runicgateway.com/PLAN.md

runicgateway.com — design of record

Status: decisions taken by the org lead on 2026-08-19 and recorded in §5. Phase 0 is complete; Phase 1 is the next build step. Date: 2026-08-19 (revision 3) Repo: RunicGateway/runicgateway.com AI-assisted: yes — drafted by Claude, per the org's AI-usage disclosure policy.

The public website and documentation site for Runic Gateway. Two audiences: server administrators who want to understand and install the platform, and developers who want to build modules and integrations for it. A third arrives with the Android closed beta: players, who want the app.

Revision 1 asked five open questions. Revision 2 records the answers as decisions and designs against them. Revision 3 marks Phase 0 shipped and adds D13, which settles the contact address §14 N2 had left open — the domain still has no mailbox, and the org lead chose to publish the existing one rather than delay the beta.

---

Contents

1. The principle
2. Verified platform state
3. What the repositories disagree about
4. Phase 0 — fix the documentation the site will quote
5. Decisions of record
6. Runtime shape
7. Branding is bind-mounted data
8. The Android closed-beta signup
9. Legal pages
10. Information architecture
11. Visual direction
12. Accuracy machinery
13. Build phases
14. Still needed from the org lead
15. Planned, not in scope: the demo instance

---

1. The principle

The site is a presentation layer over existing project knowledge, never a competing source of technical truth.

Concretely:

• The normative contracts stay where they are and are linked, never copied: MODULE_API.md, MODULE_SYSTEM.md, link/INTEGRATION.md, link/v3.md, link/v4.md, website/TEAMS.md, website/BACKEND_DESIGN.md, installer/INSTALL.md.
• The site authors the journey — install, configure, administer, extend — which is the thing no existing document owns end to end, because the repos are organised by component and an operator is not.
• Where marketing language and technical documentation conflict, the technical documentation wins, and the marketing language changes.
• No feature is advertised that does not exist. §2 records what does; the "not built" list is as load-bearing as the rest.

A process rule, learned the hard way

An earlier draft of this report stated the platform was on protocol 3, with protocol 4 "feature-complete on edge, not cut over". That was wrong — protocol 4 is on main and released.

The cause: all nine checkouts in the workspace sat on feature branches whose local main ref had never been fetched, so git show main:<path> answered confidently from a months-old blob. Local main was 1–3 commits behind origin/main in seven of nine repos.

Rule for this repo: every version, protocol number or capability claim that reaches the website is verified against origin/<default-branch> or the Gitea API at the moment it is written, never against a local working tree. §12 makes that a build-time check rather than a promise, because a promise is exactly what failed.

---

2. Verified platform state

All values re-read from the Gitea API on 2026-08-19, after revision 1.

Versions

Component	Value	Authority
Wire protocol	4	link main:sidecar/src/main.rs:55 PROTOCOL_VERSION
Module API	1.6.0	website main:server/src/modules/version.js
Current bundle	2026.08.19 (protocol 4, generated 09:05:52Z)	installer branch bundles → current.json
uo-link sidecar	v2.0.0 (2026-08-19)	release; in bundle 2026.08.19
Plugin overlay	v1.0.0 (2026-08-19)	release; in bundle 2026.08.19
Installer	v0.1.1 (2026-08-24)	release
module-uo	v1.0.2 (2026-08-25)	release
Android app	v0.5.0 (2026-08-08), id com.runicgateway.app	release; app/build.gradle.kts
ServUO	57.4 — min version, and the only version the patch tier is verified against	bundle overlay.servuo
website	no releases — ships as container images, never tagged	Gitea releases API (empty)

Three corrections to revision 1, found while re-verifying:

• The bundle manifests live at the root of the bundles branch, not under bundles/: current.json, bundle-2026.08.19.json. §12's fact-checker must fetch the root path.
• website publishes no releases at all. The site must never print a "website version"; it refers to the platform by bundle and Module API version instead.
• The Android app targets no server of ours. core/net/BaseUrlHolder.kt, core/net/HostSelectionInterceptor.kt and core/prefs/ServerPreferences.kt mean the user enters the address of the deployment they belong to. This is load-bearing for §9.

What exists

Confirmed present on main, not inferred from design docs.

Core (game-agnostic): accounts and sessions; wiki; posts across four categories (news, five-on-friday, newsletter, screenshots); CMS pages and a page builder; navigation editor; appearance/theming; branding as data via BRAND_*; invites; moderation and appeals; content reports; append-only activity log; bot-scoring with automatic IP bans; TOTP two-factor with trusted devices and recovery codes; SSO over OAuth2/OIDC (Google, Discord, generic — link-only, never auto-provisioned); mobile bearer auth with rotated hashed refresh tokens; push through self-hosted ntfy; the module loader and the admin Modules screen; OpenAPI 3.0 at /api/docs.

Teams — merged to main (PR #161, cutover 4/6). Team core and roster; forums with announcements, discussion threads, replies, an edit window, post moderation and member-raised abuse reports; notifications across web, push and email; Discord slash commands; one voice channel per Team, granted by role.

module-uo (72 routes) — capabilities shard · atlas · market · governors · guilds · houses · champs · cliloc: shard status, economy, houses and IDOC decay, character sheets, the player-vendor marketplace, a spawn atlas built from the shard's own spawn tables, champion boards, points/loyalty leaderboards, town crier, account linking, and the staff control consoles.

The bridge — ServUO plugin → loopback TCP 127.0.0.1:7788 (newline-delimited JSON, the shard dials out) → uo-link sidecar (Rust, SQLite) → website over authenticated WebSocket + REST. The installer deploys the shard side as a protocol-checked bundle and registers a service under an unprivileged account.

What does NOT exist — and will not be advertised

• Matrix. TEAMS.md §8 is explicit research, not a roadmap item: Matrix has no channel-with-overwrites, no role object, no voice channel (voice is a MatrixRTC session needing a LiveKit SFU the homeserver does not ship), and no slash-command registration. Of the five capabilities a shared interface would name, a Matrix implementation could honestly provide two. The settled outcome was a capability contract, not an integration. The original brief for this site listed Matrix as a feature; it is not one.
• Multi-module deployments. TEAMS.md Part 1 puts this explicitly out of scope — one active module per deployment. module_id columns exist to keep it later-friendly; nothing exercises them.
• A second game module. module-rust is a paper dry-run (docs/modules/rust-dryrun.md), deliberately unimplemented — it exists to test that the contract generalises. Its absence is also the exit criterion for the Integration Kit's draft status (§5, D8).
• A finished Integration Kit. It describes itself as a draft: "finished when someone outside this project builds a working module for a new game by following it alone. That has not happened yet."
• macOS and Windows-on-ARM installer builds. Deliberately absent — the shard and sidecar must share a host, and no ServUO host is either.
• A public demo instance. Planned (§15), not built, and not linked until it is.

---

3. What the repositories disagree about

Verified against origin/main. None of these are caused by this project; all of them would be inherited and amplified by a website that quotes them. All ten are now in scope — see §5, D3.

#	Conflict	Location	Severity
1	INTEGRATION.md contradicts itself. Its event catalog fully documents the protocol-4 kinds — guild.roster, guild.leave, a subsection headed "On Protocol 4" — while its normative §2 still declares X-UOLink-Version: 3, "protocol": 3 in /health and ws.hello, and a worked client example sending "3".	docs/link/INTEGRATION.md lines 36, 58–63, 95, 118, 1015	High — this is the contract a third-party integrator implements against
2	INSTALL.md still tells operators protocol 3. Line 413 is the "Protocol version 3" value in the four-value handoff an operator pastes into Admin → Shard. The document's own troubleshooting table says a stale number is answered with 409 and "looks exactly like your shard going offline".	docs/installer/INSTALL.md lines 159, 163, 164, 413, 516, 726, 862	High — actively misconfigures a real install
3	v4.md says the cutover has not happened. "Status: Feature-complete on edge… edge → main is the v4 cutover" — but link v2.0.0 and overlay v1.0.0 are released from main and bundle 2026.08.19 pairs them at protocol 4.	docs/link/v4.md line 3	Medium
4	The org landing page prints protocol 3 in the installer's four-value handoff block.	.profile/README.md line 179	Medium
5	Stale component versions throughout INSTALL.md — bundle 2026.08.04, overlay v0.1.1, sidecar 1.1.0, including in copy-pasteable curl commands in Appendix A.	docs/installer/INSTALL.md lines 119, 512, 522–523, 657–658, 705–706, 725, 862	Medium
6	Two architecture diagrams disagree. docs/website/ARCHITECTURE.md places shardIngest.js and uoLinkClient.js inside the website backend; website/README.md places them in the module. ARCHITECTURE.md predates the module system.	docs/website/ARCHITECTURE.md	Medium
7	website-README.md has drifted from the live README — the snapshot lacks "Three ways in, and none of them is a build" (admin panel / MODULES env / by hand), which is now the primary module-install story.	docs/website/website-README.md	Low
8	docs/README.md's index is stale — no entry for TEAMS.md, ARCHITECTURE.md, TRUSTED_DEVICES_MFA.md or MODERATION_APPEALS.md.	docs/README.md	Low
9	BACKEND_DESIGN.md is titled "UOMysticmoon Website — Backend Design" although core is game-agnostic and UO is a module.	docs/website/BACKEND_DESIGN.md	Low
10	Module-uo/README.md shows Phase 4 as ⬜ though the module system shipped 2026-08-12 and the admin Modules screen exists.	Module-uo/README.md line 41	Low
11	Product name spelling. All prose is "Runic Gateway"; identifiers are RunicGateway. Settled by the org lead: keep "Runic Gateway" in prose. Recorded because the site's original brief mandated the opposite.	everywhere	Settled

Also found while re-verifying, and not a documentation conflict but a real constraint on §10:

gitea.whitlocktech.com has registration disabled. Repositories are publicly readable and /explore answers anonymously, but the sign-up page returns "Registration is disabled". Nobody outside the org can open an issue. See §14, N3.

---

4. Phase 0 — fix the documentation the site will quote

Conflicts 1–5 are protocol-accuracy bugs that exist independently of this website, and the site cannot be written honestly on top of them: "Connect a game server" has to tell an operator which number to type into Admin → Shard, and today the canonical guide says the wrong one. Conflicts 6–10 are hygiene; the org lead elected to fix them in the same pass.

Phase 0 lands before the site quotes anything. Each is a separate PR in its own repo, using the verified values from §2.

PR	Repo	Change
0.1	docs	link/INTEGRATION.md — protocol 3 → 4 in §2 and every normative example (X-UOLink-Version, /health, ws.hello, the JS client). Resolves the self-contradiction with its own protocol-4 event catalog.
0.2	docs	link/v4.md — status line: the edge → main cutover is done; released as link v2.0.0 + overlay v1.0.0 in bundle 2026.08.19.
0.3	docs	installer/INSTALL.md — protocol 3 → 4 and the stale component versions → bundle 2026.08.19 / sidecar 2.0.0 / overlay 1.0.0, including Appendix A's curl commands.
0.4	.profile	README.md — protocol 3 → 4 in the handoff block; the module-uo release link moves off v0.3.0.
0.5	docs	website/ARCHITECTURE.md — the diagram moves shardIngest.js and uoLinkClient.js into the module, matching website/README.md and the module system as shipped.
0.6	docs	website/website-README.md — re-sync the snapshot with the live README, including "Three ways in, and none of them is a build".
0.7	docs	README.md — index entries for TEAMS.md, ARCHITECTURE.md, TRUSTED_DEVICES_MFA.md, MODERATION_APPEALS.md.
0.8	docs	website/BACKEND_DESIGN.md — retitle for the game-agnostic core.
0.9	Module-uo	README.md — Phase 4 marked complete.
0.10	docs	SECURITY.md — the public vulnerability contact moves from a personal Gmail to security@runicgateway.com. Withdrawn 2026-08-19 by D13. The domain has no mailbox and the org lead elected to publish the existing address rather than wait; SECURITY.md already carries it and needs no change.

0.5–0.9 are grouped where they touch one repo, so the real PR count is smaller than the row count; docs PRs 0.5–0.8 land as one hygiene PR.

Nothing in Phase 0 changes code or a contract. Each PR corrects documentation to match what the code already does, and each cites the source of truth in its description.

