LaTeXExporter._escape_latex mangles its own replacements: a backslash in a title renders as literal \{}

Start in citation_formatter.py at LaTeXExporter._escape_latex and compare it with _escape_bibtex and the _BIBTEX_ESCAPES map around line 94. Trace export_to_latex for the title path, then add coverage for a backslash-bearing title and confirm the generated LaTeX preserves the intended backslash without literal escaped braces.

Summary

LaTeXExporter._escape_latex escapes by running a sequential str.replace chain. The first rule turns \ into \textbackslash{}, and the later { → \{ and } → \} rules then re-escape the braces that rule just introduced. A title containing a backslash is emitted as \textbackslash\{\}, which LaTeX renders as the literal text \{} rather than a backslash.

Reproduction

formatter = LaTeXExporter()          # citation_formatter.py
formatter._escape_latex(r"a\b")
# -> 'a\\textbackslash\\{\\}b'       # wrong; should be 'a\\textbackslash{}b'

~ and ^ escape correctly only by accident — their rules run after the brace rules, so the {} they emit is never revisited. Reorder the list and they break too, which is why this wants a structural fix rather than a reordering.

Fix

Use a single-pass character map instead of a replace chain, exactly as _escape_bibtex in the same module already does:

_BIBTEX_ESCAPES = {...}   # citation_formatter.py, ~line 94

That helper was written this way because of this bug — its comment names _escape_latex as the existing instance:

Single-pass character map. A sequential replace() chain would re-escape the braces its own replacements introduce — the bug already present in _escape_latex.

So the fix is to give _escape_latex the same treatment, plus a test that pins a backslash-bearing title through export_to_latex.

Impact

Output corruption in LaTeX export, not a security issue: the mangled form is inert text, and it cannot break out of a field the way the BibTeX brace bug could (fixed in #5685). Reachable with any title containing a backslash.

Context

Found by AI review of #5685 and confirmed by execution. Deliberately not fixed there — #5685 is already large, and this predates it and is not made more likely by it. Sibling of #5687 and #5689, which track the other export-path gaps found in that review.