Float placement bug (#1646)

* fix for #1645 (WIP)

* update some tlgs

* more tlgs

* grummel

* changelog, internal rename, ltnews41

* small editorial changes

* typo

* fix typos; added another test for rollback

* add another test and correct previous one

* overlooked one suggestion

base/changes.txt

@@ -6,6 +6,12 @@ to completeness or accuracy and it contains some references to files that are
 not part of the distribution.
 ================================================================================
 
+2025-01-29  Frank Mittelbach  <[email protected]>
+
+	* ltoutput.dtx (subsubsection{Float control}):
+	When floats are to be placed and a test fails, \@reqcolroom needs
+	to be reset to the value prior to the failed test (gh/1645)}
+
 2025-01-23  Frank Mittelbach  <[email protected]>
 
 	* ltoutput.dtx (subsection{Floats}):

base/doc/ltnews41.tex

@@ -489,6 +489,29 @@ \subsection{Make \cs{label}, \cs{index}, and \cs{glossary} truely invisible in r
 %
 \githubissue{1638}
 
+\subsection{Correct the float placement algorithm}
+
+When floats are added to the current or next page, \LaTeX{} makes
+several tests to find an area that can receive the float. One of these
+tests calculates how much space is already used on the page and how
+much additional space is needed to place the float in a particular
+area. This means that it looks not only at the height of the float but also at the
+values from \cs{intextsep} (for \texttt{h} floats) or
+\cs{textfloatsep} and \texttt{floatsep} (for \texttt{t} and \texttt{b}
+floats). The resulting space requirement is then stored in an internal
+variable and compared to the space still available on the page.
+
+If the test fails, the algorithm tries the next area. Unfortunately,
+it was reusing the value in that internal variable as the starting point
+for the next test without removing the added space for the float
+separation (\cs{intextsep}, \cs{floatsep}, or \cs{textfloatsep}). Thus
+the comparison was being made with the wrong value (i.e., to high); therefore the test
+may have incorrectly concluded that a float doesn't
+fit, even though in fact it did.
+
+This has now been corrected.
+%
+\githubissue{1645}
 
 
 \section{Documentation}