handles "
         "theorem / lemma / proof environments correctly. The body below "
         "is a typical math-paper layout."),
        fontsize=10, fontname="helv",
    )
    # Definition block (visually distinct prefix)
    y = 200
    page.insert_text((50, y), "Definition 1 (Smooth function).",
                     fontsize=10, fontname="hebo")  # bold-ish
    page.insert_textbox(
        fitz.Rect(50, y + 14, pw - 50, y + 56),
        ("A function f: R^n -> R is said to be smooth if all its partial "
         "derivatives of every order exist and are continuous on its domain."),
        fontsize=10, fontname="helv",
    )
    # Theorem
    y = 270
    page.insert_text((50, y), "Theorem 1 (Convergence).",
                     fontsize=10, fontname="hebo")
    page.insert_textbox(
        fitz.Rect(50, y + 14, pw - 50, y + 70),
        ("Let f be smooth on a bounded domain. Then the gradient descent "
         "iterates x_{k+1} = x_k - t_k grad f(x_k) with step sizes t_k "
         "satisfying the Wolfe conditions converge to a stationary point."),
        fontsize=10, fontname="helv",
    )
    # Proof block (longer)
    y = 350
    page.insert_text((50, y), "Proof.", fontsize=10, fontname="hebo")
    proof_text = (
        "By assumption f is smooth, hence its gradient is locally Lipschitz "
        "continuous. Applying the descent lemma and summing telescopically "
        "gives sum_k t_k ||grad f(x_k)||^2 < infty. Since each t_k is "
        "bounded away from zero, ||grad f(x_k)|| -> 0. By compactness of "
        "the sublevel sets, every limit point is stationary. QED."
    )
    page.insert_textbox(
        fitz.Rect(50, y + 14, pw - 50, y + 100),
        proof_text, fontsize=10, fontname="helv",
    )
    # Lemma + equation reference
    y = 480
    page.insert_text((50, y), "Lemma 1 (Cauchy-Schwarz).",
                     fontsize=10, fontname="hebo")
    page.insert_textbox(
        fitz.Rect(50, y + 14, pw - 50, y + 50),
        "For any vectors u, v in R^n, |<u, v>| <= ||u|| * ||v||.  "
        "(See equation (3.4.1) for the discrete-time analogue.)",
        fontsize=10, fontname="helv",
    )
    # Numbered equation in (n.m.p) format
    page.insert_text((50, 560),
                     "    | sum_{k=1}^{n} a_k b_k |^2  <=  "
                     "(sum a_k^2) * (sum b_k^2)         (3.4.1)",
                     fontsize=10, fontname="helv")
    # References — needed for extract_bibliography
    p2 = doc.new_page(width=595, height=842)
    p2.insert_text((50, 60), "References", fontsize=12, fontname="helv")
    p2.insert_textbox(
        fitz.Rect(50, 80, pw - 50, 220),
        "[1] Boyd, S., Vandenberghe, L. Convex Optimization. Cambridge, 2004.\n"
        "[2] Nesterov, Y. Introductory Lectures on Convex Optimization. Kluwer, 2004.\n"
        "[3] Bertsekas, D.P. Nonlinear Programming. 3rd ed. Athena Scientific, 2016.",
        fontsize=10, fontname="helv",
    )
    doc.save(str(path))
    doc.close()


def gen_longtable_multipage(path):
    """A regression-style output table that overflows across 2 pages.
    Tests whether extract_text preserves row alignment when a table
    crosses a page break (econ / stats paper failure mode).
    """
    doc = fitz.open()
    # Page 1: intro + table start
    page = doc.new_page(width=595, height=842)
    pw = 595
    page.insert_text((50, 60), "Multi-page table stress test (econ/stats)",
                     fontsize=14, fontname="helv")
    page.insert_textbox(
        fitz.Rect(50, 100, pw - 50, 180),
        ("Table 1 reports OLS estimates for the panel regression spanning "
         "2010-2023 across 28 industries. The table is intentionally long "
         "so that it overflows to page 2."),
        fontsize=10, fontname="helv",
    )
    page.insert_text((50, 200), "Table 1. OLS regression results (continued on next page).",
                     fontsize=9, fontname="hebo")
    # Header row
    headers = ["Variable", "Coef", "Std. Err.", "t-stat", "p-value", "95% CI"]
    col_x = [50, 170, 250, 330, 400, 470]
    y = 230
    for i, h in enumerate(headers):
        page.insert_text((col_x[i], y), h, fontsize=9, fontname="hebo")
    # 20 rows on page 1
    variables_p1 = [
        ("Intercept",      "2.1456",  "0.3214", "6.676",  "0.000", "[1.516, 2.775]"),
        ("GDP_growth",     "0.4823",  "0.0512", "9.420",  "0.000", "[0.382, 0.582]"),
        ("Energy_use",    "-0.1247",  "0.0331", "-3.767", "0.000", "[-0.190, -0.060]"),
        ("CO2_emissions",  "0.0891",  "0.0247", "3.607",  "0.000", "[0.041, 0.137]"),
        ("Renewable_share","-0.2134", "0.0418", "-5.107", "0.000", "[-0.295, -0.131]"),
        ("Population",     "0.0021",  "0.0008", "2.625",  "0.009", "[0.001, 0.004]"),
        ("Urbanization",   "0.1542",  "0.0287", "5.373",  "0.000", "[0.098, 0.210]"),
        ("Industrial_GVA", "0.3211",  "0.0445", "7.216",  "0.000", "[0.234, 0.408]"),
        ("R_and_D_spend",  "0.0871",  "0.0156", "5.583",  "0.000", "[0.057, 0.118]"),
        ("Trade_openness", "-0.0432", "0.0189", "-2.286", "0.022", "[-0.080, -0.006]"),
        ("Inflation",     "-0.0156",  "0.0093", "-1.677", "0.094", "[-0.034, 0.003]"),
        ("Unemployment", "-0.0298",  "0.0142", "-2.099", "0.036", "[-0.058, -0.002]"),
        ("Interest_rate", "-0.1023",  "0.0334", "-3.063", "0.002", "[-0.168, -0.037]"),
        ("Tax_rate",      "-0.0567",  "0.0211", "-2.687", "0.007", "[-0.098, -0.015]"),
        ("Education",      "0.2103",  "0.0367", "5.730",  "0.000", "[0.138, 0.282]"),
        ("Tech_imports",   "0.0934",  "0.0218", "4.284",  "0.000", "[0.051, 0.136]"),
        ("Public_invest",  "0.1567",  "0.0298", "5.258",  "0.000", "[0.098, 0.215]"),
        ("Foreign_invest", "0.0823",  "0.0223", "3.691",  "0.000", "[0.039, 0.126]"),
        ("Labor_force",    "0.0234",  "0.0089", "2.629",  "0.009", "[0.006, 0.041]"),
        ("Manufacturing",  "0.2891",  "0.0412", "7.017",  "0.000", "[0.208, 0.370]"),
    ]
    y = 245
    for row in variables_p1:
        for i, v in enumerate(row):
            page.insert_text((col_x[i], y), v, fontsize=9, fontname="helv")
        y += 14
    page.insert_text((50, y + 10), "(continued on next page)",
                     fontsize=8, fontname="hebo")