Phase 0 is complete. It shipped on 2026-08-19 as four merged PRs — docs#172 (0.1–0.3), docs#173 (0.5–0.8), .profile#4 (0.4), Module-uo#19 (0.9) — plus 0.10 withdrawn as above. Every fact the site is about to quote now agrees across the repositories.

---

5. Decisions of record

Taken by the org lead (Colby Whitlock) on 2026-08-19. Recorded so they are not re-litigated.

#	Decision	Consequence
D1	Runtime: Astro + Node adapter, hybrid. Pages prerendered; a small number of real server endpoints exist. One Node container.	§6. Makes both the bind-mounted branding and the signup form possible without a second service.
D2	Beta signup: our own form → our own store → CSV into Play. Beta purpose only, no announce list.	§8. Granular consent is unnecessary; the policy states one purpose.
D3	All ten documentation conflicts are fixed, not just the operator-facing five.	§4.
D4	Real web screenshots, captured from the local review stack, not placeholders.	§13 phase 6. Needs seeded, presentable demo content.
D5	Claude drafts /privacy and /terms from what the code actually collects; the org lead reviews before ship.	§9.
D6	Ship the image and compose file; the org lead deploys. DNS and TLS terminate at their existing reverse proxy.	§13 phase 9. Amended by D54 (2026-08-25): a merge to main now builds, publishes and deploys. What D6 protected is held by needs: build and by every check having run on the pull request.
D7	The site sends no email at all. No SMTP, no notifications, no mailbox behind the domain yet.	§8 designs the signup so it works anyway — see "The opt-in link removes the need for email". A contact address is still required; D13 supplies it.
D8	Understated honesty. The site reads as finished; factual badges appear only where they save a reader wasted effort. The Integration Kit stays marked draft until a second module is successfully built against it.	§11, §10. A status with an exit criterion, not a mood.
D9	No analytics. No tracking scripts, no third-party requests, no cookie banner.	Reverse-proxy access logs are the only traffic data.
D10	Support routes to Discord and Gitea issues. Invite: https://discord.gg/t2Jav8yT4g.	§10. Discord is the front door; Gitea takes issues from anyone with an account, which today is nobody outside the org — §14, N3.
D11	Keep the existing emblem. runic-emblem.png is the mark on both the website and the Android launcher icon; the site adopts it rather than drawing a new one.	§11. Revision 1's "draw a new geometric mark" recommendation is withdrawn.
D12	A public demo instance is planned but out of scope today — a Proxmox VM running the full stack including ServUO, with restricted settings and an hourly automatic reset.	§15. The IA reserves the slot so it lands later without a restructure.
D13	Publish the existing address. No mailbox has been created at the domain and the org lead elected not to wait for one: whitlocktech@gmail.com is the published contact wherever the site needs one — /privacy, /terms, /community, the Play listing — and docs/SECURITY.md keeps the address it already carries.	Taken 2026-08-19, superseding §14 N2 as a blocker. The address lives only in brand.json (§7), never in prose, so moving to privacy@/security@ later is a file edit and a restart. Phase 0's PR 0.10 is withdrawn, not deferred.

Decisions after D13 are recorded where they were taken, in the section describing the phase that raised them, rather than appended here — a decision is only re-litigated when its reasoning is somewhere other than the thing it decided. The count of record is fifty-seven:

#	Where	What it settled
D14–D16	§7, "How phase 2 actually built it"	The branding pipeline: one raster in, brand text applied at boot, the mark is the real emblem
D17–D19	§10, "How phase 3 built the homepage"	The data-path diagram, all five groups on the homepage, the emblem-led hero
D20–D25	§10, "How phase 4 built the marketing pages"	/features/ as the same list with detail, /architecture/ as reasons not reference, the absences as data, the two absorbed scope items, needsModule, the demo deep links
D26–D29	§8, "How phase 5 built the app and the beta"	The screenshot slot reserved for phase 9, the demo as the tester target, /beta handling its own POST, equal billing for the APK and the beta
D30–D33	§9, "How phase 6 built the legal pages"	One logging hop and no edge provider, eighteen or older, no governing-law clause, the Data Safety notes as a generated document
D34–D37	§10, "How phase 7 built the documentation journey"	One PR for all twenty pages, a self-contained install quickstart with a drift check, every admin screen walked before it was described, a thirteenth Administration page for content
D38–D41	§10, "How phase 8 built the builder and reference docs"	One PR for all twenty pages again, Reference enumerates names and checks every one of them, the docs section links to the drawn diagrams rather than importing them, plannedSidebar becomes a checked invariant
D42–D46	§10, "How phase 9 took the screenshots"	The full rig behind the imagery, a neutral demo brand, the captures beside the claims, a committed and checked capture pipeline, the world dressed in the plugin repo's scaffolding
D47–D50	§6, "How phase 10 polished it"	Search reaches the marketing pages, the CSP is a real response header from the container, robots.txt allows everything and names the sitemap, two blocks of structured data and no more
D51–D53	§6, "How phase 11 validated it"	The chrome and the head follow the brand mount while the consent sentence does not, the documentation half gets phase 10's skip-link fix, and no twelfth check
D54–D57	§6, "How phase 12 delivered it"	A merge deploys (amending D6), the proxy is documented by its requirements rather than by an example, the container trusts the forwarded headers with nothing to configure, and the operator note is its own file while the two policy files are pointers
D58	§6, "How phase 12 delivered it"	node_modules ships in three layers because Cloudflare refuses a request body over 100 MB, and the workflow counts layers before it pushes

---

6. Runtime shape

Astro with the Node adapter, output: 'server' with per-page prerender = true. Every marketing and documentation page is prerendered HTML at build time; a handful of endpoints under /api/ and the /brand/* asset route are the only things that execute per request.

Why not fully static: two requirements need a process on the box. The beta signup must accept a POST and write it somewhere (§8), and branding must be overridable by dropping a file into a bind mount without rebuilding the image (§7) — which means the bytes cannot be fingerprinted into the build output.

Amended 2026-08-24 by D28: the signup is /beta itself rather than a POST /api/beta-signup endpoint. The dynamic surface is still two routes and the reasoning above is unchanged; see §8.

Why not a separate API service: one container is one thing to deploy, one thing to patch, and one log to read. The dynamic surface is three endpoints.

┌──────────────────────────── runicgateway.com container ────────────────────────────┐
│                                                                                    │
│   Astro (Node adapter)                                                             │
│     ├── prerendered pages ......... marketing, docs, legal — plain HTML            │
│     ├── GET  /brand/*  ............ reads the bind mount, falls back to defaults   │
│     └── GET + POST /beta/ ......... renders the form; writes SQLite on the mount   │
│                                                                                    │
│   /app/brand-default ..... baked into the image (stock logo, tokens, brand.json)   │
└─────────┬──────────────────────────────────────────────┬───────────────────────────┘
          │ bind mount                                   │ bind mount
      ./brand  →  /app/brand                        ./data  →  /app/data
   logo, favicon, theme.css, brand.json           beta.sqlite, exports/

Stack: Astro for the marketing pages, Starlight for /docs (sidebar, breadcrumbs, previous/next, automatic table of contents, offline full-text search). Vite underneath, so it stays inside the org's existing tooling family. Node 22 LTS.

Security posture, matching the rest of the org:

• A strict CSP with no external origins. Self-hosted fonts, no CDN, no analytics (D9), so default-src 'self' holds with no exceptions to argue about.
• The only writing endpoint is the signup, and it is rate-limited (§8).
• No authenticated surface exists on the site at all. The CSV export is a CLI run against the bind mount, not an HTTP route — see §8.

How phase 10 polished it

Four decisions, D47–D50, taken 2026-08-25. Three of them were straightforward; the fourth turned into the phase's real work, because the thing that was supposed to be a configuration flag was broken in a dependency and broken silently.

D47 — search reaches the marketing pages, and the marketing header gets a box. The documentation had search from phase 1: Starlight builds a Pagefind index at the end of every build. The marketing pages were outside it twice over — not indexed, so a reader searching "Teams" in the docs found the architecture page and never the feature page; and with no box, so a reader who arrived on the homepage had a four-item nav and no way to ask a question. Base.astro now marks its <main> as a Pagefind body, which puts all ten in the index the docs already query, and a Search.astro in the header opens the same index in a <dialog>.

Three things about the build are worth keeping. Nothing is fetched until the dialog is opened — Pagefind's UI bundle is 120 kB before the index and the WASM, and these pages otherwise ship almost no JavaScript, so the button is inert markup and the first open injects the script. <dialog> rather than a hand-built overlay, because the browser supplies the focus trap, the inert background, Escape-to-close and the top layer, and every one of those is something an accessibility pass would otherwise find missing. And the index needed an explicit title: Pagefind titles a result from the first <h1>, and these pages have editorial ones — /app/'s is "The app for a deployment you already use", /terms/'s is "Short, and only about what we run". Correct on the page under an eyebrow that names the section; unscannable as four rows in a result list, which is exactly what the first walk of the finished search produced. data-pagefind-meta now carries the page's short name, the one already in the nav and the browser tab.

Two things about styling somebody else's widget. Pagefind's UI takes a resetStyles option; setting it to false — on the reasoning that the site's own type and colour should show through — is wrong, because that reset is what styles Pagefind's own input and buttons. Without it they fall back to user-agent defaults, which on this ground meant black text typed into a dark field and a Clear button with an outset border. The palette is bound through Pagefind's custom properties instead. And the match highlight needed one extra class in the selector: the reset declares .pagefind-ui--reset mark { all: revert }, same specificity as a plain descendant rule and injected after our stylesheet, so it won on order and put the user-agent yellow back on every result.

It also closed a note phase 2 left here. Pagefind indexes at build time, so the boot rewrite (§7, D15) reached the pages and not the search results: a site renamed through the mount would answer a search for its own name with the stock one. applyBrand.mjs now re-indexes after a rewrite — only when it actually rewrote something, so a stock deployment still pays nothing.

D48 — the CSP is a real response header, sent by the container. The alternatives were a <meta http-equiv>, which is what Astro emits by default and which silently ignores frame-ancestors — the one directive that stops the site being framed — and writing the headers into an operator's reverse-proxy configuration, which puts the strictest promise in §6 outside the artifact this repository builds and tests. Neither is good enough for a security boundary, so the Node adapter's staticHeaders is on: the build writes one policy per prerendered route into dist/_headers.json and the server sends it.

Three things fought this, and each is the same shape: correct build, broken page, no error.

1. Astro does not hash <script is:inline>. It hashes what it processes; an inline script is the author's own text, which it never parses. Starlight ships six per documentation page — the theme provider, the theme-picker sync, the mobile menu, the sidebar scroll restore. The first build with CSP enabled produced a strict, correct header and a documentation site whose theme switch and mobile sidebar did nothing, with the explanation only in a console. 'unsafe-inline' would have fixed all six and given up the single directive CSP exists to enforce, so instead the hashes are enumerated in a generated src/config/cspHashes.mjs and scripts/checkCsp.mjs verifies, per page, that every inline block is covered by that page's own policy. A Starlight upgrade that edits one byte turns the build red; npm run csp:hashes re-harvests it.

2. Expressive Code cannot be hashed at all. Around 3,700 inline style attributes across the documentation carry every syntax colour, and CSP hashes cover <style> elements, never attributes — Astro's own documentation records Shiki as incompatible with CSP for this reason. The policy therefore carries style-src-attr 'unsafe-inline', scoped to that directive: a style attribute cannot execute script, so script-src is untouched. The marketing pages emit none.

3. @astrojs/node served the wrong page's policy. Its per-request lookup is headersMap.find((h) => h.pathname.includes(baselessPathname)) — a substring test taking the first match. /modules/ was served the policy built for /docs/modules/building-a-module; /architecture/ got a docs page's; and /, a substring of every path in the file, got whichever record came first, which was /404. Since each policy is a list of per-page hashes, the browser refused each page's own stylesheet: /modules/ and /architecture/ were rendering unstyled, and the homepage looked perfect only because it happened to share a hash with the 404 page. scripts/serve.mjs — a thin wrapper npm start now runs instead of the adapter's entry — keeps the same _headers.json and matches by equality. It is small on purpose so it can be deleted whole when upstream is fixed, and it is where the non-CSP security headers live too.

   This is why test/headers.test.mjs exists. Every other check in this repository reads dist/, and every file on disk was right — the bytes on the wire were not. It starts the server and reads the responses, and reverting the wrapper to the substring lookup fails it.

