0025: The website is the document set, rendered

Status: superseded by 0033 Date: 2026-09-14 Relates to: B-068 (the publish metadata that named homepage); decisions 0013 §5.5, 0014 §2

1. The question

Cargo.toml has named https://weida.doodleshnookie.net as this workspace's homepage since B-068, and nothing was there. The owner's instruction is narrow and it decides most of this note: the site lives in this tree, and there is no public release — work continues locally.

So the question is not "what should a project website say". It is: what can a site say today without lying, where does the text come from, and what stops it from drifting away from the repository the way a hand-written landing page always does.

2. What this repository already knows about drift

Every coherence sweep this project has run found the same class of defect, and it was never in the code. The sweep of 2026-09-14 found fifteen incoherences and eleven of them were counts and claims that had become false in prose: "five frame kinds" with six on the wire, "three transports" with four implemented, IncomingMeta with "9 fields" and twelve in the struct, a mapping document still refusing a pattern that had shipped, and a README.md that said L2 "does not exist yet" with a broker in the tree. The review a few hours later found a documented rule — "a consumer pauses by restating what it has already been delivered" — that four places stated and the code could not do.

A marketing page is that defect with a bigger audience. A landing page carrying its own summary of the guarantees would be the first thing to go stale, because it is the furthest from the tests and the only page nobody re-reads while changing code.

The documents are already the product. PROTOCOL.md is normative and asserted by golden vectors; PATTERNS.md's tables name the tests that prove each row; GUARANTEES.md says what is implemented rather than what is intended; STATUS.md is regenerated from measurements. There is nothing a website could add to that set except navigation — and something that claims to be more would be less.

Two facts constrain what may appear anywhere. There is no release: no crate on any registry, no tag, no binary. And this repository's own rule for a claim is that it must be checkable; "get started in five minutes" is not, while "the quick start below is a command in this tree" is.

3. Options

	what it is	named loss
A — hand-written site beside the docs	HTML and prose written for the web, linking into the repository	a second copy of every claim, and the copy nobody re-reads. This is the drift class §2 names, adopted deliberately
B — an external generator (mdBook, Zola, Hugo)	a `book.toml` or a theme, the docs as chapters	a toolchain the gate cannot run: none of them is a Cargo dependency, so a site build would need a second install on every machine and could not be a test. mdBook would also impose its own navigation shape on a set that is not a book
C — a workspace member that renders the documents	one crate, one dependency, output to `target/site`	a Markdown dependency in the workspace, and a template that can only be as good as the documents it wraps
D — nothing yet; point the domain at the forge	zero work, honest	the `homepage` field promises a page; a redirect to a git browser is a worse answer than a rendered document set, and the repository is not public either

4. Decision

C, with the rule that makes it worth having: the site writes no prose of its own.

crates/site renders the repository's own documents and adds nothing but a shell. Navigation, a stylesheet, a per-page heading index and one notice. Every sentence a visitor reads is a sentence in docs/ or README.md, so there is no second copy to fall behind. The landing page is README.md.
A document is published, or excluded with its reason written down. The exclusions are three kinds and the distinction is the whole of the editorial policy: the loop's own bookkeeping (BACKLOG.md, NIGHTLOG.md, LOOP.md) is not documentation of a product; a named third party's requirements are theirs rather than ours to publish; and the research sheets under docs/research/ are condensed from other projects' specifications and manuals, which are not ours to serve from a page that looks like ours. The master plan is excluded on a fourth ground — it is the brief this repository is built from, normative for the loop and a plan rather than a description of what exists.

A document that is neither published nor excluded fails a test (every_document_is_published_or_excluded_on_purpose), so the site cannot quietly fall behind the directory it renders.
A link to an excluded document becomes a link into the repository, never a page that is not there. A link that resolves to nothing at all fails the build, which makes the site build a link checker for the document set — it found 58 broken links in the decision notes the first time it ran, all of them written root-relative from inside docs/decisions/.
The notice is on every page, not only on the landing page, because a direct link delivers a reader into the middle of the set: "Pre-release. weida is not published: no crate on any registry, no tag, no binary. These documents are rendered from the repository and describe work in progress." It lives in one place, weida_site::NOTICE, and a test asserts it on every rendered page.
The output is files. cargo run -p weida-site writes target/site and prints the landing page's path; every link between pages is relative, so the whole site works over file://. No server, no JavaScript, no web font, no analytics — the last three would each need a reason, and a document set has none.
Nothing deploys it. There is no publication step in this tree, no CI to run one (B-061 is blocked on a runner) and no release to announce. Serving it is one item, filed, and it needs the owner: which host serves the domain, and whether the site goes up before there is anything to install.
One dependency, and the same one rustdoc uses. pulldown-cmark, so a document rendered for the site and a doc comment rendered into the API documentation are the same Markdown dialect rather than two readings of one file. default-features = false drops the command-line parser its own binary needs; the crate is publish = false, so nothing reaches a library user's dependency graph.

5. Consequences and follow-ups

The site is in the gate. cargo test --workspace renders the whole document set four times over and fails on a dangling link or an unaccounted document, which is the property that keeps this from becoming the thing §2 warns about.
The member count moved and the README was wrong about it. Adding this crate made the count checkable again: cargo metadata lists 28 members and README.md claimed twenty-four while missing weida-broker, weida-raft and weida-py — a table that says it is "read off cargo metadata, not kept by hand" and was not. Fixed in the same pass.
B-253 is this note's implementation; B-254 is serving it, and stays blocked on the owner for the two questions above.
What the site must gain before it is public, and not before: a quick start that names a release form (there is none yet, B-068's tag is unpushed), and the overhead numbers of B-246 re-measured, because the first thing a visitor compares is bytes on the wire.
What it must never gain: a page whose claims are not checkable in this tree. If a sentence cannot live in a document that a test or a measurement backs, it does not belong on the site either.