    # Page 2: table continues + intro of next section
    p2 = doc.new_page(width=595, height=842)
    p2.insert_text((50, 60), "Table 1 (continued).",
                   fontsize=9, fontname="hebo")
    for i, h in enumerate(headers):
        p2.insert_text((col_x[i], 90), h, fontsize=9, fontname="hebo")
    variables_p2 = [
        ("Services",       "0.1456",  "0.0298", "4.886",  "0.000", "[0.087, 0.204]"),
        ("Construction",   "0.0789",  "0.0234", "3.371",  "0.001", "[0.033, 0.125]"),
        ("Agriculture",   "-0.0234",  "0.0156", "-1.500", "0.134", "[-0.054, 0.007]"),
        ("Mining",        "-0.1023",  "0.0287", "-3.566", "0.000", "[-0.158, -0.046]"),
        ("Transport",      "0.1234",  "0.0245", "5.037",  "0.000", "[0.075, 0.171]"),
        ("Utilities",      "0.0567",  "0.0198", "2.864",  "0.004", "[0.018, 0.095]"),
        ("Health",         "0.0891",  "0.0212", "4.202",  "0.000", "[0.048, 0.130]"),
        ("Education_sec",  "0.1023",  "0.0234", "4.372",  "0.000", "[0.057, 0.148]"),
    ]
    y = 105
    for row in variables_p2:
        for i, v in enumerate(row):
            p2.insert_text((col_x[i], y), v, fontsize=9, fontname="helv")
        y += 14
    # Notes + R-squared
    p2.insert_text((50, y + 16),
                   "Notes: *** p<0.01, ** p<0.05, * p<0.10. "
                   "N=8,624. R-squared = 0.842. F-stat = 1,247.",
                   fontsize=8, fontname="helv")
    p2.insert_textbox(
        fitz.Rect(50, y + 50, pw - 50, y + 120),
        ("The estimates in Table 1 indicate that the renewable share and "
         "energy use coefficients are both statistically significant. "
         "Section 4 below discusses heterogeneity across subsamples."),
        fontsize=10, fontname="helv",
    )

    # Page 3: References
    p3 = doc.new_page(width=595, height=842)
    p3.insert_text((50, 60), "References", fontsize=12, fontname="helv")
    p3.insert_textbox(
        fitz.Rect(50, 80, pw - 50, 220),
        "[1] Wooldridge, J.M. Econometric Analysis of Cross Section and Panel Data. MIT Press, 2010.\n"
        "[2] Cameron, A.C., Trivedi, P.K. Microeconometrics. Cambridge, 2005.\n"
        "[3] Hsiao, C. Analysis of Panel Data. 3rd ed. Cambridge, 2014.",
        fontsize=10, fontname="helv",
    )
    doc.save(str(path))
    doc.close()


def gen_caption_with_sublabels(path):
    """A figure whose ONE caption contains multiple (a)/(b)/(c)/(d) sub-
    labels and is much longer than the typical 1-line caption. Tests
    whether `find_caption_rect` picks up the WHOLE caption (so the clip
    rect includes the figure block) and whether the density-above
    disambiguator still works when the caption itself is multi-line.
    Biology / medicine papers commonly use this format.
    """
    doc = fitz.open()
    page = doc.new_page(width=595, height=842)
    pw = 595
    page.insert_text((50, 60), "Caption-with-sublabels stress test (biology/medicine)",
                     fontsize=14, fontname="helv")
    page.insert_textbox(
        fitz.Rect(50, 90, pw - 50, 160),
        ("This page tests whether the caption picker handles long captions "
         "with multiple (a)/(b)/(c) sub-labels in the SAME caption body — "
         "common in biology / medicine."),
        fontsize=10, fontname="helv",
    )
    # Four sub-figures in a 2x2 grid
    panels = [
        (100, 200, 270, 340, "(a)"),
        (300, 200, 470, 340, "(b)"),
        (100, 360, 270, 500, "(c)"),
        (300, 360, 470, 500, "(d)"),
    ]
    for x0, y0, x1, y1, label in panels:
        shape = page.new_shape()
        shape.draw_rect(fitz.Rect(x0, y0, x1, y1))
        shape.draw_circle(fitz.Point((x0+x1)/2, (y0+y1)/2), 25)
        shape.finish(color=(0, 0, 0), width=1.0)
        shape.commit()
        page.insert_text((x0 + 6, y0 + 14), label,
                         fontsize=9, fontname="hebo")
    # MULTI-LINE caption with sub-labels — the failure mode is that
    # the picker stops at the first ". (a)" or similar.
    caption_block = (
        "Fig. 1. Comparison of treatment outcomes across four cohorts. "
        "(a) Control cohort at 24 h: baseline morphology preserved with "
        "no observable apoptosis. (b) Low-dose cohort at 24 h: ~12% of "
        "cells show early apoptotic markers; nuclear condensation visible "
        "in subset. (c) High-dose cohort at 24 h: extensive cytoplasmic "
        "fragmentation; >60% of cells positive for Annexin V staining. "
        "(d) Recovery cohort at 72 h post-washout: partial restoration of "
        "baseline morphology; residual marker expression in ~15% of cells. "
        "Scale bar: 50 micrometers. All images are representative of n=6 "
        "biological replicates per condition."
    )
    page.insert_textbox(
        fitz.Rect(50, 520, pw - 50, 720),
        caption_block, fontsize=9, fontname="helv",
    )
    # Page 2: references (Vancouver-numbered, biomed style)
    p2 = doc.new_page(width=595, height=842)
    p2.insert_text((50, 60), "References", fontsize=12, fontname="helv")
    p2.insert_textbox(
        fitz.Rect(50, 80, pw - 50, 220),
        "1. Kerr JF, Wyllie AH, Currie AR. Apoptosis: a basic biological "
        "phenomenon with wide-ranging implications in tissue kinetics. "
        "Br J Cancer. 1972;26(4):239-257.\n"
        "2. Vermes I, Haanen C, Steffens-Nakken H, Reutelingsperger C. A "
        "novel assay for apoptosis. J Immunol Methods. 1995;184(1):39-51.\n"
        "3. Galluzzi L, Vitale I, et al. Molecular definitions of cell death "
        "subroutines. Cell Death Differ. 2012;19(1):107-120.",
        fontsize=10, fontname="helv",
    )
    doc.save(str(path))
    doc.close()