D49 — robots.txt allows everything and names the sitemap. The sitemap has covered all fifty URLs since phase 1 (Starlight bundles @astrojs/sitemap) and nothing pointed at it; a crawler finds one either from this file or from a search console, and D9's posture extends to not having an account with anyone. Nothing is disallowed: there is no authenticated surface (§6), /brand/* is derived images with no text, and /beta/ is a page a person is meant to find. The 404 is kept out of the search index instead, with data-pagefind-ignore, which is the right layer for it.

D50 — two blocks of structured data, and no more. Organization so the project's name resolves to an entity rather than to whichever page ranks, and SoftwareApplication because what the site describes is software someone installs. No ratings, no counts, no invented aggregateRating — §11's understated honesty applies to markup a reader never sees, and inventing a rating is what gets structured data ignored. Breadcrumb and Article markup on the forty documentation pages was rejected: Starlight already renders breadcrumbs a reader can see, and it would be forty more places for a fact to go stale. Every value is read from brand.json or platform.json, so checkFacts.mjs already guards them.

It is a <script type="application/ld+json">, which is a data block: no browser executes it and no CSP hash covers it. Both checkCsp.mjs and applyBrand.mjs had to be taught that explicitly — the first would have demanded a hash for text that changes whenever a fact does, and the second would have refused to rewrite the homepage at all, which is §7 failing on the page that matters most.

What the walk found. Ten marketing pages and a documentation sample, at 390, 768 and 1280 in real Chrome. No horizontal overflow at any width, on any page — the responsive work of phases 3 and 4 held, including with a search button added to the header. The CSP violations above. The consent checkbox on /beta measured 17×17 against WCAG 2.2 SC 2.5.8's 24px minimum, and is now 24 — the one control on the site a person must hit precisely, on the page a phone is most likely to arrive at. And following the skip link moved the scroll but not the focus, because a <main> is not focusable; Chrome papers over that and not every browser does, so it now carries tabindex="-1".

checkA11y.mjs is the eleventh check, and the eighth in CI. Seven structural rules over every built page, ours and Starlight's forty. Structural on purpose: a static check cannot measure contrast on a rendered page or find a focus trap, and one that pretended to would be trusted for things it cannot see. Its own first run reported every marketing page as having two <main> landmarks — this repository comments its markup heavily, and one of those comments quotes the tag it is explaining, so comments are stripped before anything is counted. It was then verified by breaking each of its rules in turn.

How phase 11 validated it

Three decisions, D51–D53, taken 2026-08-25. The phase's list in §13 is a list of things to run, and running them was the easy half: npm run verify was green on main before this phase started and is green now. The half that mattered was the part no script does — a real browser, at three widths, across every page, and a signup and a brand mount walked end to end against the built server.

That found three things, two of which are fixed here. All three are recorded, including the one that was left, because a defect nobody wrote down is one the next phase re-discovers.

What was run. Fourteen steps: astro check, the production build, all eleven checks (sidebar, screens, tokens, brand, data safety, links, facts, quickstart, reference, a11y, CSP) and both test suites — 36 tests offline and 5 against a running server. Then, outside CI: fifty pages at 390, 768 and 1280 in the installed Chrome; the signup's whole decision path against a scratch store; the export CLI including a deletion; a full brand mount, applied and restarted; and every off-site link the built site publishes.

The browser walk came back clean on everything phase 10 had claimed. No horizontal overflow at any width on any page. No CSP violation anywhere — the wrapper in scripts/serve.mjs holds for all fifty policies. No failed request, no image without an alt, no image that failed to decode, and no console error other than the 404 page's own 404. Search opens, reaches both chromes, and closes on Escape at both widths; the documentation theme switcher and mobile sidebar work, which is the thing a wrong CSP breaks first and silently; twelve tab stops on the homepage all draw a focus ring and follow the visual order.

D51 — the chrome and the head follow the mount; the consent sentence does not. The brand walk mounted a complete brand.json — a different site name, tagline, contact address, Discord invite, Gitea org, demo URL and Play opt-in URL — restarted, and asked every page whether any stock string survived. Forty-nine came back clean. /beta did not.

The reason is structural and was written down in brand.mjs before it was true: applyBrand.mjs rewrites files in dist/client, and /beta renders per request, so its HTML never exists as a file to rewrite. liveBrand() was added in phase 5 for exactly this, and /beta used it — for betaOptInUrl, and nothing else. Everything around the form came from the shared chrome, and the shared chrome was baked: the page's <title>, its og:site_name, og:title and og:image:alt, the header lockup, and the footer's Source and Discord links. On a mounted deployment the one page that asks a person for their address under a stated identity was the one page still stating the wrong one.

Both accessors were right for a page that renders one way. The chrome is neither, so renderBrand(Astro) picks by Astro.isPrerendered — the stock value where the boot rewrite will reach it, the mount where it will not — and Base.astro, Header.astro and Footer.astro call it. Doing it by discriminator rather than by calling liveBrand() everywhere matters: unconditional live reads would also change the forty-nine, where a build machine that happened to have a mount would bake mounted text into HTML the boot rewrite then has nothing to replace.

CONSENT_TEXT is deliberately excluded. It names the operator of the list inside a sentence a person agrees to, and it is stored verbatim in their row — so making it follow a mounted name would change the recorded text of a consent already given. It stays a constant and moves only with a consent-version bump, which test/legal.test.mjs already enforces.

Two strings on /beta are knowingly left as they are: "A Runic Gateway deployment to connect to" and "If you already run a Runic Gateway deployment". Both name the platform the app connects to rather than the operator of this site, which is the one thing on the page a rebrand does not change.

D52 — the documentation half gets phase 10's skip-link fix. Phase 10 found that following the skip link moved the viewport but not the keyboard focus, because a <main> is not focusable, and fixed it on the marketing chrome. The walk found the identical defect standing on the other forty pages: Starlight's skip link targets the page <h1>, and an <h1> is no more focusable than a <main>. A PageTitle override adds tabindex="-1" and suppresses the ring on an element that is reachable by exactly one deliberate route and is never in the tab sequence. It is Starlight's own implementation with one attribute added, because the override mechanism replaces a component rather than decorating it; the drift risk that creates is named in the file.

The heading anchor links were checked and left. Starlight's ¶ links measure 23.98×34.8 beside an h2 and 19.86×28.8 beside an h3 at 390px, under WCAG 2.2 SC 2.5.8's 24px. They are exempt under that criterion's Equivalent clause: the mobile table of contents on the same page offers a 388×34.5 link to every one of the same anchors. Recorded rather than fixed, so the next walk does not re-raise it.

D53 — no twelfth check. Two of the phase's rigs were good enough to be tempting. An external link sweep found 73 of 74 destinations alive (the 74th, gnu.org's licence text, is unreachable from this network rather than gone), and a brand-mount walk would have caught D51 and would guard it. Both were left as throwaway scripts. A link check makes the build depend on other people's uptime, and a browser walk needs Chrome on the runner — CI would gain two ways to be red for reasons that are not about this repository. The eleven checks stand, and what the rigs found is written down here instead.

What the signup walk proved. Every branch of the decision path, each against the store rather than against the page it renders: an address is added and its consent text stored verbatim; the same address again is a duplicate and not a second row; a filled honeypot stores nothing and is indistinguishable from success; a malformed address and an unticked consent box are both refused on the server with the browser's validation disabled; an unsigned form token stores nothing; and a burst is stopped at perHour with a message that says the count is attempts rather than signups. The minSeconds gate is real enough to be worth knowing about — the first pass of the walk recorded four too-fast rows and nothing else, because a script fills a form faster than a person can.

The export CLI writes both files Play needs, marks the rows exported, refuses to re-export without --all, and remove overwrites the address, the IP hash and the user agent rather than flagging the row — which is what /privacy promises.

And the mount itself. The rewrite reached 51 files and re-indexed all 50 pages for search, so results agree with the pages. /brand/* served the mounted theme.css and fell back per key to the default wordmark.svg. Three path-traversal shapes were refused. The demo slot opened and twelve per-capability deep links pointed into the mounted demo. And an opt-in URL pasted into the mounted file reached the confirmation screen on the next request, with no restart — which is the one part of §7 that has to be true on the day the closed test opens.

How phase 12 delivered it

The last phase, and the one that turns a repository into a deployment: a two-stage Dockerfile, a pull-only docker-compose.yml carrying both bind mounts, .env.example, the publishing workflow, CONTRIBUTING.md with the AI-disclosure requirement, the community-health files this repository was the only one in the organisation to lack, and DEPLOY.md. Four decisions, D54–D57, taking the count of record to fifty-seven — and a fifth, D58, added when the merge that shipped the phase could not publish its own image.

It also found a defect that would have made the closed beta impossible, and it is the only phase that could have found it. Everything before this ran the site the way a developer runs it: one process, plain HTTP, an origin the browser and the server agree about by construction. The site does not run that way. Standing a real container behind a real proxy is a different question, and it had a different answer.

D54 — a merge deploys, which amends D6

D6 said "ship the image and the compose file; the org lead deploys", and it was written before there was a host to deploy to. The org lead amended it on 2026-08-25: .gitea/workflows/build-image.yml builds, pushes runicgateway-site:latest and :sha-<7>, and then rolls the container over on a runner labelled rgcom on the site's own host, out of /opt/runicgateway.com. There is no release step and no promotion, so a merge is a publication.

What D6 was protecting is held by something else now. Every one of the eleven checks and both test suites have already run on the pull request; the deploy job is needs: build, so a failed build never reaches the host at all; and the job then waits for the container's own healthcheck rather than for up -d to return, because npm start runs the brand rewrite before the server starts and "running" therefore arrives well before "serving". A pinned IMAGE_TAG survives an automatic deploy — Compose recreates on the pinned tag — because a pin is a decision and a merge should not quietly undo it.

The runner registers with the label rgcom:host. The :host suffix is what makes jobs run on the machine rather than inside a job container; without it the deploy fails on the cd, having no Docker socket and no compose directory.

D55 — a generic proxy, documented by its requirements rather than by an example

The site runs on its own host, behind whatever reverse proxy the org lead puts there. So DEPLOY.md does not carry a worked Caddyfile or nginx block that would be wrong for three readers out of four; it states the four things the proxy must do, and the container binds to 127.0.0.1 by default so that the safe configuration is the default one.

Two of the four are worth repeating here because they are silent when wrong:

• X-Forwarded-For is load-bearing. The signup rate-limits per client from the first entry of that header, falling back to the connection's peer address. Behind a proxy that does not set it, the peer address is the proxy — so every visitor on earth shares one bucket and the third signup of any hour closes the form for everybody. It fails toward refusing signups rather than toward accepting abuse, which is the right direction and still a broken page.
• A second Content-Security-Policy from the proxy breaks every page. Browsers enforce the intersection of duplicate policies, and this site's is a list of per-page hashes, so a generic policy added at the proxy forbids the page's own stylesheet. The container already sends CSP, X-Content-Type-Options, Referrer-Policy, X-Frame-Options and Permissions-Policy (D48), so the proxy's job is to add none of them. The same paragraph names the three Cloudflare features that rewrite HTML — Rocket Loader, HTML minification and email obfuscation — because each of them breaks the hashes in exactly the same way, from a dashboard rather than from a config file.

D56 — the container trusts the forwarded headers, with nothing to configure

The defect. @astrojs/node builds the URL of every request from the connection and the Host header alone. In astro/app/node's createRequestFromNodeRequest:

const isEncrypted = "encrypted" in req.socket && req.socket.encrypted;
const protocol = isEncrypted ? "https" : "http";

x-forwarded-proto is never consulted on that path — and security.allowedDomains, which sounds like the answer, is not: on this code path it gates only whether Astro.clientAddress may come from x-forwarded-for. Behind a proxy that terminates TLS, the browser sends Origin: https://runicgateway.com and the container computes http://runicgateway.com, because its own socket is plaintext. Astro's CSRF middleware then compares the two for equality:

const isSameOrigin = request.headers.get("origin") === url.origin;

So every beta signup, from every visitor, is answered 403 Cross-site POST form submissions are forbidden — on the one route that accepts a POST, on the site whose nearest real deadline is a closed test that cannot start without it. No proxy configuration fixes it; a proxy cannot make the container's socket encrypted. Fifty pages look perfectly healthy while the form silently refuses everyone.

scripts/serve.mjs already exists to wrap this adapter's mistakes (D48), so the fix went there: x-forwarded-proto: https marks the socket encrypted, and x-forwarded-host replaces Host for the proxies that rewrite it to the upstream address instead of passing it through.

