DOC-3470: Improve index.md quality with DOM preprocessing pipeline

[kemister85]

Summary

• Adds a DOM preprocessing pipeline to generate-markdown.mjs that cleans Antora HTML before dom-to-semantic-markdown conversion, fixing broken tables, about:blank anchors, admonition formatting, live demo noise, and leaked <style>/<script> elements.
• Refactors the script to use const arrow functions with a composable transform array, following tinymce/tinymce-premium coding conventions.

Changes

Transform	Before	After
stripNonContent	<style>, <script>, signup promos leaked into markdown	Stripped before conversion
rewriteAdmonitions	Broken table rows: | | text |	Clean blockquotes: > **Note:** text
rewriteLiveDemos	Tab list noise + duplicated raw HTML	Single JS code block with Example label
stripHeadingAnchors	about:blank#section-name in all headings	Clean ## Section Name
rewriteCardTables	Malformed markdown tables from layout grids	Bulleted list: **[Link](url)** — description

Test plan

• Ran against full build (1,442 pages converted, 16 skipped)
• Verified byte-identical manifest output between old and new implementations
• Verified 30 representative pages across content types (plugins, config options, migration, release notes, API docs, getting started)
• Global sweep: zero about:blank, <style>, signup-promo, kapa-widget, or Edit on CodePen leaks across all generated .md files

[kemister85]

@TheSpyder

I've been looking at the quality of our generated index.md files from V1 release, the Markdown endpoints that LLMs and AI coding tools consume when fetching TinyMCE documentation. The dom-to-semantic-markdown library does a solid job with straightforward content, but several Antora-specific HTML patterns were producing degraded output:

• Admonition blocks (Note/Warning/Tip) rendered as broken table rows instead of blockquotes
• Live demo sections produced noisy tab lists and duplicated raw HTML textarea content
• Heading anchors resolved to about:blank#section-name instead of #section-name
• Card layout tables (like on Getting Started) converted to malformed markdown
• <style>, <script>, and signup promo blocks leaked through into the markdown

The fix adds a DOM preprocessing pipeline that runs before d2m conversion — each transform is a small, named function in a composable array, so adding or reordering transforms is straightforward. The script was also refactored to follow our established conventions (const arrows, single-purpose functions, verb-first naming).

Tested against all 1,442 pages with zero regressions. Happy to walk through the changes if anyone has questions.