def main():
    print(f"Generating stress-test PDFs to {OUT}")
    gen_two_column(OUT / "stress-2col.pdf")
    print("  wrote stress-2col.pdf")
    gen_subfigs(OUT / "stress-subfigs.pdf")
    print("  wrote stress-subfigs.pdf")
    if gen_gbt7714(OUT / "stress-gbt7714.pdf"):
        print("  wrote stress-gbt7714.pdf")
    gen_theorem_proof(OUT / "stress-theorems.pdf")
    print("  wrote stress-theorems.pdf")
    gen_longtable_multipage(OUT / "stress-longtable.pdf")
    print("  wrote stress-longtable.pdf")
    gen_caption_with_sublabels(OUT / "stress-caption-sublabels.pdf")
    print("  wrote stress-caption-sublabels.pdf")


if __name__ == "__main__":
    main()

{
  "skill_name": "pdf-to-chinese-latex",
  "evals": [
    {
      "id": "eval-0-tan-mdevrpstw",
      "prompt": "Translate the academic PDF into a Chinese LaTeX project that compiles in Overleaf with XeLaTeX. Preserve equations (with original numbering), tables, figures, and bibliography 1:1 with the source. Use the ctex package for Chinese typesetting. The output directory must contain main.tex (full Chinese translation), images/ (extracted figures), and a short README with Overleaf compile steps. Assumptions for this run (since you can't ask the user interactively): extract figures from the original PDF and embed them with Chinese captions; do a full translation (abstract, every body section, references). At minimum produce title, authors, abstract, introduction, model/methods section, and the bibliography; continue with results/conclusion as time permits.",
      "files": ["test_inputs/eval-0-input.pdf"],
      "expected_output": "Folder containing main.tex (ctex), images/ with extracted figures, README.md. main.tex should compile under XeLaTeX without errors and render Chinese correctly."
    },
    {
      "id": "eval-1-liu-drone-hbdn",
      "prompt": "Translate the academic PDF into a Chinese LaTeX project that compiles in Overleaf with XeLaTeX. Preserve equations (with original numbering), tables, figures, algorithm pseudocode, and bibliography 1:1 with the source. Use the ctex package for Chinese typesetting. The output directory must contain main.tex (full Chinese translation), images/ (extracted figures), and a short README with Overleaf compile steps. Assumptions for this run (since you can't ask the user interactively): extract figures from the original PDF and embed them with Chinese captions; do a full translation. This paper contains vector network diagrams (not raster); pay special attention to preserving them. At minimum produce title, authors, abstract, introduction, problem formulation, the algorithm section with pseudocode, and the bibliography; continue with experiments/conclusion as time permits.",
      "files": ["test_inputs/eval-1-input.pdf"],
      "expected_output": "Folder containing main.tex (ctex), images/ with extracted figures (5+ figures expected; some are vector), README.md. main.tex should compile under XeLaTeX without errors and render Chinese, math, and algorithm pseudocode correctly."
    },
    {
      "id": "eval-2-stress-2col",
      "prompt": "This is a layout stress test against a synthetic 2-column-layout PDF. Run `python scripts/render_figures.py evals/test_inputs/stress-2col.pdf --out <out>/images --auto` and verify the resulting fig1.png contains only the right-column figure (rectangle + circle + caption), with NO left-column body text bleeding into the crop. Also run `python scripts/extract_bibliography.py` on the extracted raw.txt and verify all 3 references are parsed.",
      "files": ["test_inputs/stress-2col.pdf"],
      "expected_output": "fig1.png cropped to right column only (clip x ≈ 326..524); 3 bibitems parsed."
    },
    {
      "id": "eval-3-stress-subfigs",
      "prompt": "This is a layout stress test against a synthetic PDF where one Fig. 1 caption sits under three sub-figures (a)/(b)/(c) arranged horizontally. Run `python scripts/render_figures.py evals/test_inputs/stress-subfigs.pdf --out <out>/images --auto` and verify the resulting fig1.png contains ALL THREE sub-panels plus the caption (default behaviour — sub-figures are not split unless the user explicitly requests it).",
      "files": ["test_inputs/stress-subfigs.pdf"],
      "expected_output": "Single fig1.png covering all three sub-figures + the shared caption; clip aspect ≈ 2:1 landscape."
    },
    {
      "id": "eval-4-stress-gbt7714",
      "prompt": "This is a bibliography parsing stress test against a synthetic Chinese-source paper with GB/T 7714 numbered references. Run `python scripts/extract_text.py` then `python scripts/extract_bibliography.py` against evals/test_inputs/stress-gbt7714.pdf. Verify all 5 references (mix of Chinese [J]/[M]/[C]/[D] entries and one English entry) are parsed with content preserved.",
      "files": ["test_inputs/stress-gbt7714.pdf"],
      "expected_output": "5 bibitems parsed; Chinese characters intact; [J]/[M]/[C]/[D] document-type markers preserved in entry text."
    },
    {
      "id": "eval-5-publisher-arxiv",
      "prompt": "Pipeline regression on an arXiv preprint (no publisher watermark, no journal page header — only the right-edge vertical 'arXiv:NNNN.NNNNN' stamp). Run extract_text + render_figures --auto + extract_bibliography against evals/test_inputs/eval-5-arxiv-input.pdf. Verify all three stages exit 0 with non-trivial counts.",
      "files": ["test_inputs/eval-5-arxiv-input.pdf"],
      "expected_output": "extract_text writes ≥ 50K chars; render_figures --auto detects ≥ 8 figures with 0 fallback and 0 invalid PNGs (magic bytes); extract_bibliography parses ≥ 30 entries."
    },
    {
      "id": "eval-6-publisher-mdpi",