Both are trusted unconditionally, with no flag to set — the org lead's call, and the right one. The image is meant to be deployed and work; an operator who has to discover a TRUST_PROXY variable to make the signup work is an operator who ships a dead form, and the broken configuration would be the default. It costs nothing: a cross-site form submission cannot make a victim's browser send x-forwarded-proto, so the CSRF check is exactly as strong as it was, and the site has no cookie, session or credential to protect in the first place. Two assertions in test/headers.test.mjs hold both halves — a proxy-shaped POST is accepted, and a genuinely cross-origin one is still refused.

D57 — the operator note is its own file, and two policies are pointers

DEPLOY.md rather than a README section: the README is for someone working on the site and was already long, and deployment is a different task for a different sitting — it is also the file that gets opened on the host. It carries what the site needs, first deploy, the proxy, DNS and TLS, branding without a rebuild, the tester-list CLI, updating, rolling back, backups, and a symptoms table.

This repository was also the only one of the ten with no CONTRIBUTING.md, no CODE_OF_CONDUCT.md, no SECURITY.md and no issue or pull-request templates, so phase 12 added them. The two policy files are pointers to the organisation's copies in docs, not copies — because a copy would hard-code the contact address in a tenth place, and D13's whole promise is that the address lives only in brand.json and moves for the cost of a file copy. checkFacts.mjs scans src/ and scripts/; these files honour the same rule voluntarily, and say so, so that the next person does not "fix" the missing address.

SECURITY.md is not only a pointer, though. It names what is actually worth reporting here — the signup's store, rate limit and signed form token; path traversal out of the branding mount; and a page served no policy or another page's — and states plainly that there is no authenticated surface to attack.

No twelfth check. D53 was re-tested against this phase's new drift risk — the environment variables the code reads, .env.example declares and the README tabulates — and the org lead held the line. Eleven checks stand.

D58 — the image ships in three layers, because of the registry rather than the site

The merge that landed phase 12 could not publish the image it had just built. docker push answered 413 Payload Too Large on one blob and stopped; nothing reached the registry, and needs: build meant nothing reached the host either. The site was merged and undeployed, and the log said only that a digest was too large.

The limit is Cloudflare's, not Gitea's. gitea.whitlocktech.com is proxied, and Cloudflare refuses a request body over 100 MB on every plan below Enterprise — a plan limit, not a setting. docker push uploads each layer as a single monolithic PUT, so the ceiling applies per layer, and the rejection happens at the edge with Gitea never seeing the request. It is invisible from the Gitea side and unfixable from it.

Measured on the merge commit, one layer was over and only just: COPY node_modules at 108.8 MB compressed, against a 188.6 MB image whose next largest layer was the 47.6 MB Node base. Two packages account for it, @pagefind (the search binaries) and @img (sharp's libvips), and both are needed at run time — the boot rewrite re-indexes the site and re-derives the brand images — so what could move was where they land, not whether they ship.

The build stage now moves those two aside after npm prune, and the runtime stage copies them as their own layers: 46.8 + 50.5 + 11.5 MB in place of 108.8, largest layer 50.5, and the image exactly the same total size, because the same bytes are simply divided differently. Moving rather than copying twice is what keeps the three disjoint — whatever remains in node_modules is the remainder by construction, so a dependency added later needs no maintenance here.

The workflow now counts layers before it pushes, because a split is a margin and not a guarantee: docker save, re-compress anything over 8 MB the way the push would, and fail at 90 MB — not 100, since the blob is not the only thing in the request — naming the layer and what would have happened. It was tested in both directions, against the fixed image and the broken one, and the number it reports for the broken layer (108 MB) agrees with what the registry recorded.

This is a workflow step and not a twelfth check script: it needs a built image rather than a source tree, which is the one thing the eleven never have. D53 holds.

---

7. Branding is bind-mounted data

The requirement: swapping a logo or recolouring the site is a file copy and a container restart — never a rebuild, never an image push. This mirrors the product's own posture (BRAND_* env, appearance/theming in the admin panel), so someone who has themed a Runic Gateway deployment already knows how to theme this site.

The mechanism

Two directories. /app/brand-default is baked into the image and always complete. /app/brand is the bind mount and may be empty, partial, or full. Every asset resolves against the mount first and the defaults second, per file — so overriding only theme.css leaves every logo stock, and an empty mount produces the stock site exactly.

GET /brand/* serves them at stable, unhashed URLs with an ETag and a short cache TTL. These files are deliberately not imported through Vite, because Vite would fingerprint the filename into the build and the mount could never replace them.

File	Purpose
logo.png / logo.svg	The emblem — header mark, hero
wordmark.svg	Optional horizontal lockup (emblem + "Runic Gateway"); falls back to emblem + type
favicon.ico, icon-192.png, icon-512.png	Browser and install icons
og-image.png	Link preview card
theme.css	Custom-property overrides only — appended last, so it wins
brand.json	Text and links: site name, tagline, Discord invite, Gitea org URL, contact address

brand.json exists so that renaming the product, changing the Discord invite or adding a contact address does not require a rebuild either — the same class of change as swapping a logo.

How phase 2 actually built it

Three decisions taken during the build (org lead, 2026-08-20). They refine the mechanism above rather than change what it promises.

D14 — one raster in, every size out. Only logo.png, wordmark.svg, og-image.png, theme.css and brand.json are baked into brand-default/. Every other image in the table above — all the logo sizes, both install icons, the apple-touch icon, the favicons and the .ico — is derived at request time from whichever logo.png is in force, cached in memory, and limited to an allowlist of sizes. Precomputing them would have meant an operator producing fifteen files to change a mark, and the realistic outcome of that is a deployment with a new header and the old favicon. "A file copy" now means one file.

D15 — brand text is applied at boot, not at render. §6 prerenders every page, so a value read at build time is baked into HTML the mount cannot reach; §7 promises otherwise. npm start runs scripts/applyBrand.mjs before the server opens a socket, rewriting the built HTML from what was baked to what the mount says. Every page stays prerendered, Pagefind still has static HTML to index, and the documentation is covered by the same pass as the marketing pages. The alternatives — server -rendering the brand-bearing pages, which is the whole site because of the footer, or accepting build-time text — were rejected. The script rewrites from a record of what it last applied rather than from the defaults, because the naive version works exactly once and then silently ignores every later edit.

D16 — the mark is the real emblem (D11 carried through). The header shows runic-emblem.png, not phase 1's placeholder glyph, so the site, the product and the Android launcher icon are one mark. The cost, accepted: it is raster art, so theme.css cannot recolour it — changing the mark means replacing logo.png.

The rule that keeps the promise true

Every colour, radius, shadow and font in the site's stylesheet is a CSS custom property defined in one file. theme.css in the mount only ever redefines those properties. A scripts/checkTokens.mjs fails the build if a colour literal appears anywhere outside the token definition file.

Without that check, "one CSS file changes the appearance" decays into "one CSS file changes most of the appearance, and then there is a hardcoded #0e1318 in the footer". The check is the mechanism; diligence is not.

scripts/checkBrand.mjs is the second half of it, added in phase 2: it fails the build if brand-default/ is incomplete, if any /brand/* URL in the source would 404 against the route's own allowlist, or if a brand string is short enough that replacing it blindly at boot could corrupt a page.

The mounted stylesheet wins by cascade layer, not by link order. tokens.css is wrapped in @layer tokens and theme.css is unlayered, so the mount takes precedence wherever the browser encounters it. The first attempt relied on theme.css being linked last, and it did not work: Astro emits its own stylesheet after the head markup, so the site's tokens landed after the operator's and every override was silently a no-op.

Token names deliberately match website/client/src/styles/theme.css where the concepts line up (--bg, --panel-a, --accent, --ink, --line, --radius-card, …), so a theme written for one is legible in the other.

---

8. The Android closed-beta signup

What Google Play actually requires

Established before designing, because it constrains everything:

• Closed testing accepts testers as pasted email lists (up to 200 lists, 2,000 addresses each, 50 lists per track) or as a Google Group. There is no API to add an individual tester — any form we build ends in a human pasting a CSV.
• Every tester must opt in themselves through a web opt-in link, whichever method is used. Being on the list is necessary and not sufficient.
• An individual (non-organisation) developer account needs 12 testers opted in continuously for 14 days before production access. That number is a design target, not a footnote — the page should be written to convert.
• The store listing requires a privacy policy URL and a contact email address (§9, §14).

These are current as of 2026-08-19 and should be re-checked in the Play Console before launch; Play's testing requirements have changed more than once.

The opt-in link removes the need for email

D7 says the site sends no email. That looks fatal — with the email-list method, Google does not notify testers; the developer distributes the opt-in link. But the closed-testing opt-in URL is safe to publish, because it only works for addresses already on the tester list. Anyone else who opens it is refused.

So: the confirmation screen after a successful signup shows the opt-in link and says plainly "open this with the same Google account once you've been added — we'll announce each batch in Discord." Discord (D10) is the notification channel that email would otherwise be. No SMTP, no mailbox, no deliverability problem, and no unsubscribe machinery for a list nobody is mailed from.

This is worth stating in the plan because it is the decision that makes D7 workable rather than merely accepted.

The page

/beta — what the beta is, what it needs (an Android device, a Google account, and the willingness to stay opted in), current status, the form, and what happens next. Honest about the wait: batches are added by hand.

The store

SQLite at /app/data/beta.sqlite (bind-mounted), via better-sqlite3.

signups
  id             INTEGER PRIMARY KEY
  email          TEXT NOT NULL UNIQUE COLLATE NOCASE
  created_at     TEXT NOT NULL          -- ISO 8601 UTC
  ip_hash        TEXT NOT NULL          -- salted SHA-256, never the raw address
  user_agent     TEXT
  consent_text   TEXT NOT NULL          -- the exact wording they agreed to, versioned
  status         TEXT NOT NULL          -- new | exported | removed
  note           TEXT

The raw IP is never stored — only a salted hash, which is enough to rate-limit and not enough to identify. The salt lives in the container environment, so rotating it destroys the linkage deliberately. consent_text stores the wording itself rather than a version number, so a record can always answer "what exactly did this person agree to" without archaeology.

Phase 5 added the second table this section describes in prose but does not draw — attempts (ip_hash, at, outcome), which is the token bucket below, persisted as the events themselves rather than as a counter that would need a decay schedule and a clock it trusts. It is pruned on write, so nothing has to remember to run.

One consequence of remove that follows from the promise rather than from a separate choice: the address is overwritten, not flagged, so afterwards the store cannot tell a removed address from one it has never seen. Somebody who left and signs up again is an ordinary new row. Keeping a hash so the form could say "you were removed" would mean retaining a derived identifier for the one person who explicitly asked not to be retained.

Abuse resistance without a third party

D9 and §6's CSP forbid external requests, so no captcha service. Instead:

• A honeypot field, hidden from real users, that bots fill.
• A minimum time-to-submit — a submission under ~2 seconds after page render is a script.
• A per-ip_hash token bucket persisted in SQLite: a few signups per hour, a couple of dozen per day.
• A global cap on total rows, above which the form closes and says so, so the box cannot be filled.
• Strict validation, and a duplicate submission answered idempotently ("you're already on the list") rather than with an error that leaks whether an address is enrolled.

Getting the list into Play

A CLI, not an HTTP route:

docker compose exec site node scripts/beta.mjs export        # → /app/data/exports/<date>.csv, marks rows exported
docker compose exec site node scripts/beta.mjs export --all  # everything, including already-exported
docker compose exec site node scripts/beta.mjs remove <email>  # deletion request
docker compose exec site node scripts/beta.mjs stats

Phase 5 note: export writes two files, not one. The .csv is the record (id, address, date, status, the consent wording); the .txt beside it is one address per line, which is what Play's tester list actually wants pasted. Producing only the CSV would mean hand-editing it before every paste, which is where a mistake would come from.

Deliberately not an admin page. An authenticated HTTP surface on a marketing site is a login form, a session, a password to rotate and a thing to patch — for an operation performed by the one person who already has shell on the host, against a file already on their disk. The CSV lands in the bind mount and is opened locally.

remove exists because §9 promises deletion on request and a promise needs a mechanism.

