feat: a showcase page for the archive, and export --ui to write it

[khaneight]

Stacked on #74. The interface: published work, the graph, and the work in flight.

sentinel export --out ./content --flat --ui ./showcase
# → showcase/index.html + showcase/bundle.json

Two files. Copy them to a subdomain, an S3 bucket, a docs/ folder — no build
step, no dependencies, one network read (its own JSON). It works offline, under
a strict content policy, and in five years.

Why one HTML file rather than a Vite/React app

You said the host should be flexible. A build toolchain in a Rust repository is
a second dependency tree to keep alive, and this is a graph viewer over a few
hundred nodes — canvas and forty lines of Fruchterman-Reingold do it. If you'd
rather have React, say so and I'll build it; --data already emits the bundle
for anyone who wants to render it themselves.

What it shows

• the graph, nodes sized by how connected they are
• what the clone writes from — affirmed traits, evidence behind each, and
  what has been written from it
• what is in flight — unpublished, awaiting your approval, unconfirmed
  traits, concepts the wiki has named and not written
• growth — articles and links over time, from meta/progress.jsonl

Work the clone wrote is never rendered as though you wrote it. Different
colour, a count called out at the top, and on selection a note naming the claims
it was written from. The bundle carries extrapolated on every node for exactly
this, and the page branches on it in words, not only in a hue.

Keeping a front end from rotting

Nothing compiles a page like this, and a renamed field in the exporter turns a
panel blank rather than failing anything. So the page declares every field it
reads in one CONTRACT object; it uses that at runtime to say which key is
missing instead of rendering an empty panel, and tests/ui.rs parses it out of
the HTML and asserts it against a bundle from the real command.

I checked the guard bites — renaming extrapolated to extrapolatd:

the page reads `node.extrapolatd`, which the bundle does not publish.
Fields present: {"domain", "extrapolated", "inbound", ...}

Also asserted: the page is self-contained (no external URL, exactly one
fetch), --ui writes both files, the written page is byte-identical to the
one in the tree, --dry-run writes nothing, and the bundle respects the
approval gate — an unapproved article never appears in the graph.

Two defects found by opening it, not by testing it

Five nodes huddled in the middle of a 1400px canvas: the ideal edge length was a
constant tuned against one archive. Deriving it from the available area and node
count fixed that and then overshot, stringing five nodes across the whole
stage — which fills the space and is much harder to read than a compact cluster.
Bounded above as well as derived.

531 tests (+6), clippy and fmt clean.

🤖 Generated with Claude Code