      "prompt": "Pipeline regression on an MDPI open-access paper (Drones journal). Verify the boilerplate Creative Commons / 'Copyright © by the authors' license text in the body doesn't break bibliography section detection, and figures crop correctly despite MDPI's narrow margins. Run the full pipeline against evals/test_inputs/eval-6-mdpi-input.pdf.",
      "files": ["test_inputs/eval-6-mdpi-input.pdf"],
      "expected_output": "extract_text writes ≥ 50K chars; render_figures --auto detects ≥ 5 figures with 0 fallback; extract_bibliography parses ≥ 30 entries."
    },
    {
      "id": "eval-7-publisher-springer-scirep",
      "prompt": "Pipeline regression on a Springer Nature paper (Scientific Reports) — distinct page-header style versus Elsevier/Wiley, longer bibliography. Run the full pipeline against evals/test_inputs/eval-7-springer-scirep-input.pdf.",
      "files": ["test_inputs/eval-7-springer-scirep-input.pdf"],
      "expected_output": "extract_text writes ≥ 60K chars; render_figures --auto detects ≥ 5 figures with 0 fallback; extract_bibliography parses ≥ 50 entries (Sci. Rep. papers often have 60+)."
    },
    {
      "id": "eval-8-publisher-chinese-journal",
      "prompt": "Pipeline regression on a Chinese-language journal paper. Verifies that the Chinese caption keywords (图 N. / 图N. / 图 N: / 图N: / 图 N：/ 图N： / 图 N / 图N) work, and that mixed-language bibliography is parsed. Run the full pipeline against evals/test_inputs/eval-8-chinese-journal-input.pdf.",
      "files": ["test_inputs/eval-8-chinese-journal-input.pdf"],
      "expected_output": "extract_text writes ≥ 40K chars of mostly Chinese text; render_figures --auto detects ≥ 3 figures via Chinese caption pattern with 0 fallback; extract_bibliography parses ≥ 15 entries."
    },
    {
      "id": "eval-9-stress-theorem-proof",
      "prompt": "Discipline stress test (math / theoretical physics). Synthetic fixture with `Theorem 1` / `Lemma 1` / `Proof` / `Definition 1` blocks rendered as labelled body text, plus an equation with (n.m.p) numbering. Run extract_text and verify the structural markers survive in raw.txt so a translator can map them to `\\begin{theorem}` / `\\begin{proof}` etc. The skill's latex_template.tex already preloads amsthm + \\newtheorem{theorem}{定理}/{lemma}{引理}/etc., so the produced main.tex must compile.",
      "files": ["test_inputs/stress-theorems.pdf"],
      "expected_output": "raw.txt contains 'Theorem 1', 'Lemma 1', 'Proof', 'Definition 1', and '(3.4.1)' each at least once; latex_template.tex already defines \\newtheorem environments for all of these in Chinese."
    },
    {
      "id": "eval-10-stress-longtable",
      "prompt": "Discipline stress test (econometrics / statistics). Synthetic fixture with an OLS regression output table containing 28 variable rows that overflows from page 1 to page 2, including a header row, numeric coefficients, standard errors, t-stats, p-values, and 95% CI brackets. Run extract_text and verify row alignment and numeric data survive the page break.",
      "files": ["test_inputs/stress-longtable.pdf"],
      "expected_output": "raw.txt contains all 28 variable names (Intercept ... Education_sec), the GDP_growth coefficient '0.4823', the page-break marker 'continued on next page', the '95% CI' header on page 2, and bracketed CI ranges like [0.382, 0.582] intact."
    },
    {
      "id": "eval-11-stress-caption-sublabels",
      "prompt": "Discipline stress test (biology / medicine). Synthetic fixture with one Fig. 1 whose caption contains FOUR sub-panel sub-labels (a)/(b)/(c)/(d), each followed by long descriptive prose totaling ~10 lines. Verify that render_figures --auto produces ONE fig1.png covering all four panels (not 4 separate files) and that the long caption text survives in raw.txt so the translator can reproduce the (a)/(b)/(c)/(d) sub-descriptions in Chinese.",
      "files": ["test_inputs/stress-caption-sublabels.pdf"],
      "expected_output": "Single fig1.png with all four sub-panels visible; raw.txt contains 'Fig. 1', '(a)', '(b)', '(c)', '(d)', 'Scale bar', and 'Annexin V' (the last sub-label's terminal phrase) each at least once."
    }
  ],
  "notes": {
    "test_inputs_provenance": "eval-0 / eval-1 inputs are real published papers and not committed to this repo (separate workspace). eval-2 / eval-3 / eval-4 inputs are synthetic fixtures generated by `evals/_gen_stress_pdfs.py` and committed under `evals/test_inputs/`. eval-5..8 are real published papers committed to test_inputs/ as a publisher-coverage regression set (arXiv / MDPI / Springer Nature / Chinese journal).",
    "publisher_coverage": "After eval-5..8 the regression set covers 8 distinct publisher / format combinations: Hindawi single-col (eval-0), Wiley single-col (eval-1), arXiv (eval-5), MDPI (eval-6), Springer Nature (eval-7), Chinese-language journal (eval-8), plus synthetic 2-col / sub-figures / GB/T 7714 stress fixtures (eval-2..4). Real Elsevier coverage lives in the project's `Elsevier批量测试输出/` folder (6 papers; see _summary.csv there).",
    "discipline_coverage": "eval-9..11 add discipline-specific structural stress: theorem/proof environments (math/physics), multi-page longtable (econometrics/stats), and long captions with (a)/(b)/(c)/(d) sub-labels (biology/medicine). These were generated synthetically because the skill operates on structural primitives (text + figures + bib + math envs) rather than discipline semantics — most cross-discipline differences (chemistry symbols, electrical diagrams, gel images) are content-level not structure-level and don't expose new failure modes in the pipeline."
  }
}

