Port upstream 4935: render same-line double-dollar display math

[Renakoni]

Summary

Port of marktext#4935 (issue marktext#4904, authored upstream by this repo's owner) into the vendored Muya tree. $$...$$ display math only rendered when the fences sat on their own lines; markdown copied from pandoc/GitHub-style sources showed raw $$ instead. On Android, paste-from-other-apps is a primary input path, so this lands as a paste-fidelity fix.

Changes (adapted onto our vendored tree, byte-equivalent to upstream)

• inlineRenderer/rules.ts + lexer.ts — a dedicated inlineDisplayMathRule recognizes non-empty single-line $$...$$ spans, tried before the unchanged $...$ rule. Empty $$$$, unclosed openers, and multiline fences are excluded — multiline $$ still goes through the block parser, and single-dollar math (including soft line breaks) is untouched.
• inlineRenderer/renderer/inlineMath.ts — $$ tokens render through KaTeX in display mode; the KaTeX cache key now includes the mode; preview content offsets use the real marker length instead of a hardcoded 1.
• config/index.ts + assets/styles/inlineSyntax.css — new mu-display-math class: the hidden preview renders as a full-width centered block, while the editing state (caret inside) stays inline-block so the source is editable in place.
• block/base/format.ts — format/cursor offsets for inline_math derive from token.marker.length, so selecting or toggling a $$x$$ span targets the content correctly.
• Same-line $$ stays an inline token (never promoted to math-block): source round-trips losslessly and paste behavior is preserved.

Tests: upstream's Muya unit specs ported (sameLineDisplayMath.spec.ts, plus formatToggle/pasteBlockMerge additions). Upstream's real-Chromium e2e (our vendored tree has no Muya e2e harness) is replaced by an app-side Playwright spec (mobile-display-math.spec.ts) that verifies what happy-dom cannot — the shipped CSS lays the preview out as a full-width centered block in the real editor, and $x$ stays inline.

node_modules/@muyajs/core synced per the vendor contract (the contract test passes in the app suite).

Verification

• Focused Muya specs: 3 files / 48 tests passed.
• Full Muya app gate: 212 files / 1435 tests passed. Full app unit suite: 75 files / 577 tests passed (includes the vendor-sync contract).
• pnpm typecheck and focused ESLint passed. Muya tsc --noEmit reports only the two pre-existing import.meta.glob errors in utils/prism/loadLanguage.ts (verified identical on a clean tree — zero new type errors). Muya's lint:css cannot run in the vendored tree (the monorepo-root stylelint config was never vendored); the CSS addition is byte-identical to upstream's, which passed stylelint there.
• New e2e passed. Failure path proven: with node_modules reverted to main's Muya files, the spec fails (.mu-math.mu-display-math never appears — the raw-$$ behavior); restored, it passes.

Notes

• Upstream PR is still open; divergence from any post-review upstream changes is accepted (Android product first).
• Labeled "feature" upstream; adopted here as a paste-fidelity fix per the repo owner's scope decision.

[Renakoni]

Addressed the Codex P1 (super-linear backtracking in the display-math rule) in 92c86bb.

Fix. inlineDisplayMathRule is replaced by execInlineDisplayMath, a linear-time scanner with the same semantics: content ends at the first unescaped same-line $$, a backslash always escapes the following character (the regex's greedy pair-first behavior, so an escaped closer extends the content), and a returned match always has an unescaped closer — the escape-parity group is always empty, so the lexer's isLengthEven gate is unchanged. The eslint-disable regexp/no-super-linear-backtracking suppression is gone.

Equivalence evidence. An exhaustive parity sweep over all 349,525 strings of length ≤ 9 from the alphabet {$, \, a, \n} (prefixed with $$) compares the old regex (through the parity gate) against the scanner: 0 mismatches on both the accepted set and the matched content.

Performance evidence. Reproduced the reported blow-up, then eliminated it: $$ + 131,072 backslashes + $ took the old regex 10,045ms; the scanner takes 0ms. (65,536: 2,581ms → 0ms.)

Regression tests added (spec now 15 tests): the adversarial 131,072-backslash input asserted under 500ms and never matched as display math; escaped-closer-extends-content; even-backslash-run content; escaped-only-closer rejected; dangling backslash / backslash-before-newline rejected.

Full re-verification: Muya app gate 212 files / 1440 tests, app suite 75 files / 577 tests (vendor-sync contract included), typecheck, and the display-math e2e all pass.

Note: the pre-existing single-dollar inline_math rule (and upstream's copy of both) shares the same backtracking shape; that's out of this PR's scope but worth carrying upstream together with this fix.