Fix MathML rendering: placeholder approach + HTML escaping + \frac OCR fix

Three bugs fixed:
- Python-Markdown stopped capturing multi-line <div> blocks at blank lines.
Switch to placeholder approach: extract math → replace with ZZKEY tokens
→ run markdown → swap keys back for MathML/HTML.
- Unescaped '&' in LaTeX fallback blocks broke HTML (e.g. \begin{aligned}&...).
Use html.escape() in the <code> fallback in _to_mathml.
- OCR error \frac{1/2} (single-argument fraction) caused latex2mathml to fail.
Pre-process with re.sub to split into proper \frac{1}{2} form.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -11,6 +11,7 @@ import fitz
 import re
 import numpy as np
 import base64
+import html as html_lib
 import markdown as md_lib
 import latex2mathml.converter
 
@@ -153,45 +154,59 @@ math[display="block"] { display: block; overflow-x: auto; max-width: 100%; }
 
 def _to_mathml(latex: str, display: bool) -> str:
     """Convert a LaTeX string to MathML. Falls back to a code block on error."""
+    # Fix OCR error: \frac{n/m} (single-argument fraction) → \frac{n}{m}
+    latex = re.sub(r'\\frac\{(\d+)/(\d+)\}(?!\s*\{)', r'\\frac{\1}{\2}', latex)
     try:
         mathml = latex2mathml.converter.convert(latex)
         if display:
             mathml = re.sub(r'<math\b', '<math display="block"', mathml, count=1)
         return mathml
     except Exception:
+        escaped = html_lib.escape(latex)
         if display:
-            return f'<pre class="math-fallback"><code>{latex}</code></pre>'
-        return f'<code class="math-fallback">{latex}</code>'
+            return f'<pre class="math-fallback"><code>{escaped}</code></pre>'
+        return f'<code class="math-fallback">{escaped}</code>'
 
 def to_math_html(text: str) -> str:
     """Convert model markdown output to HTML with server-side MathML rendering.
 
-    LaTeX is converted to MathML by latex2mathml (pure Python, no JS required).
-    The MathML is embedded directly in the HTML before the markdown pass so the
-    markdown engine never touches LaTeX backslashes or delimiters.
+    Uses a placeholder approach: math is extracted and replaced with unique
+    tokens before the markdown pass, then swapped back afterwards. This avoids
+    Python-Markdown mishandling multi-line <div> blocks that contain blank lines.
     """
     if not text:
         return ""
 
-    # --- display math \[...\] → block MathML wrapped in <div> ---
-    # Insert as a proper block (blank lines around the div) so Python-Markdown
-    # treats it as a raw HTML block and leaves it untouched.
-    def display_block(m):
-        mathml = _to_mathml(m.group(1).strip(), display=True)
-        return f'\n\n<div class="math-display">{mathml}</div>\n\n'
+    blocks: dict[str, str] = {}
+    counter = [0]
 
+    def display_block(m):
+        key = f'ZZDISPLAYMATH{counter[0]}ZZ'
+        counter[0] += 1
+        blocks[key] = f'<div class="math-display">{_to_mathml(m.group(1).strip(), display=True)}</div>'
+        return f'\n\n{key}\n\n'
+
+    def inline_math(m):
+        key = f'ZZINLINEMATH{counter[0]}ZZ'
+        counter[0] += 1
+        blocks[key] = _to_mathml(m.group(1).strip(), display=False)
+        return key
+
+    # Replace display math \[...\] with placeholder tokens
     text = re.sub(r'\\\[(.+?)\\\]', display_block, text, flags=re.DOTALL)
-
-    # Remove orphaned \[ with no matching \] (truncated model output).
+    # Remove orphaned \[ with no matching \] (truncated model output)
     text = re.sub(r'\\\[.*', '', text, flags=re.DOTALL)
+    # Replace inline math \(...\) with placeholder tokens
+    text = re.sub(r'\\\((.+?)\\\)', inline_math, text)
 
-    # --- inline math \(...\) → inline MathML ---
-    text = re.sub(r'\\\((.+?)\\\)',
-                  lambda m: _to_mathml(m.group(1).strip(), display=False),
-                  text)
-
-    # --- remaining text: standard markdown (tables, bold, headings, images…) ---
+    # Run markdown on text that now contains only safe placeholder tokens
     html = md_lib.markdown(text, extensions=['tables', 'fenced_code', 'sane_lists'])
+
+    # Swap placeholders back for MathML/HTML (handle <p>KEY</p> wrapping too)
+    for key, value in blocks.items():
+        html = html.replace(f'<p>{key}</p>', value)
+        html = html.replace(key, value)
+
     return f'<div class="math-preview">{html}</div>'
 
 def embed_images(markdown, crops):