MIT License

Copyright (c) 2026 the pdf-to-chinese-latex authors

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

% !TEX program = xelatex
% =========================================================
%  中文译稿模板 -- pdf-to-chinese-latex skill
%
%  使用方法：
%   1. 把所有 <...> 占位符替换为译稿内容；
%   2. 在 images/ 目录下放置 fig1.png, fig2.png, ... ；
%   3. Overleaf 编译方式：Menu -> Compiler 选 XeLaTeX -> Recompile。
%
%  本模板预加载的包足以支撑大部分学术论文：
%   - ctex          中文支持（XeLaTeX）
%   - amsmath/amssymb/amsthm  公式
%   - graphicx + booktabs + tabularx + longtable + multirow  图与表
%   - algorithm + algpseudocode  算法伪代码
%   - hyperref + url  超链接
%   - enumitem    自定义列表
% =========================================================
\documentclass[a4paper,11pt]{article}

\usepackage[UTF8,heading=true]{ctex}
\usepackage[a4paper,margin=2.5cm]{geometry}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{array}
\usepackage{longtable}
\usepackage{caption}
\usepackage{enumitem}
\usepackage{algorithm}
\usepackage{algpseudocode}
\usepackage{xcolor}
\usepackage[hidelinks,colorlinks=false]{hyperref}
\usepackage{url}
\usepackage{tabularx}
\usepackage{multirow}
\usepackage{float}        % provides the [H] specifier to FORCE a figure
                          % into position (no floating). Use only when a
                          % regular [!htbp] still defers the figure.

\graphicspath{{images/}}  % so \includegraphics{fig1.png} resolves to images/fig1.png

% --- Figure inclusion convention (read before pasting figures) -----------
% Default to [!htbp] placement and pick width from the PNG's aspect ratio:
%
%   aspect ≥ 2.5  → \includegraphics[width=0.85\textwidth]{figN.png}  (very wide banner)
%   1.5 ≤ < 2.5   → \includegraphics[width=0.75\textwidth]{figN.png}  (standard landscape)
%   0.9 ≤ < 1.5   → \includegraphics[width=0.65\textwidth]{figN.png}  (square-ish)
%   < 0.9         → \includegraphics[width=0.55\textwidth]{figN.png}  (portrait)
%
% render_figures.py emits an `_figure_includes.tex` sidecar with these
% widths already filled in for every extracted figure — paste from there.
% -------------------------------------------------------------------------

\graphicspath{{images/}}
\captionsetup{font=small,labelfont=bf}
\setlength{\parskip}{0.4em}

% 自定义环境
\newtheorem{definition}{定义}
\newtheorem{assumption}{假设}
\newtheorem{theorem}{定理}
\newtheorem{lemma}{引理}
\newtheorem{proposition}{命题}

\title{\bfseries <中文标题>}
\author{<罗马字作者 1>\textsuperscript{*}\quad <罗马字作者 2>\\[2pt]
\small\itshape <作者机构 1>\\
\small\itshape <作者机构 2>\\
\small 通讯作者邮箱：\href{mailto:<email>}{<email>}}
\date{原文发表于 \emph{<期刊全名>}，<年份>，第 <卷> 卷 第 <期> 期，pp.\ <起>--<止>\\
\small DOI:\ \href{<DOI URL>}{<DOI>}}

\begin{document}
\maketitle

\begin{abstract}
\noindent <摘要正文翻译>

\noindent\textbf{关键词：}<关键词 1>；<关键词 2>；<关键词 3>
\end{abstract}

% =========================================================
\section{引言}
\label{sec:intro}

<引言正文翻译...>

% =========================================================
\section{<第二节标题>}
\label{sec:methods}

<正文翻译...>

% --- 公式示例 ---
% 用 \label{eq:cN} 对应原文的 (N) 式，正文里用 \eqref{eq:cN} 引用。
\begin{equation}
\label{eq:c1}
\min z = \sum_{i=1}^{n} c_i x_i.
\end{equation}

约束如式 \eqref{eq:c2} 所示：
\begin{equation}
\label{eq:c2}
\sum_{j=1}^{m} a_{ij} x_j \le b_i,\quad \forall i.
\end{equation}

% --- 图示例 ---
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{fig1.png}
\caption{<图 1 的中文标题>}
\label{fig:f1}
\end{figure}

% --- 表示例（用 tabularx + booktabs）---
\begin{table}[ht]
\centering
\caption{<表 1 的中文标题>}
\label{tab:t1}
\begin{tabularx}{0.85\textwidth}{c X r}
\toprule
参数 & 含义 & 数值 \\
\midrule
$\alpha$ & <含义说明> & 1.0 \\
$\beta$  & <含义说明> & 0.5 \\
\bottomrule
\end{tabularx}
\end{table}

% --- 算法示例 ---
\begin{algorithm}[ht]
\caption{<算法名称>}
\label{alg:a1}
\begin{algorithmic}[1]
\State \textbf{输入：}<输入说明>
\State \textbf{输出：}<输出说明>
\For{$i=1$ to $n$}
  \If{<条件>}
    \State <操作>
  \EndIf
\EndFor
\State \Return <返回值>
\end{algorithmic}
\end{algorithm}

% =========================================================
\section{结论}
\label{sec:conclusion}

<结论翻译...>

\subsection*{致谢}
<致谢翻译...>

% =========================================================
\begin{thebibliography}{99}
\small
\bibitem{Key1} <Author>, <Title>, <Journal>, <Year>. doi: <DOI>.
\bibitem{Key2} <Author>, <Title>, <Journal>, <Year>. doi: <DOI>.
\end{thebibliography}

\end{document}

# -*- coding: utf-8 -*-
"""Static-validate a LaTeX file before sending it to Overleaf.

Checks:
    * \\begin{env} / \\end{env} balance per environment
    * \\cite{key} <-> \\bibitem{key} correspondence
    * \\ref / \\eqref <-> \\label correspondence
    * each line's $...$ count is even (catches stray $ in prose)
    * every \\includegraphics{...} target exists AND has valid image
      magic bytes (catches placeholder text files masquerading as PNGs)

Exits 0 if everything looks fine, 1 otherwise. Use this before declaring
a translation "done" -- a clean static check catches the majority of bugs
that would otherwise force a round-trip to Overleaf.

