fix(blog): read heading ids from the rendered html

[iamtatsuki05]

WHY

Follow-up noted in #135. parseMarkdownFile computed the headings slugs from mdast, separately from the ids rehype-slug puts on the rendered headings, and the two disagreed.

For the bonsai post:

before   h2 agent--に-ai-用の前提を集める  |  .agent  に AI 用の前提を集める
after    h2 agent-に-ai-用の前提を集める   |  .agent に AI 用の前提を集める
rendered <h2 id="agent-に-ai-用の前提を集める">

Three ways it drifted:

• it joined only the direct text and inlineCode children with a space, so a heading containing inline code gained a doubled separator
• a heading whose text sits inside a link or **strong** lost that text entirely
• it slugged only h2/h3, while rehype-slug slugs every heading, so the dedupe counter could fall out of step (## Notes under # Notes is notes-1 in the html)

Nothing renders headings today — the table of contents reads the DOM — so there is no visible bug. But anything that starts using the field would link to anchors that do not exist, and it is adjacent to the legacyAnchors map added in #135, where copying a slug from here would produce a confusing build failure.

WHAT

• Collect the headings from the tree right after rehype-slug with a small collectHeadings plugin, so the ids come from the same place the html does
• Drop the separate mdast pass, along with the github-slugger and mdast type imports it needed
• Three tests covering inline code, the dedupe counter, and heading text inside a link or **strong**

VERIFICATION

• tsc, biome, vitest lib 103 / components 100, bun run build (73 pages), Playwright chromium E2E 59 — all pass
• Rendered heading ids are unchanged; only the reported headings values move
• Compared the field before and after against the same post to confirm the doubled separator is gone