Closes #5483.

Source mode now highlights exactly the math syntaxes the editor is currently
reading. Three things stood between it and that, all in the same file and all
the same missing wiring.

1. The mode never read a preference

codeMirror/index.ts registers markdown-math at module load and
sourceCode.vue asked for it by name, so the source view delegated $…$ to
stex whether or not texMathDollars was on. That went unnoticed because the
preference ships enabled — but #5446 added three more math syntaxes and they
ship disabled, so the same shortcut would be wrong in the opposite direction
for them.

The mode now takes its regions from the mode spec. A fresh CodeMirror instance
reads the preferences at mount (index.vue builds the source view behind a
v-if, so every entry is a new one); a watch covers a preference changed while
the view is already open, mirroring the one #5111 added for
sourceCodeLineNumbers. The factory's defaults reproduce the old behaviour, so
a caller still passing the bare string 'markdown-math' is unaffected.

Visible change, and the intended one: with texMathDollars off, source mode
stops treating $ as a delimiter — matching what the editor has done since
#5450, and what #2002 and #5243 asked for.

2. The $ open guard had drifted

It claimed to mirror Muya's inline_math rule and pointed at
src/muya/lib/parser/rules.js — a path belonging to the engine this editor no
longer uses. It drifted further when #5449 gave the editor pandoc's three
tex_math_dollars constraints, which the guard never gained. So source mode
still opened a math span on Revenue rose from $13B to $24B. while the editor
read it as prose — the currency case #5446 used to describe the problem in the
first place.

The guard now carries the same three constraints. It still requires a closer on
the line, which is this view's own rule rather than pandoc's: CodeMirror
tokenizes per line and the multiplexer's state survives across them, so an
unclosed opener would render the rest of the file as LaTeX.

Worth knowing: the existing currency test passed throughout. The old guard
opened its span at $5 and closed it before only, which is exactly what the
assertion looked for. It now checks what is inside the spans.

3. The backslash delimiters had no regions

\(…\), \[…\], \\(…\\) and \\[…\\] were not just unhighlighted but
actively mis-highlighted, since the markdown mode reads the underscores in
\(\text{F}_\text{A}\) as emphasis — the #4121 bug in its backslash spelling.

Each pair now gets a stex region behind its own preference. The double-backslash
regions are registered first, because the multiplexer takes the first spec
matching at the cursor and the literal \( matches at the second character of
\\(. The inline forms need their own guards rather than INLINE_MATH_OPEN:
part of that one requires the character before the closer not to be a backslash,
and for \) it is one. Their content rules come from Muya's own regexes,
including the two escape rules that differ between the
extensions.

Two behaviours that look wrong and are not

• - \[TODO\] item becomes a formula once texMathSingleBackslash is on. That
  is pandoc's documented drawback — the extension "precludes escaping ( and
  [" — the editor agrees, and it is why the preference ships disabled. Pinned
  by a test so nobody "fixes" it.
• \\(x\) — a double-backslash opener closed with a single-backslash closer —
  is read as math when only the single extension is on, where the editor sees an
  escaped backslash. The multiplexer searches a regexp against the rest of the
  line, so a guard cannot see the backslash in front of the opener. The shape is
  invalid in both dialects and not worth reworking shared code for; the
  well-formed \\(x\\) is unaffected, because its closer reads as an escaped
  backslash under the single-backslash rule and leaves the span unclosed.
  Commented at the guards.

Also here: a comment pass over the same file

The mode file had grown to 42% comments, several restating the code or
recording archaeology. Trimmed to what the code cannot say, and the three
delimiter pairs folded into one region helper instead of three copies of the
same literals: 156 lines to 111, no behaviour change. It edits the file this PR
rewrites, so it belongs here rather than on its own branch.

A separate bug found while verifying this — the two backslash preferences
rendering the same label, because vue-i18n collapses \\ — is its own PR,
#5485.

4. GitHub inline math had no region either

$`e=mc^2`$ opens on $, so the dollar rule claimed the span and handed
the backticks to stex as part of the formula — the mis-reading #5457 fixed in
the editor, still on show in the source view and unmoved by texMathGfm. It
now has its own region, offered ahead of the dollar ones because its opener is
the longer of the two.

5. The delimiters are set back from the formula

$, \(, \[ and their closers carried the multiplexer's delimStyle
classes but nothing ever painted them, so a delimiter was as loud as the
formula it wrapped. They now keep the surrounding colour at reduced opacity —
fading what is inherited rather than naming a hue is what makes this work
across the three CodeMirror themes and the thirty-odd palettes behind them.

0.65 is the lowest value clearing 3:1 in every shipped theme. Measured: light
7.0, graphite 6.85, ulysses 6.72, tokyo-night-light 6.0, oxocarbon-dark 7.28,
synthwave-84 6.57, material-dark 5.43, one-dark 3.67. one-dark sets the floor —
its body text only reaches 6.57:1.

Not in scope

```math is a fenced code block rather than a delimiter pair, so it reaches
stex by a different route than a multiplexer region.

\\[…\\] with texMathDoubleBackslash off is styled as a markdown link: \\
is an escaped backslash and CodeMirror's markdown mode speculatively opens a
link on the [ that follows, closing at the ]. Not math-specific — any
unclosed [ does it — and identical on develop.

Test plan

• pnpm run lint and pnpm run typecheck
• test/e2e/source-math.spec.ts — 13 cases: the four existing $ ones
  (one strengthened), the currency regression, the backslash forms per
  preference, the unclosed opener, the escaped bracket, and both the
  live-toggle and remount paths for every preference
• source-code-line-numbers, source-mode-flush, issue-781-source-undo,
  issue-2372-source-selection-color — 18 passing together
• Both open guards checked against the case tables they mirror before being
  written in: the new $ guard agrees with fix(muya): apply pandoc's tex_math_dollars constraints to $…$ (#5446) #5449 on all ten rows, where the
  old one was wrong on four

🤖 Generated with Claude Code