Fix aligned environments breaking KaTeX: collapse math to single line

Two root causes for equations not rendering after the first few:

1. LaTeX \\ (line-break inside \begin{aligned}) was being processed by
Gradio's markdown engine as a hard line-break (<br>) before KaTeX saw
the content. Multi-line $$ ... $$ blocks may also not be recognised by
the underlying marked.js parser. Fix: collapse all whitespace (including
real newlines) inside \[...\] content to a single space before wrapping
in $$content$$. \\ is two backslash characters, not whitespace, so it
is preserved and KaTeX receives valid aligned-environment LaTeX.

2. Truncated model output (page content cut off mid-equation) left an
unclosed \[ in the text with no matching \], which appeared as raw LaTeX.
Fix: after converting all complete \[...\] pairs, strip any remaining
orphaned \[ to end-of-string.

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

app.py

@@ -116,11 +116,21 @@ def to_math_md(text):
     """Convert model output to markdown with $$/$ delimiters for gr.Markdown + KaTeX."""
     if not text:
         return ""
-    # \[...\] → $$\ncontent\n$$ (block display math)
+    # Collapse all whitespace inside \[...\] to a single line before converting to
+    # $$content$$.  This fixes two problems:
+    #   1. Markdown parsers that don't recognise multi-line $$ ... $$ blocks.
+    #   2. LaTeX \\ (line-break inside aligned) being processed by the markdown
+    #      engine as a hard line-break (<br>) before KaTeX ever sees it, which
+    #      breaks every aligned/cases/gathered environment.
+    # Note: \\ is two backslash characters — not whitespace — so it survives split().
     text = re.sub(r'\\\[(.+?)\\\]',
-                  lambda m: f'\n\n$$\n{m.group(1).strip()}\n$$\n\n',
+                  lambda m: f'\n\n$${" ".join(m.group(1).split())}$$\n\n',
                   text, flags=re.DOTALL)
-    # \(...\) → $...$ (inline math)
+    # Remove any orphaned \[ without a closing \] (model output truncated mid-equation).
+    # After the substitution above all complete pairs are gone, so any remaining \[
+    # is unclosed and would appear as raw LaTeX in the output.
+    text = re.sub(r'\\\[.*', '', text, flags=re.DOTALL)
+    # Inline math
     text = re.sub(r'\\\((.+?)\\\)', lambda m: f'${m.group(1).strip()}$', text)
     return text