How phase 5 built the app and the beta

Four decisions taken before the pages were written (org lead, 2026-08-24), plus what the repositories said when the plan above was checked against them.

Three things this section had assumed that turned out not to hold. §10 promised /app/ "the 14 existing screenshots"; they exist and are the wrong fourteen (D26). §8 never said what a tester would point the app at, and the app points at nothing by default (D27). And /app/ can offer a download today, because every Android-app release attaches a signed APK — which §8 and §10 both omitted, having been written as though Play were the only delivery path.

D26 — the screenshots are reserved for phase 9, and the slot ships empty. docs/android/screenshots/ is a trusted-device and recovery-code smoke test from 2026-07-22: captured against a development instance with no seeded content, before the theming work changed how every screen looks, and five of the fourteen are two-factor prompts. Shipping them would break D4 and would show an app that no longer looks like that. Phase 9 already stands up the review stack and seeds content for the web screenshots, so it gains an emulator pass and the phone shots then show the same deployment on the same day. src/components/app/Screenshots.astro exists now, rendering nothing, so filling it is a data change rather than a design task. Rejected: shipping the fourteen, and pulling phase 9's rig forward into phase 5.

Filled in phase 9: six captures, from an emulator pointed at the same seeded deployment the web screenshots came from, on the same day — see "How phase 9 took the screenshots" in §10. The component now reads src/data/screens.mjs rather than a list of its own, which is what made it a data change in the end.

D27 — the public demo is the tester target, so the beta waits for it. ConnectScreen.kt on Android-app main is unambiguous — nothing in the app runs until a valid Runic Gateway site has been entered and validated — so an installed app with no deployment behind it is a text field. The alternative considered and declined was naming UOMysticmoon, which would have opened the beta to players immediately at the cost of publishing a private shard's address on a public page. The consequence is accepted rather than hidden: the beta cannot start until §15's demo VM exists, which is the second of the two gates /beta states outright. It costs nothing today, because D28's other gate — no closed test track — is open anyway.

D28 — /beta handles its own POST; there is no /api/beta-signup. §6 specified an endpoint, and an endpoint cannot report a validation error without JavaScript: it answers with JSON, which makes the form script-only, or with a redirect, which returns a person who mistyped an address to a blank form with no explanation. Both are poor on a page whose job is conversion, and the first is worse on a site with no analytics — a form that silently does nothing for a reader with scripts off has no way of telling anyone it is broken. Handling the POST in the page costs one on-demand route and buys a form that works with JavaScript disabled, renders every outcome in the real layout, and needs no client-side code, so nothing on it argues with the CSP.

D29 — the APK and the beta get equal billing, and the APK link is currently off. Two panels of the same weight: sideload today, or join the closed test for Play delivery and automatic updates. The beta argues for itself on convenience rather than on being the only door. But the org lead reports that the published v0.5.0 build does not work, so platform.json's androidApk.serviceable is false and the panel renders a plain statement that the build is being replaced rather than a link. That flag is the one value in platform.json with no authority to check it against, and deliberately so — no fetch can tell whether an APK runs. checkFacts.mjs asserts the two assets still exist and that minSdk still says what "Android 10 or newer" claims, so the link is correct the moment a working build flips the boolean. The panel is not removed while the link is off: a page that omitted sideloading would read, to somebody who knows the APK exists, as a page hiding it.

Three mechanisms this phase added that the plan did not anticipate.

• liveBrand(), because a server-rendered page cannot use the boot rewrite. §7's mechanism rewrites files in dist/client; an on-demand route's HTML never was a file, so /beta reading brand would show stock values forever. It reads the mounted brand.json itself, guarded by an mtime check. That is strictly better where it applies — pasting the opt-in URL into the mount takes effect on the next request, with no restart. Phase 11 found it had been applied to one field and not to the chrome around it, and added renderBrand() — see D51.
• checkLinks.mjs learned what an on-demand route is. /beta is the first on-demand page, and rule 1 resolves links against the build, where it has no file. The fix is not a PLANNED_ROUTES entry — that list's reverse check fires when a route has been built, and an on-demand route never produces a file, so the entry could never rot out and would become the permanent exemption the two-way check exists to prevent. Instead the routes are derived from the source: a page exporting prerender = false is one. Delete beta.astro and the links fail again.
• A test suite, for the first time in this repository. The five checks of §12 all read built output, and none of this phase's logic appears there — a honeypot can stop working entirely and produce a build identical to one where it works. node --test, named file rather than directory (node --test test/ fails on Node 22, which is what CI runs).

---

9. Legal pages

Both drafted from what the code actually collects (D5), reviewed by the org lead before ship. Not legal advice — accurate and specific beats generated boilerplate, and the org lead decides whether it is sufficient.

/privacy

Three separately-scoped sections, because Runic Gateway is self-hosted software and conflating them would be wrong in both directions:

1. This website. No cookies. No analytics. No third-party requests of any kind. The one thing collected is a beta signup: email address, timestamp, salted IP hash, user agent, and the consent wording — used solely to add the address to the Google Play tester list, never sold, never mailed to. How to have it removed. Reverse-proxy access logs and their retention.
2. The Android app. The point that makes this unusual and must be stated precisely: we operate no server the app talks to. The app connects to an address the user enters (ServerPreferences / BaseUrlHolder), which is run by whoever runs that community. Data the app holds on the device — session and refresh tokens, the trusted-device token, the selected server, the ntfy push registration — and what leaves it, and to whom. Google Play itself collects its own data as the distributor; that is Google's policy, not ours.
3. Self-hosted deployments. A Runic Gateway deployment collects account data, IP addresses for bot scoring and IP bans, session records and audit logs. The operator of that deployment is the data controller for it, not us. This section exists so an operator understands the responsibility they take on, and so no player mistakes this policy for the one governing their community's site.

The Play Data Safety declaration is filled from section 2, and section 2 is written knowing that is what it is for.

/terms

Short and honest: the software is GPL-3.0-or-later and the licence governs its use; this site is informational and warranty-free; the beta is a beta and may break or end; acceptable use of the signup form; and how to get in touch. It does not attempt to govern anyone's self-hosted deployment, because it cannot.

Both pages are linked from the footer on every page, and /privacy is the URL given to Play.

How phase 6 built the legal pages

Four decisions taken before either page was written (org lead, 2026-08-24).

D30 — One hop in front of the site, and the page says so. §9 requires /privacy to state the access logs and their retention, which needed a fact rather than a guess. The domain's DNS is on Cloudflare but the records are DNS-only: no edge provider terminates the connection, so the reverse proxy on the org lead's own host keeps the only access log there is — IP, path, user agent, timestamp — read when something is broken or being attacked, rotated on the proxy's own schedule. The page describes it qualitatively rather than quoting a retention number, because the number belongs to the proxy's configuration and a policy that states one the deployment does not enforce is worse than one that does not. If the record is ever proxied, this section is wrong and has to be rewritten — an edge provider that terminates TLS is a processor, and D9's "no third-party requests" would still be true of the browser while ceasing to be the whole story.

D31 — Eighteen or older. Play asks, and the answer decides whether consent alone is a lawful basis in the EEA. Eighteen was chosen over thirteen (Google's own account minimum, but below the children's-consent threshold in several EEA states, so a 13–15 year old's consent would need a parent's — which this form cannot obtain) and over sixteen (sufficient, but no simpler to state). The number lives in src/data/legal.mjs because four surfaces render it: /terms, /privacy, the eligibility list on /beta, and the consent sentence itself. Nothing verifies it and no surface implies otherwise — the pages say in as many words that ticking the box is the whole of it, which is both accurate and the only claim the code supports.

Adding the clause changed CONSENT_TEXT, which is stored per row rather than versioned — so rows written from now on carry the new sentence and older ones keep theirs. CONSENT_VERSION gained a suffix rather than a new date, because the change landed on the day the original wording was written and two different sentences must not share the label an operator groups a CSV by.

D32 — No governing-law clause. Nothing of value is contracted for on this site: it sells nothing, the software is free under a licence that carries its own terms, and the beta is a list of addresses people asked to be on. A jurisdiction clause here would be decoration, and §9's standard for these pages is that accurate and specific beats boilerplate. It stays available: adding one later is a clause, not a rewrite.

D33 — The Play Data Safety notes are a generated repository document. §9 says the declaration is "filled from section 2, and section 2 is written knowing that is what it is for" — so the two are one array, src/data/collection.mjs, rendered by /privacy as prose and by scripts/playDataSafety.mjs as the console's own questions into a committed PLAY_DATA_SAFETY.md. --check regenerates and fails if the committed copy differs, and CI runs it, so a hand edit is a red build that names the data file to edit instead. The document is operator-facing rather than published: it is a form's worth of console vocabulary no visitor is looking for, and /privacy already says the same things in prose.

Two properties of that file are worth keeping. It does not pretend to know Play's current definitions — there is no API to read them from and the requirements have changed more than once (the same reason playPolicy carries a verifiedOn date), so it holds the facts arranged as the console arranges its questions, with the answer each fact supports and why; a person reads the console's wording against them. And a test asserts that every mapped row answers "not collected, not shared", failing with the reason rather than a diff: "we operate no server the app talks to" is the premise of the whole section, and a telemetry endpoint added later must not be able to produce a row that quietly contradicts the lede three inches above it.

What phase 6 also closed. /privacy/ and /terms/ were the last two entries in checkLinks.mjs's PLANNED_ROUTES; building them emptied the list, and its reverse check is what forced the deletion. The list itself stays, because §10's documentation routes land in phases 7 and 8 under the same convention.

---

10. Information architecture

Organised by what a reader is trying to do. A reader should never need to know that link, servuo-plugins and installer are three repositories in order to connect a game server.

Marketing pages

Route	Purpose
/	Hero, the data path, grouped capabilities, the self-hosted argument, get-started CTA. Reserves a "See it running" slot for the demo (§15) — laid out now, hidden until it exists
/features/	Grouped capability presentation, explicit about what is core and what a module supplies
/architecture/	The system explained visually, for a technical evaluator deciding whether to run it
/modules/	What a module is, module-uo as the worked example, writing your own, the Integration Kit (draft-badged per D8)
/integrations/	Discord, mobile + ntfy push, SSO — with an explicit "not built" list
/app/	The Android app: what it does, the signed-APK download beside the beta CTA, and six phone captures phase 9 filled (D26 — the 14 existing screenshots were the wrong fourteen)
/beta/	The closed-beta signup (§8). The one page that handles its own POST (D28)
/community/	Discord (discord.gg/t2Jav8yT4g) as the front door, the Gitea org for code and contributions, the brand.json contact address for vulnerabilities (D13) — the split in §14 N3
/privacy/, /terms/	§9

Feature grouping, using project terminology:

• Community — Teams, Team forums, notifications, wiki, news and newsletter, player self-service
• Game intelligence (module-supplied; module-uo today) — shard status, economy, character sheets, points and loyalty boards, player-vendor marketplace, houses and IDOCs, spawn atlas, champion boards, guilds, city governors
• Administration — roles, moderation and appeals, content reports, audit log, bot scoring and IP bans, module management, the shard connection
• Integration — modules, the sidecar bridge, Discord (slash commands, notifications, voice), mobile and push, SSO
• Infrastructure — self-hosted, Docker, prebuilt pull-only images, branding as data, OpenAPI

Guilds and city governors were added to Game intelligence in phase 3: module-uo declares them as capabilities and the site was omitting two of the eight. That correction is now mechanical rather than editorial — see D18.

Community is core machinery, but two of its six need a module to fill them. Teams and Team forums are marked as such (D24). Core owns every part of the Team machinery and cannot create a Team: they arrive from the installed module, so on a deployment with no module the feature is present and permanently empty. The group's summary says so; /features/ says why.

How phase 3 built the homepage

Three decisions taken before the page was written (org lead, 2026-08-20).

D17 — the data path is drawn generically, and captioned specifically. The diagram's nodes read "your game server", "sidecar", "Runic Gateway", "browser and app", because a reader should not have to know this org's repository layout to understand the picture, and because the tagline promises a platform. It does not hide what ships: the sub-labels and the caption name ServUO and uo-link outright, since there is exactly one implementation of the shape today and §1 says the technical truth wins. Rejected: naming the real components in the nodes (reads as a UO product), and omitting UO entirely (advertises a generality one module proves).

