normalize display math fences attached to content / nested in lists

[dykhgit]

Fix display math swallowing text after the formula

Problem

normalizeDisplayMath in lib/markdown.ts rewrites display math so that
remark-math (which requires $$ to sit on their own lines) can parse it.
Three common LLM output shapes are currently normalized incorrectly, and each
makes the text after the formula render as a giant red katex-error block
(ParseError: Can't use function '$' in math mode):

1. Single-line $$…$$ nested in a GFM list item

   - 嵌入模型几乎一律用**余弦相似度**，因为要对向量做 L2 归一化：
     $$\mathrm{sim}(q, x) = \frac{q \cdot x}{\|q\| \, \|x\|}$$
   - 归一化后相似度范围 $[-1, 1]$，再除以 $\tau$ 放大后过 softmax。…

   The split emits $$ / content / $$. Inside a list item the
   column-0 content line is a lazy continuation, which breaks remark-math's
   math-flow tokenizer: the opening $$ becomes an empty math node and the
   closing one becomes a new opener that swallows all following text.

2. Multi-line block with delimiters glued to content

   $$\frac{1}{x_{2023}^2}\le 4+4044+\frac52\sum_{k=2}^{2023}\frac1k
   <4048+\frac52\cdot\ln 2023<4048+\frac52\cdot 8=4068<4096=64^2.$$

   remark-math never finds a valid fence (delimiters must start their own
   lines), so the whole block — and everything after it — is consumed as math.

3. Indented bare $$ fences inside a list item with column-0 content:
   same lazy-continuation breakage as (1).

Fix

• Re-indent content lines together with the $$ fence when the block is
  indented, keeping it inside the list item (no lazy continuation).
• Normalize multi-line $$ blocks with glued delimiters to
  $$\n<content>\n$$.
• Detach a glued closing $$ and re-indent lazy content of indented bare
  fences.
• Fenced code, inline code, raw HTML tags and \[…\] handling are unchanged;
  a top-level block whose delimiters are already on their own lines is left
  untouched.

Tests

Adds lib/markdown.test.mjs (12 cases) using the built-in node:test runner —
no new dependencies, no lockfile changes. Run with npm test.

Verified end-to-end through the app's render pipeline
(remark-gfm + remark-math → rehype-raw → rehype-sanitize → rehype-katex):
both affected real-world answers drop from 3 katex-error spans and dozens of
stray $ to 0; a regression pass over all local session texts reports 0 new
failures. tsc --noEmit is clean.