Usage:
    python check_tex.py path/to/main.tex
"""

import argparse
import io
import os
import re
import sys
from collections import Counter


# Image format magic-byte signatures. The check looks for one of these
# at the start of the file; anything else (including 12-byte ASCII
# placeholders saved as `placeholder\n`) fails as "not a real image".
IMG_MAGIC = [
    ("PNG",  b"\x89PNG\r\n\x1a\n"),
    ("JPEG", b"\xff\xd8\xff"),
    ("PDF",  b"%PDF"),
    ("GIF87", b"GIF87a"),
    ("GIF89", b"GIF89a"),
    ("TIFF_LE", b"II*\x00"),
    ("TIFF_BE", b"MM\x00*"),
    ("EPS",  b"%!PS"),
    ("BMP",  b"BM"),
    ("WebP", b"RIFF"),  # generic RIFF header; webp specifically is RIFF....WEBP
]

# Extensions LaTeX will silently auto-append when \includegraphics{X}
# is given without one.
LATEX_GRAPHICS_EXTS = (".png", ".jpg", ".jpeg", ".pdf", ".eps", ".tif", ".tiff", ".bmp")


def check_image_file(path: str) -> "tuple[str | None, str]":
    """Return (kind, error). kind is None when the file isn't a valid image."""
    if not os.path.isfile(path):
        return None, "file not found"
    try:
        size = os.path.getsize(path)
    except OSError as e:
        return None, f"stat failed: {e}"
    if size == 0:
        return None, "0-byte file"
    try:
        with open(path, "rb") as f:
            head = f.read(16)
    except OSError as e:
        return None, f"read failed: {e}"
    for name, sig in IMG_MAGIC:
        if head.startswith(sig):
            return name, f"OK {name} ({size} B)"
    # Provide a hint about what was actually in the file
    try:
        snippet = head.decode("ascii", errors="replace").replace("\r", "").replace("\n", "\\n")
    except Exception:
        snippet = repr(head[:8])
    return None, f"unrecognized header {snippet!r} ({size} B) -- placeholder or corrupted?"


def expand_inputs(path: str, _seen: "set[str] | None" = None, _depth: int = 0) -> str:
    """Read `path` and recursively inline every `\\input{X}` / `\\include{X}`
    referenced from it. Resolves X relative to `path`'s directory; tries the
    target as-is, then with a `.tex` extension. Catches circular includes
    via `_seen`. Comments (lines that start with %) are skipped so that
    commented-out \\input statements don't trigger expansion.

    Universal: lets check_tex.py handle modular main.tex projects that
    split body / bibliography / preamble across files via \\input{}.
    Previously, `\\bibitem` entries in an \\input'd file were invisible to
    cite/bibitem matching, causing false-positive MISSING reports.
    """
    if _seen is None:
        _seen = set()
    abs_path = os.path.abspath(path)
    if abs_path in _seen or _depth > 10:
        return ""  # circular or absurd depth — bail
    _seen.add(abs_path)
    base_dir = os.path.dirname(abs_path)
    try:
        with open(path, encoding="utf-8") as f:
            txt = f.read()
    except OSError:
        return ""

    inc_re = re.compile(r"\\(?:input|include)\s*\{([^}]+)\}")

    def replace(m: "re.Match[str]") -> str:
        target = m.group(1).strip()
        # try as-is, then with .tex appended
        for cand in (target, target + ".tex"):
            full = cand if os.path.isabs(cand) else os.path.join(base_dir, cand)
            if os.path.isfile(full):
                return "\n" + expand_inputs(full, _seen, _depth + 1) + "\n"
        # leave the directive in place if we can't find the file — the
        # downstream checks will flag the resulting missing bibitems / etc.
        return m.group(0)

    # Mask out commented lines so commented \input doesn't get expanded.
    def is_commented(start: int) -> bool:
        line_start = txt.rfind("\n", 0, start) + 1
        prefix = txt[line_start:start]
        i = 0
        while i < len(prefix):
            if prefix[i] == "\\" and i + 1 < len(prefix):
                i += 2; continue
            if prefix[i] == "%":
                return True
            i += 1
        return False

    out = []
    pos = 0
    for m in inc_re.finditer(txt):
        out.append(txt[pos:m.start()])
        if is_commented(m.start()):
            out.append(m.group(0))
        else:
            out.append(replace(m))
        pos = m.end()
    out.append(txt[pos:])
    return "".join(out)


def check(path: str) -> int:
    # Inline \input{}/\include{} so bibitems / labels / images defined in
    # split-out files are visible to the static checks below.
    txt = expand_inputs(path)
    if not txt:
        # Fall back to raw read if expansion path failed
        with open(path, encoding="utf-8") as f:
            txt = f.read()

    issues = 0

    # 1. begin/end balance
    begs = re.findall(r"\\begin\{([^}]+)\}", txt)
    ends = re.findall(r"\\end\{([^}]+)\}", txt)
    bc, ec = Counter(begs), Counter(ends)
    for k in set(list(bc) + list(ec)):
        if bc[k] != ec[k]:
            print(f"  MISMATCH \\begin{{{k}}}={bc[k]} \\end{{{k}}}={ec[k]}")
            issues += 1
    print(f"  begin total={sum(bc.values())}, end total={sum(ec.values())}")