D18 — all five groups on the homepage, named only. Not three with a link out: Integration and Infrastructure carry the module and self-hosted arguments, which are the differentiators, and hiding them until phase 4 would have made the front page look smaller than the product. The per-capability argument stays /features/'s job so there is one copy of it.

The list is data with a check behind it (src/data/capabilities.mjs). Every Game-intelligence item names the module-uo capability slug it comes from, and the build fails if the page and platform.json disagree in either direction. Closing that loop needed a fifteenth fact in checkFacts.mjs: §12 listed the capability list as an externally-sourced fact and nothing re-read it, so the whole chain rested on someone remembering. Manifest → platform.json → page is now checked end to end.

D19 — the hero leads with the emblem. Chosen over a type-only hero: the mark is already the site logo, the Android launcher icon and the Play listing, and showing it large is what makes the three read as one product (D11). It costs what D16 already accepted — raster art a mounted theme.css cannot recolour — but every size is derived from whichever logo.png is in force (D14), so the hero, the header, the tab icon and the installed icon still change together from one file.

A convention, not a decision: the homepage links the final routes — /features/, /modules/, /integrations/ — which phases 4 to 6 have not written yet. The header and footer already did this from phase 1. Nothing is deployed until phase 12, so no visitor meets a 404, and nothing has to be rewritten later. Links into the documentation are the exception: they point at /docs/, because phases 7 and 8 own those slugs and a guessed one would be a stale URL nothing checks.

How phase 4 built the marketing pages

Six decisions taken before coding (org lead, 2026-08-20), plus two scope items the phase table had never assigned to anyone.

D20 — /features/ is the homepage's list with a detail line, not a second list. Every capability in src/data/capabilities.mjs gained a sentence or two of argument; / renders the label, /features/ renders the label and the detail. Rejected: slicing the page by reader (players / staff / operators / builders), which reads better but makes the same capability appear twice and breaks the one-to-one mapping the coverage check depends on; and deep-diving only the differentiators, which would have left the page looking smaller than the homepage promised. assertDetailCoverage() fails the build on a capability with no detail — the homepage would still look right, and /features/ would render a heading with nothing under it.

D21 — /architecture/ draws reasons, not reference. Three new inline SVGs, each drawing one boundary: two hosts and two installs, the public/staff allowlist, and the core/module seam. It carries no endpoint tables, no configuration keys, no schema and no event catalog — phase 8 owns those, they are canonical in docs/, and a second copy here is a copy that goes stale (§1). Rejected: reusing the homepage's data-path diagram larger (a visitor arriving from / meets the same picture twice), and adding a component/version table (starts becoming the Reference section). The vocabulary the four diagrams now share moved to src/styles/diagram.css.

D22 — the deliberate absences are one data file. src/data/notBuilt.mjs, each entry tagged with the pages that render it, because the homepage already promises a reader they will find the list on both /features/ and /integrations/ and two hand-written copies is how the inconvenient half stops appearing on one of them. Every entry carries a resolvedBy: an absence with an exit condition is a position, an absence without one is a hole. That generalises what D8 already required of the Integration Kit's draft badge.

D23 — phase 4 absorbs /community/ and checkLinks.mjs. Neither had a phase. §10 specifies the page and §14 N3 specifies its contents, and the header and footer have linked it since phase 1 — a page the site pointed at that no phase built. checkLinks.mjs is specified in §12 and phase 4 is what makes it load-bearing: it roughly quadrupled the internal link count and added the first outbound links into the repositories.

D24 — needsModule, because "core" and "module-supplied" were not enough. Writing the /features/ detail for Teams exposed a claim phase 3 had shipped: the Community group said "everything here works on a deployment with no game module installed at all", and that is false. teams.module_id is NOT NULL on website main, there is no create route anywhere under /api/v1/admin/teams, and sync is gated on teamProvider.providerModuleId(). Core owns the whole Team machinery — tables, roster resolver, forums, notification streams, Discord bridge, voice, activity feed, /admin/teams — and deliberately cannot originate a Team, because core does not own the word for one. On a bare core the feature is present, correct and permanently empty. Teams and Team forums are marked; the group summary was requalified; the homepage changed by one sentence and kept D18's five named groups. Rejected: a sixth group for Teams (says it loudest, costs the five-group grid phase 3 tuned), and fixing only the wording (leaves the distinction one sentence deep and unguarded).

D25 — the demo affordance on /features/ is a per-capability deep link. brand.json had promised one since phase 2 without defining it. Capabilities with a stable public route carry a link appended to the mounted demoUrl; the rest carry nothing, and that asymmetry is honest — a character sheet is reachable only by the account it belongs to, and a Team forum lives behind an id no static page can know. Paths are read from the real route tables on main, never guessed, which also means they are the module's routes: a deployment running a different module deep-links somewhere else.

That needed the branding pipeline extended, because the phase-3 slot could not express it. The slot is a literal swap of a whole URL, so it can only ever put the demo's root in an href, and reversing it would not even find a deep link — whose href is the root plus a path, matching no literal the script knows. applyBrand.mjs gained a second pass that recomputes all three attributes from the immutable data-demo-path, making it idempotent and exactly reversible, and checkBrand.mjs gained a guard that lifts the pattern out of applyBrand.mjs and runs it against the stock markup, so the two cannot drift. Both directions were proved against a real mount.

One thing the checks caught about each other. A scoped :has([data-demo-url='']) rule, added to hide the wrapper around a hidden demo link, made checkBrand.mjs fail: it cannot tell a CSS selector from an attribute, and it should not have to. The right fix was to delete the wrapper and let the link be the flex item, so the existing hide rule takes the margin with it — a case where the check being blunt pointed at simpler markup rather than at a needed exemption.

And one thing no check caught. [data-demo-url=''] { display: none } is specificity 0,1,0, and so is the scoped class Astro puts on the same element — so a component that sets display wins on source order, because component styles are emitted after global.css. /features/'s .demo-link set display: inline-flex for its arrow, and twelve links to a demo that does not exist rendered on the page, each pointing at href="" — which a browser resolves to the page it is already on. checkBrand.mjs was green throughout: the attributes were perfect and the defect was three files away, in the cascade. It was found by looking at the rendered page at 390px, which is not a mechanism, and it is the argument for keeping the live browser pass in every phase. The rule is now !important and says why in the stylesheet: while there is no demo these elements do not render, and no component may overrule that by accident.

Documentation

Getting started    What is Runic Gateway? · Requirements · Install the site ·
                   First run · Install a game module · Connect a game server ·
                   Verify the whole stack

Administration     Configuration · Branding and theming · Navigation and pages ·
                   Content · Users and roles · Authentication · Teams · Moderation ·
                   Notifications and email · Managing modules ·
                   The shard connection · Maintenance and upgrades · Troubleshooting

Modules            The module system · Installing modules · Module lifecycle ·
                   The module manifest · The module API · Building a module ·
                   The Integration Kit · Testing and release

Architecture       System architecture · The bridge · Authentication architecture ·
                   Teams architecture · Protocol versions

Reference          Environment variables · Installer CLI · sidecar.toml ·
                   Bridge.cfg · HTTP API · Event catalog · Canonical documents

Forty pages — thirty-nine planned, plus the Content page D37 added in phase 7. (This said "roughly 38, 37 planned" until phase 8 counted the tree: 7 + 13 + 8 + 5 + 7. checkSidebar.mjs now keeps the count honest.) Every Reference page is a navigable summary plus a link to the canonical document — never a re-specification, per §1.

The installation path

The repos treat these as separate deployments; this site is the first place they appear as one sequence. Worth stating plainly, because it is a real trap: a "Runic Gateway install" is two independent installs. The installer binary sets up the shard side only and never contacts the website; the website is a separate Docker deployment.

1. Requirements
2. Install the site — Docker Compose, pull-only
3. First run — first admin, maintenance → live
4. Install a game module — admin panel, MODULES env, or by hand
5. Connect a game server — the installer binary on the shard host (ServUO-specific today)
6. Paste the four values into Shard (uo-link), /admin/uo/link — protocol 4, per §2. (Not /admin/shard: the screen belongs to the module now, and the installer still prints the old path — see "How phase 7 built the documentation journey" below)
7. Verify — [bridge status in game, /health reporting plugin_connected: true, then doctor
8. Configure authentication and integrations

Each step states what the operator should expect to see, and links the failure modes to Troubleshooting.

How phase 7 built the documentation journey

Four decisions taken before a page was written (org lead, 2026-08-24), and three things the live site disproved while it was being written.

D34 — one PR for all twenty pages. Twenty, not nineteen: see D37. The alternative on the table was splitting Getting started from Administration so the installation path could land first; the org lead kept the phase whole, as every phase before it has been.

D35 — the install page is self-contained. /docs/getting-started/install-the-site/ prints a complete Compose file and a complete .env that an operator copies without going to another repository first. §1 argues at length against exactly this — it is a second copy of somebody else's file, free to rot — so the copy is not trusted, it is checked. src/data/quickstart.mjs holds both files and the page renders them; scripts/checkQuickstart.mjs re-reads website's own docker-compose.yml and .env.example from main over the Gitea API and fails the build on any disagreement, in both directions:

• every value the quickstart states must match upstream's;
• every service and variable upstream has must be either included or listed as deliberately omitted, with a reason, so a new variable in .env.example turns this repo red until someone decides whether a first install needs it;
• and an entry in either omission list that upstream no longer has fails too, so the lists cannot rot into permanent exemptions.

Same mechanism and same intent as checkFacts.mjs. It caught two stale entries on its first run — TOTP_ISSUER and MODULES, which are commented suggestions upstream rather than keys — which is the check earning its place before the page had shipped.

D36 — every Administration screen was walked before it was described. Not read from source: opened, in a browser, on a real deployment. The rig was the quickstart itself — the exact two files from D35, against the published image — so one run proved the install page and produced the screenshots' worth of detail the admin pages needed. Three of the four defects below came from that walk, and no check could have found any of them.

D37 — a thirteenth Administration page. §10's planned twelve named no home for Posts, Pages, Wiki, Activity, Invites, the Hero editor or Web Bot Activity, all of which are real admin nav rows. Rather than mirror the panel one page per row — which would organise the docs by the app's menu, against this section's own principle — content authoring became one page, Content, and the other four folded into the page that already owned their subject: Invites into Users and roles, the Hero editor into Branding and theming, Web Bot Activity into Authentication.

What the live deployment disproved.

• The documented Compose deploy does not boot. SECRET_ENC_KEY is required in production — utils/secretBox.js throws at require time, so the container crash-loops before it listens — and it is missing from website's root .env.example, the file Compose actually reads. It is present in server/.env.example, which is the file local development copies, which is why this has never bitten anyone in dev. The quickstart carries it, declared as an upstream omission so the check fails the day it is fixed. Fixed in website#163 (merged 2026-08-24), which also adds BOT_INTERNAL_KEY to the README's "set at least" list — required in production even on a deployment running no bot. The declaration did exactly what it was built to do: this repo went red on the next run, and the entry is deleted here.
• The installer points operators at a screen that no longer exists. It prints <site>/admin/shard, and INSTALL.md §5 repeats it. Since the module-system cutover a module owns one path segment, and the screen is /admin/uo/link, labelled Shard (uo-link). The old path does not even 404 — the SPA sends the operator to the dashboard, so the link looks like it worked and the four values have nowhere to go. Fixed in installer#22 (the path is a named constant and both handoff tests assert it) and docs#174, both merged 2026-08-24, and shipped in installer v0.1.1. Getting there found a fourth defect, in installer's release pipeline: the run for the fix built every artifact and pushed tag v0.1.1, then took a 500 from POST /releases one second later, leaving an orphan tag and no binaries. Re-running the workflow published it — the failure was a race with the tag push, not a structural one — so the note here names v0.1.0 as the version that prints the old path rather than describing the installer as currently wrong.
• The admin "Restart the server" button opens a window.confirm. Its text is the honest warning that a deployment with no supervisor does not come back — which is exactly why restart: unless-stopped is called out as load-bearing on the install page rather than left as boilerplate.

