fix(kit): restore whitespace eaten by \(...\) and $$ KaTeX markers

[hungnnvidia]

Summary

• Fixes the KaTeX preprocessor dropping the leading whitespace that \(...\) and $$...$$ require to match, which glued formulas to the preceding text and collapsed consecutive \( ... \) lines (reported in Inline \(...\) math swallows the preceding whitespace; the custom delimiter regexes also diverge from standard KaTeX #809).
• $...$ already captured and restored surrounding spaces; the other two delimiters now do the same.
• Adds node --test coverage for markKatex (npm test in kit/).

This does not take on the rest of #809 (switching to remark-math / supporting \[...\] / relaxing $...$ punctuation boundaries). Those are larger syntax-policy changes.

Test plan

• cd kit && npm test
• uv run --python 3.10 python -m pytest -n 1 --dist=loadfile -q ./tests/ (138 passed; unrelated to this change)
• Confirm Deep RL course page (units/en/unit2/mc-vs-td) no longer renders one stepS_{t+1}

Fixes #809 (whitespace-eating part)

[coyotte508]

🤖 Automated bot review — this review was generated by an AI agent (moon-coder). It may contain mistakes; please verify findings before acting on them.

This PR fixes a real whitespace-eating bug in markKatex (kit/preprocessors/mdsvex/index.js) by capturing the mandatory leading \n/\s before $$...$$ and \\(...\\) and re-emitting it in the replacement — exactly the pattern the $...$ branch already used — and adds the first node --test unit tests for the preprocessor. The change is minimal, correct, and well-scoped (it deliberately defers the larger remark-math/\[...\] questions from #809).

Findings

• 🟡 New test suite isn't wired into CI. .github/workflows/lint-svelte-kit.yml runs npm run lint and npm run check (lines 29–35) but never npm test, so the new kit/package.json "test" script only runs locally. Suggest adding an npm test step (with working-directory: kit) so the regression test actually guards against regressions.
• 🟡 Test imports a module with heavy side effects. markKatex.test.js imports ./index.js, whose module scope calls hljs.addPlugin(...), constructs mdsvex(...), and does a top-level await findSvelteComponentNames("./src/lib") (index.js:166–167) — a filesystem scan relative to process.cwd(). The test therefore requires npm ci in kit/ and only works when run from the kit/ directory (npm test guarantees that, but bare node --test from repo root would fail). Consider extracting markKatex/renderKatex into a small pure module (e.g. katex.js) so the unit test has no import side effects.
• 🔵 Known limitations left as-is (fine, but worth a line in the PR body): math at the very start of the content still won't match (both regexes require a preceding \n/\s, index.js:105–106), and $...$ still consumes its trailing space from the scanner's perspective, so $a$ $b$ with a single separating space only marks the first. Both pre-existing, not regressions.
• 🔵 The display regex comment says "leading \s/\n" collectively; the display marker specifically requires \n only. Trivial wording nit in the new comment block (index.js:98–101).

Correctness itself checks out: tracing the replacement flow in index.js:100–128, restoring the captured char cannot create spurious matches in the two subsequent .replace passes (markers contain no $, \, or whitespace), and the four tests cover the reported repro (one stepS_{t+1}), consecutive-line collapse, $$ paragraph boundaries, and the untouched $...$ path. The two-backslash test inputs correctly mirror what convert_rst_to_mdx/notebook conversion emits. Style (tabs, double quotes) matches the file.

Verdict: LGTM with minor suggestions — the fix is correct and consistent with the existing $...$ handling; the only meaningful gap is that CI doesn't run the new tests.