v0.2.3

Adds a feed kit: a site may declare one content stream, served as RSS 2.0,
Atom 1.0 and JSON Feed 1.1 at /rss.xml, /atom.xml and /feed.json.

A feed is not a search ranking signal, and this does not pretend otherwise. What
it buys is real but narrower: Google accepts an RSS or Atom file as a sitemap
and treats it as the recency signal beside the full inventory, and syndication to
readers and aggregators produces inbound links, which are.

Feeds

• WebServer::feed is omittable — a site with no stream serves no feed
  documents and emits no autodiscovery links, so existing sites are untouched.
  Calling it without frontend fails the boot rather than silently serving
  nothing.
• Entries are supplied, not derived: a blog is a dynamic_page_group, whose
  template data is built per request, so there is nothing to read at boot.
  Supply them in any order and any number — sorted newest-first, truncated to 20.
• Channel metadata (author, language, icons, copyright, self and alternate
  links) is derived from what the site already declares.
• FeedEntry is a named-field struct, so shipping a teaser feed is something you
  type (content_html: None) rather than something you forget.

Correctness

• Content URLs are rewritten, not copied: cache-busted, then absolutised, with a
  bare #fragment resolved against the post. A reader renders your entry on
  feedly.com, where every relative URL is broken. Rewriting is confined to real
  elements, so a post that shows markup in a code fence is left alone;
  srcset fails the boot rather than being half-rewritten.
• Characters outside XML 1.0's Char production are stripped, not escaped — no
  escape exists for them, and one 0x0B rejects an entire feed on a validating
  parser.
• Documents are byte-stable: dates come from the newest entry, never the
  clock, and ordering ties break on path. A redeploy that changed no post
  re-sends nothing to any subscriber.

Serving and discovery

• Feeds bypass the never-cache layer, which strips ETag/If-None-Match and
  stamps no-store — wrong for the most-polled document a site serves. They
  carry public, max-age=1800, a boot-computed strong ETag and
  Last-Modified, and answer a conditional request with a bodiless 304.
• Three rel="alternate" links on every page, ordered RSS, Atom, JSON.
• robots.txt names only /atom.xml, between the sitemap index and the chunk
  lines — Atom has a real per-entry <updated>, RSS carries publication dates
  only, and JSON Feed is not a supported sitemap format.

Also

• Every resource and its cache tier is named once at boot, logical path to
  hashed path, so a tier that failed to wire up is visible in the first screen
  of a deploy's logs.
• Pages::paths, so a feed advertising an undeclared page fails the boot rather
  than linking a 404.
• bin/, cache-buster.json, static/script/generated/ and .DS_Store are
  excluded from the Docker build context, so a host-generated manifest cannot
  reach the image and feed the JS build a stale one.

Upgrading

No consuming project needs to change anything. Frontend::build gained a
parameter and FrontendRuntime, WellKnown and TemplateData each gained a
field, so the public surface did move — but all of them are constructed by the
library and read by the layout, and no project builds one. Kept a patch bump
deliberately so existing 0.2.x pins pick the feed kit up without an edit.

The TypeScript library is untouched and stays at 0.1.1.

Full Changelog: v0.2.2...v0.2.3