And the fourth defect, the one only a look found — three phases running. The .env block's prose says every highlighted line must be changed, and mark given the variable names highlighted the names alone, leaving the values a reader has to replace unmarked. The build passed, every check passed, and the page was quietly wrong about its own highlighting. Marking the whole KEY=value string fixed it. See phase 4 (cascade), phase 5 (literal backticks) and phase 6 (the card void) for the same lesson.

---

How phase 8 built the builder and reference docs

Twenty more pages — Modules (8), Architecture (5), Reference (7) — completing the tree §10 planned. Four decisions, taken by the org lead before anything was written.

D38 — one PR for all twenty pages, again. The alternative on the table was splitting the prose (Modules + Architecture) from Reference, since only Reference needed new checking machinery. Rejected for the same reason D34 was: the three sections cross-reference each other heavily, and a split means either landing pages whose links point at nothing yet or writing the links twice.

D39 — Reference enumerates the NAMES, and checks every one of them. This is the phase's central decision, because §1 forbids re-specifying a contract and a Reference section is exactly where that rule is most tempting to break.

The line drawn: names are on the page, semantics are not. Every environment variable, config key, installer command, visibility rung and canonical document is listed, with one terse line saying what it is for. Shapes, defaults that matter, interactions and every "why" stay in the canonical document.

That is only safe because scripts/checkReference.mjs compares each list against the repository that owns it — six sources, over the Gitea API, never from a working tree — as a set comparison in both directions. The second direction is the one that earns its keep: a reference page does not usually rot by describing something that vanished, it rots by quietly not mentioning the three things added since it was written.

The alternative considered was strict summary-plus-link with nothing enumerated. It needs no machinery and cannot rot — but a Reference section that cannot answer "what variables are there?" without a click-through is a link farm, and the checking machinery turned out to be one script.

Descriptions are deliberately not checked, and the script says so. Nothing can know whether a one-line summary is still true; keeping them short enough to re-read is the mitigation, not a check.

D40 — the docs link to the drawn diagrams rather than importing them. /architecture/'s three diagrams are Astro components carrying marketing chrome and depending on src/styles/diagram.css, which Starlight does not load. Reusing them inside the docs would have coupled the two layouts for one page's benefit. The docs use text diagrams in code blocks — which are also copy-pasteable into an issue — and link out to the drawn versions.

D41 — plannedSidebar stops being a checklist and becomes a checked invariant. It was written in phase 1 so phases 7 and 8 had their checklist where they would be working. With every page now written it is a second, hand-maintained copy of the live tree, which is the exact shape §1 warns about — so checkSidebar.mjs asserts the two agree on groups, labels and order.

Order, because the order of "Getting started" is the installation path, and a reordering nobody noticed would be a worse defect than a missing page.

What the checks found, before any of the pages shipped.

• plannedSidebar had already drifted. Phase 7 added the Content page under D37 and never updated the planned list. Nothing failed, because nothing read it — which is the whole argument for D41. Reproduced by deleting the entry again and watching the new check catch it.
• The page count in this document was wrong, and had been since §10 was written: it said "roughly 38 — 37 planned", where the tree it describes is forty.
• module.json's mounts and the SPA's paths are different mechanisms, which is not stated plainly in any one place. module-uo declares admin: ["/shard", "/uo-link"] and its screen lives at /admin/uo/link; API routes are deliberately not namespaced while SPA routes are. That is the distinction the installer got wrong in v0.1.0, and it now has a named home on The module system.

The check was verified by breaking it, not by watching it pass. It went green on its first run, which is the least trustworthy possible outcome, so seven mutations were fed through it — a stale name, an omitted name, a renamed key in each of three sources, a canonical document that moved, and the visibility ladder reordered with its membership unchanged. All seven failed the build. The ladder case is the one worth keeping: it is a security boundary, and a set comparison alone would have passed it.

---

How phase 9 took the screenshots

D4 said real screenshots from the review stack rather than placeholders, and left the how open. Five decisions settled it, taken by the org lead before the rig was built.

D42 — the full rig: a real shard, a real sidecar, a real site. ServUO with the bridge overlay on this machine, the Rust sidecar beside it, website main with module-uo installed, and the demo database seeded on top for what a fresh shard cannot produce.

The alternatives were cheaper and both of them lie a little. Sidecar-only screenshots the degraded state — a reachable bridge with nothing behind it. Everything-database-seeded produces pages that look identical to the real thing and were produced by nothing: the marketplace would be rows somebody typed. This is the one option where the marketplace rows are player vendors the game actually holds, the atlas is parsed from the shard's own spawn files, and "Candlewick House is now IDOC" happened.

D43 — a neutral demo brand. The deployment is "Runic Gateway Demo", not UOMysticmoon. The screenshots show the platform rather than one private community, which is the same instinct as D27's refusal to publicise a real shard — and §15's demo instance can wear this identity the day it exists, so the imagery stays true rather than becoming a period piece. The name says "Demo" deliberately: nobody should have to wonder whether they are looking at a server they could join.

D44 — the captures sit beside the claims they support, in two places. A figure set on /features/, one on the homepage, and inline shots on the phase-7 administration pages that describe a screen in prose. Eleven web captures.

The administration pages are where a screenshot does the most work, because phase 7 described thirteen screens it could not show. A dedicated /screenshots/ gallery was rejected for the reason galleries usually are: a page nobody visits does less than a figure sitting under the sentence it proves.

D45 — the rig is committed, not remembered. Three files rather than a folder of images: scripts/seedDemo.mjs puts the content there by driving the site's own API, src/data/screens.mjs declares every capture with its route, viewport, scroll offset and caption, and scripts/captureScreens.mjs turns the second into files. scripts/checkScreens.mjs is the ninth check script and runs in CI.

The argument is the same one D35 made for the install quickstart: the way real screenshots rot is that the recipe for taking them lives in somebody's memory. Re-taking the set after a redesign is now npm run screens:capture, and the check fails if an entry has no file, a file is the wrong size, a file is orphaned, or a declared screen is rendered nowhere.

Why the seed drives the API and never the database. Every row it creates could have been an INSERT, and every INSERT would be a second implementation of a rule the website owns — how a body is sanitized, which excerpt is derived, how a password is hashed. A seed that writes SQL produces a database the product could not have produced, and screenshots of that database show a product that does not exist.

D46 — the world gets dressed in servuo-plugins/tools. BridgeSeeder builds a world at realistic scale; it never needed the world to look like anything, so a vendor traded as "Seed Shop 810" and a character was "Seed004A" — and every one of those strings travels the whole bridge and lands on the marketplace, the guild roster and the housing pages. BridgeDemoDress renames them in place and seeds nothing, drawing names from fixed tables hashed off each object's serial, so a re-run reproduces the same world and a screenshot can be retaken later and still match.

What the rig found. A screenshot rig is an integration test with a human in the loop, and this one turned up six things nothing else had:

• A fresh module-uo install pinned wire protocol 3 while the sidecar speaks 4, so a new deployment 409s on every shard read until an admin edits the number by hand. The protocol-4 cutover bumped link, servuo-plugins and docs and missed the module's own default. Fixed upstream and released as module-uo v1.0.2 — which is what this repository's own facts check then noticed, since platform.json still said v1.0.1.
• A renamed guild member never reaches the site. The plugin folds name, abbreviation, leader, member count and alliance into the signature it compares, and re-emits the roster only when the member set changes — so renaming a member leaves the published roster stale indefinitely.
• A guild deleted while the shard is offline is a ghost row forever. The "gone" pass compares against a cache that is cleared on reconnect, so nothing emits guild.remove. The demo's guild board was showing two guilds the world no longer had, a week after they went.
• "Houses in danger" cannot show a house that was already collapsing. The ingest writes that column only from the house.decay transition feed, while the registry frame's stage is deliberately left alone so the two cannot clobber each other. A house already in IDOC when the site connects is therefore invisible — the page said none while the shard had two.
• The Android news list prints raw ISO timestamps. Found while choosing the phone captures; the news screen was dropped from that set rather than shipping a picture of it.
• The app says "1 players online". shard_online_count and ShardEventText.kt both interpolate a count into a fixed plural. Found in the retake after a character was signed in, and it is in the shipped phone capture — a plurals resource is the fix, in the app.

The first is fixed. The rest are raised as product observations, with the rig working around them: the guilds are built after the rename, and the IDOC staging is two passes with a wait between them so the site watches the collapse happen. All of that is scaffolding under servuo-plugins/tools/, which is never deployed.

The emulator pass (D26), and the AVD that would not take it. The six phone captures come from an emulator pointed at the same deployment on the same day, signed in as an ordinary player, reached through adb reverse — the app's debug network policy permits cleartext to localhost only, which is a better default than the one that would have made 10.0.2.2 work. The device is API 35 rather than the API 36 the plan named: the API 36 image on this machine had 200 MB free and refused the install, and wiping somebody's development device to take a screenshot is not a trade worth making.

The shard screen is the one worth having. It shows the two houses entering IDOC in its live activity feed — the same event that reached /uo/houses in the browser, on the same rig, in the same minute.

The character that had to be logged in by hand. The org lead asked for a player in the world, and the scaffolding does its half — it sets a known password on a seeded account, because BridgeSeeder gives every account a random GUID nobody kept. Driving the client is where automation stopped. ClassicUO stores its password crypted, so a plaintext one in settings.json decrypts to garbage and auto-login fails; posted mouse clicks reach the client but posted text does not; and the remaining route — taking the foreground and typing — was tried once, failed to take focus, and typed into the browser window the person at this machine was using. It was not tried again.

The org lead signed in instead, and the two frames that depended on it were retaken: the shard page now reads 1 player online, in Britain, and the app's shard card agrees. Both came from the same npm run screens:capture shard-status app-shard, which is the whole point of D45 — the thing that changed was the world, not the recipe.

Two things that pass is worth noticing here. Presence reaches the public page as counts and regions, not names, which is the visibility framework doing its job unprompted. And the guild board's "online" column did not move: it is refreshed only when a guild's signature changes, which is the same defect as the stale roster above wearing a different hat.

A layout decision worth recording. The /features/ figures are one-up at the column's full width, not a two-column grid. Two-up was built first and is the obvious layout for a set of figures — but these are screenshots of a dense interface, and halving the width puts the product's own type at about a third of its real size, which reads as a thumbnail of something rather than a picture of it. A long section of legible evidence beats a tidy grid of unreadable tiles.

---

11. Visual direction

"Modern infrastructure software with an arcane identity." Dark-first. Marketing pages are single-theme by design; the docs honour the reader's light/dark preference.

The mark is the existing emblem (D11). runic-emblem.png — a gold-and-ruby ring around a glowing cyan portal — is already the site logo and the Android launcher icon, so adopting it makes the three surfaces one product. Revision 1 argued for replacing it; the org lead likes it, and a mark already carrying recognition beats a better-drawn one that carries none.

Work needed on it, none of which is a redesign:

• Web derivatives from the 1024px source: it is a 1.4 MB PNG, far too heavy for a header. WebP and AVIF at header, hero and OG sizes, plus a real multi-resolution favicon.ico and the 192/512 PWA icons.
• A horizontal lockup — emblem beside "Runic Gateway" set in Cinzel — for the header and OG card.
• All of it lands in /app/brand-default (§7), so the org lead can replace any of it with an updated logo by copying a file.

Palette, revised. Revision 1 proposed a muted violet secondary. That was invented; the emblem already has a palette and it is better. The site descends from the product's tokens (website/client/src/styles/theme.css) and takes its accents from the mark:

• Near-black ground and panels — the product's --bg family, unchanged.
• The product's steel-blue #7f99bd for interface and links, unchanged.
• Gold and cyan from the emblem as the accent pair — gold for emphasis and rules, the portal's cyan for the glow behind the diagrams and the live-state signal. Both derived from the artwork by sampling, not guessed, and both held to WCAG AA against the ground.
• The product's existing live/maintenance signal colours (--mode-live, --mode-maint) reused verbatim, so a status pill means the same thing on both sites.

Type: Cinzel — already the project's display face, and already in the Android app's res/font — for the wordmark and hero only; a modern variable sans for everything else. Self-hosted, so there are no external font requests and §6's CSP needs no exception.

Motif: hand-drawn SVG geometry — a gateway glyph derived from the emblem's concentric rings, thin luminous topology lines, layered panels. Subtle, structural, used where it explains something. No AI-generated rune artwork anywhere.