    # 2. cite ↔ bibitem
    cites = re.findall(r"\\cite[a-zA-Z]*\{([^}]+)\}", txt)
    cite_keys = set()
    for c in cites:
        for k in c.split(","):
            cite_keys.add(k.strip())
    bibs = set(re.findall(r"\\bibitem\{([^}]+)\}", txt))
    missing = cite_keys - bibs
    unused = bibs - cite_keys
    if missing:
        print(f"  MISSING bibitems for: {sorted(missing)}")
        issues += 1
    if unused:
        print(f"  UNUSED bibitems: {sorted(unused)}")
        # unused is usually a soft warning, not an error
    print(f"  cited={len(cite_keys)} bibitems={len(bibs)}")

    # 3. ref / eqref ↔ label
    labels = set(re.findall(r"\\label\{([^}]+)\}", txt))
    refs = set(re.findall(r"\\ref\{([^}]+)\}", txt)) | set(
        re.findall(r"\\eqref\{([^}]+)\}", txt)
    )
    missing_lbl = refs - labels
    if missing_lbl:
        print(f"  REFS without matching \\label: {sorted(missing_lbl)}")
        issues += 1
    print(f"  labels={len(labels)} refs={len(refs)}")

    # 4. per-line $...$ parity
    odd_lines = []
    for i, ln in enumerate(txt.split("\n"), 1):
        s = ln
        cnt = 0
        j = 0
        while j < len(s):
            if s[j] == "\\" and j + 1 < len(s):
                # skip escaped char (e.g., \$)
                j += 2
                continue
            if s[j] == "$":
                cnt += 1
            j += 1
        if cnt % 2:
            odd_lines.append((i, cnt, ln[:80]))
    if odd_lines:
        print(f"  ODD $ count on {len(odd_lines)} line(s):")
        for i, c, l in odd_lines[:10]:
            print(f"    line {i} (count={c}): {l}")
        issues += 1

    # 5. \includegraphics target must be a real image, not a placeholder.
    # Honors \graphicspath{{dir1/}{dir2/}} from the preamble.
    base_dir = os.path.dirname(os.path.abspath(path))
    # Collect search dirs: main.tex dir first, then each graphicspath entry.
    search_dirs = [base_dir]
    gp_re = re.compile(r"\\graphicspath\s*\{((?:\{[^}]*\}\s*)+)\}")
    for gp_match in gp_re.findall(txt):
        for d in re.findall(r"\{([^}]*)\}", gp_match):
            if d:
                resolved = d if os.path.isabs(d) else os.path.join(base_dir, d)
                search_dirs.append(resolved)
    inc_re = re.compile(r"\\includegraphics(?:\s*\[[^\]]*\])?\s*\{([^}]+)\}")
    img_paths = inc_re.findall(txt)
    bad_imgs = []
    for ref in img_paths:
        ref = ref.strip()
        # If ref has no extension, LaTeX would try extensions in order.
        extensions_to_try = [""] if any(
            ref.lower().endswith(e) for e in LATEX_GRAPHICS_EXTS
        ) else [""] + list(LATEX_GRAPHICS_EXTS)
        kind = None
        last_err = "no candidate matched"
        for ext in extensions_to_try:
            candidate_name = ref + ext if ext else ref
            for d in search_dirs:
                abs_path = candidate_name if os.path.isabs(candidate_name) \
                    else os.path.join(d, candidate_name)
                k, err = check_image_file(abs_path)
                if k:
                    kind = k
                    break
                last_err = err
            if kind:
                break
        if not kind:
            bad_imgs.append((ref, last_err))
    if bad_imgs:
        print(f"  BAD \\includegraphics targets ({len(bad_imgs)}):")
        for r, e in bad_imgs[:10]:
            print(f"    {r}: {e}")
        issues += 1
    print(f"  images={len(img_paths)} valid={len(img_paths) - len(bad_imgs)}"
          + (f"  (search_dirs: {[os.path.basename(d) or '.' for d in search_dirs]})"
             if len(search_dirs) > 1 else ""))

    if issues == 0:
        print("OK: no issues detected.")
    else:
        print(f"FAIL: {issues} issue(s) detected.")
    return issues


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    parser.add_argument("tex", help="path to main.tex")
    args = parser.parse_args()

    if sys.stdout.encoding and sys.stdout.encoding.lower() != "utf-8":
        sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8")

    sys.exit(1 if check(args.tex) else 0)


if __name__ == "__main__":
    main()

# -*- coding: utf-8 -*-
"""Compile a Chinese LaTeX project (ctex + XeLaTeX) to PDF locally.

Runs xelatex twice (second pass resolves cross-references) on the
main.tex inside the given project directory. If xelatex is not in PATH,
prints installation guidance and the Overleaf-upload fallback instead.

Usage:
    python compile_pdf.py path/to/project_folder
    python compile_pdf.py path/to/project_folder --main main.tex --keep-aux

By default the script removes auxiliary files (.aux/.log/.out/.toc/.synctex.gz)
after a successful compile, keeping only main.pdf next to main.tex. Pass
--keep-aux to preserve them (useful for debugging compile errors).

Requires: a local TeX distribution with XeLaTeX. Recommended:
  - Windows : winget install MiKTeX.MiKTeX
  - macOS   : brew install --cask mactex
  - Linux   : apt install texlive-xetex texlive-lang-chinese
"""

import argparse
import io
import os
import shutil
import subprocess
import sys
from pathlib import Path
from typing import Optional


AUX_SUFFIXES = (".aux", ".log", ".out", ".toc", ".synctex.gz", ".fls", ".fdb_latexmk")


def find_xelatex() -> Optional[str]:
    """Return path to xelatex.exe/xelatex, or None if not in PATH."""
    p = shutil.which("xelatex")
    if p:
        return p
    # On Windows, MikTeX may not be in current shell's PATH right after install.
    # Try common install paths as a last resort.
    if sys.platform == "win32":
        candidates = [
            os.path.expandvars(r"%LOCALAPPDATA%\Programs\MiKTeX\miktex\bin\x64\xelatex.exe"),
            r"C:\Program Files\MiKTeX\miktex\bin\x64\xelatex.exe",
            r"C:\texlive\2024\bin\windows\xelatex.exe",
            r"C:\texlive\2025\bin\windows\xelatex.exe",
        ]
        for c in candidates:
            if os.path.isfile(c):
                return c
    return None


