Two \column's in a frame no longer cost a page break

[tannevaled]

\beamer@colclose is not just an \end{minipage}:

\def\beamer@colclose{\end{minipage}\hfill\end{actionenv}\ignorespaces}
                                       (beamerbaseframecomponents.sty:283)

A body scanner that meets the next \column has honoured only its first half — the
\end{minipage} — but emptied the whole macro, dropping the \end{actionenv}
that pairs with the \begin{actionenv} the column opened. Two columns then left an
environment group open, the \end{frame} that followed closed that instead of
the frame, and the next frame never began its own page. With material after
\end{columns} the pages even came out in the wrong order — the text following the
columns landed on the page before the frame's own content.

Now only the honoured half is dropped.

Why this took two goes

The first attempt at this (branch beamer-colclose, measured in #205) looked like a
regression: it fixed the synthetic cases and one real talk while breaking another,
039a234b02c5, whose first column holds an lstlisting. That talk was not a
counter-example to the fix — it was #214: the listing was reading its body from the
character buffer and eating the column. With #214 merged, the two compose.

Measured against tectonic

	before	after
beamer page error (79 talks)	40	38
exact	62	63
0029f300	5 pages	7 (reference 7)
039a234b02c5	4 pages, 1307 glyphs	4 pages, 1493 glyphs
arXiv (157 papers)	583	583

Tests

colCloseAfterEndMinipage is unit-tested — the tail kept, a closer with no leading
\end{minipage} left alone, an empty one left empty. The whole-document test needs
the real beamer class (the built-in emulation makes \column a \par and never
opens an actionenv, so it cannot show the defect); it skips with a stated reason
where the tree is absent rather than passing for the wrong one.

Closes #205.

🤖 Generated with Claude Code