Maturity signals (D8) are typographically quiet: a small "draft" chip on Integration Kit pages, version chips wherever a component is named. No apologetic tone anywhere. The Integration Kit chip has a defined removal condition — a second module built successfully against the kit — recorded here so a future reader knows when to take it down.

---

12. Accuracy machinery

The site quotes versions, protocol numbers and capability lists. Given §1's process rule, that needs a mechanism rather than diligence:

• src/data/platform.json — one file holding every externally-sourced fact: protocol version, Module API version, current bundle tag, component versions, module-uo's capability list. Every page reads from it. No version number is ever hardcoded in prose.

• scripts/checkFacts.mjs — fetches the authority for each fact and fails the build on any disagreement:

  Fact	Authority
  Protocol version	link main:sidecar/src/main.rs → PROTOCOL_VERSION
  Overlay protocol	servuo-plugins main:overlay.toml → protocol
  Module API	website main:server/src/modules/version.js → MODULE_API_VERSION
  Module capability list	Module-uo main:module.json → capabilities (added in phase 3)
  Bundle + component pins	installer branch bundles, root current.json
  Release versions	Gitea releases API per repo

  Same mechanism and the same intent as the Integration Kit's checkCoreApi.js: when the platform moves, this repo goes red so someone updates the site. That failure is the feature.

  It carries one rule that is not a version: no email address may appear anywhere in the source except brand-default/brand.json. D13 publishes a personal address that is meant to be replaceable by a file copy, and that promise survives exactly as long as nobody types the address into a paragraph. Same argument as checkTokens.mjs and colour literals — the check is the mechanism, diligence is not.

• scripts/checkLinks.mjs — every internal link resolves; every outbound link into a RunicGateway repo points at a branch path, not a commit permalink. Built in phase 4 (D23), and it reads dist/client rather than src/: half the links these pages carry are assembled from data files and template literals, and a source scan sees an expression rather than a URL. It runs after the build for that reason, in verify and in CI. It fetches nothing — the outbound rule is about the shape of a URL, and a check that fails when someone else's host is slow is a check people learn to ignore.

  It carries one exemption list, PLANNED_ROUTES, because §10's convention is that the header, footer and homepage link the final routes rather than growing links phase by phase. That is safe only because the list is checked in both directions: a link to a route that is neither built nor listed fails, and an entry whose route has since been built also fails, so the list cannot rot into a permanent exemption once the page arrives.

• scripts/checkBrand.mjs also guards the demo slot and, since phase 4, the per-capability deep links (D25) — lifting the pattern out of applyBrand.mjs and running it against the stock markup, so a template and a script that share no code cannot drift apart. Both are invisible in a stock build, which is exactly why they need a check rather than a look.

• scripts/checkQuickstart.mjs — added in phase 7 for D35. The install page prints a Compose file and an .env verbatim, which is the one place this site knowingly copies another repo's file; this re-reads website main:docker-compose.yml and main:.env.example and fails on any disagreement. Two-directional, like PLANNED_ROUTES: a value that drifts fails, and a service or variable that appears upstream fails until it is either included or recorded as deliberately omitted with a reason. Its own first run found two stale entries.

• scripts/checkScreens.mjs — added in phase 9 for D45. src/data/screens.mjs is the one list of what the site shows of itself, and this proves every entry has a file at the size the markup declares, that nothing in public/screens/ is orphaned, and that every declared capture is rendered somewhere. The size half is the one that repays it: a re-capture taken at the wrong viewport looks perfectly fine on its own and only reveals itself as a page that reflows while it decodes. No browser and no game server — the capture tool is an authoring script whose output is committed, exactly like the brand assets.

• scripts/checkTokens.mjs — no colour literal outside the token file (§7).

• astro check plus a production build, in CI on every PR.

---

13. Build phases

Phase	Deliverable
0	The documentation fixes in §4 — all ten conflicts, across docs, .profile and Module-uo
1	Foundation: Astro + Node adapter scaffold, the token file, typography, layout shell, header/footer, docs theming and sidebar, platform.json + checkFacts.mjs + checkTokens.mjs
2	Branding pipeline (§7): /brand/* resolution, brand-default contents, the emblem's web derivatives and lockup, brand.json wiring
3	Homepage: hero, the data-path diagram as inline SVG, grouped capability sections, CTA, the reserved demo slot
4	Marketing: /features/, /architecture/, /modules/, /integrations/, and /community/ — plus checkLinks.mjs, the capability detail lines, notBuilt.mjs and the demo deep links. See D20–D25
5	The app and the beta: /app/, /beta/, the signup handler, the SQLite store, rate limiting, the export CLI (§8). Also the repository's first node --test suite, and phase 9 inherits an emulator pass (D26)
6	Legal: /privacy/, /terms/, footer links, and the Play Data Safety notes (§9)
7	Docs — the journey: Getting started (7) + Administration (13, per D37) — twenty pages in one PR (D34), with the install page self-contained and drift-checked (D35) and every admin screen walked before it was described (D36). The installation path is the priority of the whole project
8	Docs — builder and reference: Modules (8) + Architecture (5) + Reference (7) — twenty pages in one PR (D38), with Reference enumerating names and checking every one of them against its source (D39), and plannedSidebar becoming a checked invariant (D41)
9	Screenshots (D4): stand up the local review stack, seed presentable content, capture the admin panel, Teams, forums, marketplace, spawn atlas and shard console; build the screenshot components. Plus an emulator pass against the same seeded stack to fill /app/'s reserved slot (D26)
10	Polish: responsive, accessibility, SEO/OpenGraph/sitemap/robots, full-text search, CSP headers. See D47–D50 — the CSP was the work, because @astrojs/node served every page another page's policy
11	Validation: astro check, production build, all eleven check scripts (tokens, brand, links, facts, quickstart, data safety, reference, sidebar, screens, a11y, CSP) plus both test suites, mobile layout verified in a real browser, a signup walked end to end. See D51–D53 — the scripts were green before the phase started; the browser walk and a real brand mount are what found the three defects
12	Delivery: two-stage Dockerfile, pull-only docker-compose.yml with both bind mounts, .env.example, the Gitea Actions workflow that publishes and deploys (D54), README, CONTRIBUTING, the community-health files this repo alone lacked, and DEPLOY.md (D55, D57). It found the one defect that only a container behind a proxy can find — see D56

Phases 5 and 6 are deliberately adjacent and early: the beta cannot start without /privacy, and the closed test is the nearest real deadline.

All twelve are built, as of 2026-08-25. The rgcom runner was registered on the host the next day, in host mode as §7 of DEPLOY.md requires. What is left is not a phase: point the DNS record at the host (§14, N1), and — when the demo VM exists (§15) and the Play track is open — put two URLs into the mounted brand.json. Neither is a code change, which was the point.

One thing did need a code change. The merge that landed phase 12 built its image and then could not publish it: Cloudflare rejected the largest layer with 413 Payload Too Large, so the registry stayed empty and the deploy never ran. See D58 — the layer is split, and the workflow now counts layers before it pushes.

---

14. Still needed from the org lead

None of these block starting Phase 0 or Phase 1.

N1 — Resolved. runicgateway.com is registered through Cloudflare, with DNS on Cloudflare. The domain does not resolve to anything yet. Phase 12 shipped everything needed to point it: the A record, the proxy requirements and the three Cloudflare features that break a hash-based CSP are in DEPLOY.md §4. Creating the record is the org lead's, on the day the host is up.

N2 — Settled by D13, and no longer blocking anything. As of 2026-08-19 no mailbox exists at the domain and the org lead chose not to wait for one. whitlocktech@gmail.com is the published contact, and docs/SECURITY.md keeps the address it already carries.

The site is built so this is reversible for the cost of a file edit. The address is a brand.json field (§7), read by /privacy, /terms, /community and the OG/structured metadata; it appears in no page's prose and in no committed markdown, and checkFacts.mjs (§12) fails the build if a literal @ address is found outside brand-default/brand.json. Changing the published contact is then the same operation as changing the logo — copy a file, restart the container.

The recommendation stands for whenever it is wanted: Cloudflare Email Routing, now that N1 confirms the domain is on Cloudflare. Free, unlimited, receive-only, and it adds its own MX and SPF records automatically. Three addresses, all forwarding to the existing inbox:

Address	Would replace the Gmail in
hello@runicgateway.com	The Play store listing's contact email; /community
privacy@runicgateway.com	/privacy — deletion and data requests (§9)
security@runicgateway.com	docs/SECURITY.md (a one-line docs PR, whenever)

Pair it with Gmail's "Send mail as" to reply from the addresses rather than the underlying inbox. Routing is receive-only, which is sufficient — the site sends nothing (D7). If double opt-in is ever wanted later, Resend or Brevo's free tier plus DKIM/DMARC would cover sending.

One consequence to accept knowingly: a Play store listing's contact email is public, and so is this site, so the address will be scraped. That is the trade the org lead took against delaying the beta; the mitigation is that switching later costs one file.

N3 — Gitea registration. Open registration previously drew a flood of spam accounts, so it is disabled, and "file an issue" is therefore a wall for anyone outside the org. Reopening it safely on Gitea 1.24.7 is a config change, not a gamble:

[service]
DISABLE_REGISTRATION           = false
REGISTER_MANUAL_CONFIRM        = true    ; an admin approves each account before it can sign in
REGISTER_EMAIL_CONFIRM         = false   ; manual confirm is ignored unless this is off
ENABLE_CAPTCHA                 = true
CAPTCHA_TYPE                   = cfturnstile
CF_TURNSTILE_SITEKEY           = <from Cloudflare — free, and the domain is already there>
CF_TURNSTILE_SECRET            = <ditto>
DEFAULT_ALLOW_CREATE_ORGANIZATION = false
DEFAULT_KEEP_EMAIL_PRIVATE     = true

[repository]
MAX_CREATION_LIMIT             = 0       ; -1 is unlimited; 0 means no user may create a repository

• REGISTER_MANUAL_CONFIRM is the control that actually stops the flood. Registrations queue in Admin → User Accounts and cannot sign in, comment or open issues until approved, so a spam run produces a list to delete rather than content to clean up.
• MAX_CREATION_LIMIT = 0 is defence in depth, and is what "users cannot create repos by default" means concretely. Raise it per-account in Admin → User Accounts for anyone who should be able to. Set the org lead's own account to -1 explicitly before flipping the global, rather than assuming admin accounts bypass it.
• Turnstile is free and the domain is already on Cloudflare, so it costs one form.
• Caveat worth weighing: with no mailer configured on Gitea, nobody is notified when an account is pending, and the applicant is not told when they are approved. That makes Gitea a slow channel for someone who just wants to report a bug.

Consequently, /community should describe an honest split rather than one channel: Discord (discord.gg/t2Jav8yT4g) is the front door for questions, bug reports and the beta announcements — instant and unauthenticated. Gitea is where the code lives, is publicly readable without an account, and takes issues and pull requests from contributors willing to request one. The brand.json contact address (D13) is the private channel for vulnerabilities. That page is written the same way whether or not N3 is actioned; only one sentence changes.

---

15. Planned, not in scope: the demo instance

D12. Recorded now so the IA reserves room for it and nothing has to be restructured later.

Shape: a VM on the org lead's Proxmox host running the full stack — website, module-uo, the uo-link sidecar and a real ServUO shard — so an evaluator clicks one link and sees the actual product with live game data flowing through the bridge, which no screenshot can convey.

Constraints that make it safe to run:

• Automatic reset roughly hourly from a golden snapshot, so nothing an anonymous visitor does outlasts the hour. Reverting the VM is simpler and more complete than any application-level reset.
• Restricted settings — the demo account reaches most of the admin panel, but not the surfaces that would let a visitor break out or reach the network: SSO client secrets, the uo-link token and base URL, mail, the module installer, and anything shelling out. The specific allow-list is designed when it is built, against the admin route inventory.
• Seeded content shared with phase 9's screenshots, so the demo and the site's imagery agree.
• Isolated network segment; no path from the demo VM to anything else on the host.
• Its own subdomain, its own TLS, a noindex header, and a banner stating it resets hourly.

Site-side preparation done now, for free: the homepage lays out a "See it running" slot and /features/ a per-capability "try it" affordance, both rendered only when brand.json carries a demo URL. When the VM exists, the site gains a working demo by way of one line in a bind-mounted file — no rebuild, consistent with §7.

---