def print_install_help() -> None:
    print("[ERROR] xelatex not found on PATH.\n")
    print("To compile locally, install a TeX distribution with XeLaTeX:")
    print("  Windows : winget install MiKTeX.MiKTeX")
    print("  macOS   : brew install --cask mactex")
    print("  Linux   : sudo apt install texlive-xetex texlive-lang-chinese\n")
    print("Or, as a fallback, upload to Overleaf:")
    print("  1. Zip up {main.tex, images/, README.md}")
    print("  2. New Project -> Upload Project on https://www.overleaf.com")
    print("  3. Menu -> Compiler -> XeLaTeX -> Recompile\n")


def compile_one(project_dir: Path, main_name: str, keep_aux: bool, xelatex: str) -> int:
    main_tex = project_dir / main_name
    if not main_tex.is_file():
        print(f"[ERROR] {main_tex} does not exist")
        return 2

    print(f"[1/2] xelatex pass 1: {main_tex.name}")
    r1 = subprocess.run(
        [xelatex, "-interaction=nonstopmode", "-halt-on-error", main_name],
        cwd=str(project_dir),
        capture_output=True, text=True, encoding="utf-8", errors="replace",
    )
    if r1.returncode != 0:
        print(f"[ERROR] pass 1 failed (exit={r1.returncode}). Last lines of stdout:")
        for ln in (r1.stdout or "").splitlines()[-25:]:
            print(f"    {ln}")
        return r1.returncode

    print("[2/2] xelatex pass 2 (resolve cross-references)")
    r2 = subprocess.run(
        [xelatex, "-interaction=nonstopmode", "-halt-on-error", main_name],
        cwd=str(project_dir),
        capture_output=True, text=True, encoding="utf-8", errors="replace",
    )
    if r2.returncode != 0:
        print(f"[ERROR] pass 2 failed (exit={r2.returncode}). Last lines of stdout:")
        for ln in (r2.stdout or "").splitlines()[-25:]:
            print(f"    {ln}")
        return r2.returncode

    # Inspect the resulting PDF
    pdf_path = main_tex.with_suffix(".pdf")
    if not pdf_path.is_file():
        print("[ERROR] xelatex exited 0 but no PDF was written.")
        return 3

    size_kb = pdf_path.stat().st_size / 1024
    n_pages = "?"
    try:
        import fitz
        with fitz.open(str(pdf_path)) as d:
            n_pages = len(d)
    except Exception:
        pass
    print(f"[OK]  {pdf_path}  ({size_kb:.1f} KB, {n_pages} pages)")

    # Clean up aux files unless --keep-aux
    if not keep_aux:
        stem = main_tex.stem
        for sfx in AUX_SUFFIXES:
            f = project_dir / f"{stem}{sfx}"
            if f.is_file():
                try:
                    f.unlink()
                except OSError:
                    pass
    return 0


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    parser.add_argument(
        "project_dir",
        help="directory containing main.tex (and images/)",
    )
    parser.add_argument(
        "--main", default="main.tex",
        help="main .tex filename inside project_dir (default: main.tex)",
    )
    parser.add_argument(
        "--keep-aux", action="store_true",
        help="keep .aux/.log/.out/etc. after compile (useful for debugging)",
    )
    args = parser.parse_args()

    if sys.stdout.encoding and sys.stdout.encoding.lower() != "utf-8":
        sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8")

    xelatex = find_xelatex()
    if not xelatex:
        print_install_help()
        sys.exit(127)

    project_dir = Path(args.project_dir).resolve()
    if not project_dir.is_dir():
        print(f"[ERROR] {project_dir} is not a directory")
        sys.exit(2)

    print(f"Using xelatex: {xelatex}")
    sys.exit(compile_one(project_dir, args.main, args.keep_aux, xelatex))


if __name__ == "__main__":
    main()

# -*- coding: utf-8 -*-
"""Extract page-by-page text from a PDF.

Usage:
    python extract_text.py input.pdf --out output_dir/raw.txt

The output uses `========== PAGE N ==========` delimiters so a translator
(human or LLM) can find specific pages by line search.

Requires: pdfplumber (`pip install pdfplumber`).
"""

import argparse
import io
import os
import sys

import pdfplumber


def extract(pdf_path: str, out_path: str) -> int:
    """Extract text. Returns total chars written. Raises SystemExit if the
    PDF appears to have no text layer (scanned image PDF — needs OCR first).
    """
    os.makedirs(os.path.dirname(os.path.abspath(out_path)) or ".", exist_ok=True)
    total_chars = 0
    with pdfplumber.open(pdf_path) as pdf, open(out_path, "w", encoding="utf-8") as f:
        n = len(pdf.pages)
        for i, page in enumerate(pdf.pages, 1):
            f.write(f"\n========== PAGE {i} ==========\n")
            text = page.extract_text() or ""
            total_chars += len(text)
            f.write(text)
        print(f"Wrote {n} pages -> {out_path}  ({total_chars} chars)")

    # Loud-failure guard: if a 5+ page PDF has < 50 chars in total, it's
    # almost certainly a scanned image PDF with no text layer. The rest
    # of the skill (caption search, bibliography extraction, equation
    # alignment) all depend on real text and will silently produce
    # garbage. Fail here with an actionable message instead.
    if n >= 5 and total_chars < 50:
        sys.stderr.write(
            "\n[ERROR] This PDF appears to have no text layer "
            f"({total_chars} chars across {n} pages).\n"
            "This skill needs a born-digital PDF or one that has been OCR'd.\n"
            "Quick check:    pdftotext input.pdf -\n"
            "If empty, run OCR first (e.g. `ocrmypdf input.pdf input-ocr.pdf`)\n"
            "and re-run this skill with the OCR'd file.\n"
        )
        sys.exit(2)
    return total_chars


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    parser.add_argument("pdf", help="path to the input PDF")
    parser.add_argument(
        "--out", required=True,
        help="path to the output .txt file (parent dir will be created)",
    )
    args = parser.parse_args()

    # Force UTF-8 stdout on Windows so prints don't crash on non-ASCII paths.
    if sys.stdout.encoding and sys.stdout.encoding.lower() != "utf-8":
        sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8")

    extract(args.pdf, args.out)


if __name__ == "__main__":
    main()