Optimize the lesson -> GitHub Pages conversion in build_website.py:
- BS4 single parse: thread one BeautifulSoup tree through heading-id
normalisation, TOC extraction, and link rewriting instead of
reparsing the rendered HTML 3x per page. Build drops 2.97s -> 1.77s
(~40% faster) on the ~222-page site; html.parser reparsing was the
dominant cost.
- Nav O(n^2) -> O(n): build the sidebar section grouping once
(build_nav_skeleton) and resolve per-page relative URLs in localize_nav,
removing ~49k redundant relative_link/grouping calls.
- Read each .md once: cache source text on PageInfo.content and reuse it
across collect_pages, render_pages, and copy_assets (was 2-3 reads/page).
- CI: enable uv cache and cache scripts/.vendor-cache so deploys skip cold
dep resolution and vendor (Tailwind/Mermaid/fonts) downloads.
Public string APIs (derive_page_title, render_markdown, extract_toc,
rewrite_links, normalise_heading_ids) preserved as thin wrappers over the
new soup-based cores. Full suite green (92 tests), ruff/mypy clean.
Output is semantically identical to the prior build; the only diff is raw
hand-written <img/> badges serializing as <img></img> (valid HTML5,
renders identically; nothing validates generated site HTML).
8143e37d1b