Extract text and annotations from Shamela book pages

[XIAX66]

The book page reported in #386 extracts the search panel's 73 words while dropping all five book and annotation paragraphs. The adjacent page has the same failure. Both pages keep their book text in a unique .nass container; the search panel reuses the generic #content entry point.

Add a Shamela extractor that selects the known text container through the existing contentSelector pipeline. It applies only to shamela.ws or www.shamela.ws, numeric /book/<book>/<page> paths, and a unique nonempty .nass root. Missing, empty, or ambiguous roots retain automatic extraction. An explicit caller contentSelector opts out of the new extractor. Generic hidden-content recovery is unchanged, including articles rendered inside hidden overlays/dialogs.

Related to #386. This targets the verified book-page structure; it does not establish support for every Shamela layout or change cross-site dialog heuristics.

Validation on Node 24 / macOS:

• Two anonymized HTML/expected-output fixtures, with 20-word and 202-word chapters, fail before the fix and pass afterward. Book/page identifiers are placeholders; the functional host is retained for routing.
• Regression cases cover both sides of the 50-word retry threshold, nested hidden ancestors, active/closed overlays, duplicate IDs, caller selection, URL/root boundaries, and existing hidden/expandable article recovery. Poisoned-attribute cases check sanitization with optional cleanup passes disabled.
• On the downloaded reported page and its adjacent page, HTML and rendered Markdown retain 5/5 and 10/10 book/annotation paragraphs respectively under linkedom and JSDOM, with debug enabled or disabled. Normal output contains 215 and 175 words. Offline extraction uses the cached HTML.
• npm test and npm run test:jsdom: 800 tests pass in each. npm run build, npm run lint:innerhtml, npm run lint:unused, npm run size, and git diff --check pass.
• Production core/full bundles pass 34 checks using native DOM in an isolated headless browser. The compiled CLI retains all five paragraphs on the reported page.

Cloudflare Worker, happy-dom, and Node 20/22 were not directly exercised locally. Remote CI has not yet been verified.

Prepared with AI assistance; the reproduction and validation above were run locally.