Name the contexts a change cannot survive

[phucnht]

Second slice of the roadmap's "Robust and accessible authoring" group, following #23. Measured with 26 probe documents rather than inferred, and two of the measurements changed what the work should be.

What was actually broken

Four constructs aborted compilation with an error from TeX or xcolor that never named this package:

Inside a change	Error the user saw	Same thing in \textbf	Verdict
\verb	Argument of \XC@col@rlet has an extra }	fails, but says \verb illegal in argument	inherited LaTeX limit, worse message
verbatim environment	Not allowed in LR mode	File ended while scanning use of \@xverbatim	inherited, both cryptic
spans an & tab	Missing } inserted	Argument of \check@nocorr@ has an extra }	inherited, both cryptic
\footnote	Argument of \XC@col@rlet has an extra }	works	texchanges-specific

The \textbf column is what decides fault. Three are properties of LaTeX, so the guards name them rather than pretending to fix them.

Two findings that changed the plan

The roadmap item's premise was wrong. It read "unsupported citation contexts", but \cite inside a change compiles, and so do \index, \ref, and a change within a single table cell. Nothing needed fixing there; the roadmap wording is corrected instead.

The README was stale. It still told users to avoid headings, floats and math at a larger boundary, which #23 had already made work. Left alone it would steer people away from working features. Rewritten against the probe results.

The footnote case, traced rather than assumed

Not a regression: the pre-#23 package fails identically. Root cause is \addtocontents carrying a footnote in the report line, reproduced with zero texchanges involved:

\addtocontents{x}{\protect\cmd{text\footnote{n}}}

Writing the text unexpanded does not help, and a report entry has no sensible footnote of its own, so it is diagnosed rather than fixed. The suggested workaround, putting the footnote immediately after the change, was compiled and confirmed to work.

How it is implemented

A local redefinition inside the change's own group intercepts \verb and the verbatim-like environments, so the redefinitions vanish with the group. Feasibility was proven with a standalone probe first, since these failures happen while xparse reads the argument and interception could have been impossible.

Alignment tabs are found by scanning for category code 4 using the same \tl_analysis_map_inline:nn technique \tx_report_text:n already uses. Matching on catcode rather than the character means an escaped \& in ordinary prose does not trip it, which is verified by a probe.

Tests

case_error_fixtures gains a helper asserting a fixture fails and that the log carries the expected message. Asserting only failure would pass even if the old cryptic error returned, since these documents always stopped compilation. The helper joins the log before matching, because TeX wraps at 79 columns and a line-based grep silently misses any message long enough to be useful — which it did on the first attempt.

Verification

Each guard disabled in turn makes its own fixture's assertion fail and no other; restoring passes:

• no verb guard → verb-in-change assertion fails
• no alignment scan → alignment-in-change assertion fails
• no footnote check → footnote-in-change assertion fails

Across the 26 probes from this slice and #23: 21 compile clean, 4 stop with the new diagnostics, 0 broke unexpectedly. \cite, \index, \ref and in-cell changes still compile, so the guards do not over-fire. Full suite green at 23 cases, shellcheck clean, website build green.