@aglyn/aglyn 1.0.0-beta.143

package/src/lib/app-utils/screen-route.d.ts

@@ -0,0 +1,669 @@
+/**
+ * @license
+ * Copyright 2026 Aglyn LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+import { type ScreenUid } from '../foundation';
+/**
+ * Route path of a host's root screen. The tenant matcher joins the catch-all
+ * segments and collapses an empty one to this value (`load-page-data.ts`:
+ * `slugSegments.length ? slugSegments.join('/') : SCREEN_ROOT_PATH`), so the
+ * root is `'/'` and every other path is slash-joined segments WITHOUT a
+ * leading slash (`about`, later `company/about`).
+ */
+export declare const SCREEN_ROOT_PATH = "/";
+/**
+ * Normalizes user slug input into the routing-map path format described on
+ * {@link SCREEN_ROOT_PATH}. `'/'` normalizes to the root path; anything else
+ * becomes a single lowercase url-safe segment.
+ *
+ * Returns `undefined` for EMPTY input and for anything that sanitizes away
+ * (`'###'`) — the two cases are deliberately the same answer, and the caller
+ * decides what it means. In the Screens page's slug field an empty string
+ * means "no address typed"; in a starter template it meant "home". Two
+ * meanings for one value is what made AGL-1575 possible, so this function
+ * refuses to pick one: a caller that wants the root must say `'/'`.
+ *
+ * This docstring used to claim empty input normalized to the root, which the
+ * code has never done (see the spec, which asserts the `undefined`). A fix
+ * written against the docstring instead of the behaviour would have shipped
+ * broken; if you are here to make the code match the comment, do not.
+ */
+export declare function normalizeScreenSlug(input: string | null | undefined): string | undefined;
+/**
+ * Route segments the published site CANNOT serve, whatever the routing map
+ * says (AGL-2076).
+ *
+ * Every entry was measured against production on 2026-08-19 with a fresh
+ * `x-vercel-cache: MISS`, because a stale negative always passes. Two
+ * mechanisms, both of which resolve ahead of `[host]/[scheme]/[[...slug]]`:
+ *
+ *  - `404`, `500` — Next emits `pages/404.html` and `pages/500.html` even for
+ *    an app-router-only build, and the deployed filesystem answers them as
+ *    static files (`content-disposition: inline; filename="404"`,
+ *    `accept-ranges`, an `etag`, and no `x-nextjs-prerender`). The tenant
+ *    middleware's own CSP headers are on those responses, so the middleware
+ *    ran and its rewrite still lost. There is no rewrite this app can write
+ *    that beats them, which is why the honest fix is to refuse the slug at
+ *    authoring time rather than to keep promising an address that is dead.
+ *  - `search`, `api`, `_next`, `_static` — routes and exclusions the tenant
+ *    app owns. `/search` matches `app/[host]/[scheme]/search`; the other three are the
+ *    middleware matcher's exclusions, so no host rewrite happens at all and
+ *    the path is parsed as `[host]` with an empty slug.
+ *
+ * NOT reserved, and deliberately so, because measuring found them fine:
+ * `401`, `403`, `503` (`/401` returns 200 through the catch-all — two of the
+ * four designed error screens have always been viewable at their slugs, which
+ * is what made the `404` case look like a Next.js defect), and `index`.
+ *
+ * `fonts` and `examples` USED to be dead for the third reason and are not
+ * here: they were exclusions inherited from the Vercel platforms starter kit
+ * for directories `apps/tenant/public` has never contained. They are given
+ * back rather than reserved — see the matcher in `apps/tenant/middleware.ts`.
+ */
+export declare const RESERVED_SCREEN_ROUTE_SEGMENTS: readonly string[];
+/**
+ * The reserved segment a routing-map path would collide with, or `undefined`
+ * when the path is servable. The segment rather than a boolean, so a refusal
+ * can name what it is refusing over.
+ *
+ * Reads the FIRST segment only. For `api`/`_next`/`_static` that is exactly
+ * right — the middleware excludes the whole subtree, so `api/docs` is as dead
+ * as `api`. For `404`/`500`/`search` it over-refuses in principle (`/404/why`
+ * would serve), and that is deliberate: a child path only exists when its
+ * PARENT screen is published, the parent would sit at `404`, and this same
+ * rule refuses that. A second list to express a case that cannot arise would
+ * be a second list to keep correct.
+ *
+ * Takes a composed routing-map path (`about`, `company/about`, `'/'`), not a
+ * raw slug field — `'/'` is the home screen and is never reserved.
+ */
+export declare function reservedScreenRouteSegment(path: string | null | undefined): string | undefined;
+/**
+ * The refusal an author reads when {@link reservedScreenRouteSegment} names a
+ * segment. One sentence, in one place, because the Screens page, the version
+ * view and the besigner all have to say the same thing — three surfaces
+ * wording a platform constraint three ways is how AGL-2093's precheck drifted.
+ */
+export declare function reservedScreenRouteMessage(segment: string): string;
+/**
+ * Does a raw slug value carry a path separator the one segment it describes
+ * cannot hold (AGL-2572)?
+ *
+ * {@link normalizeScreenSlug} answers with a SINGLE segment, and it gets there
+ * by deleting every character outside `[a-z0-9-_]` — an interior `/` included.
+ * Whitespace becomes a `-`, but a `/` becomes nothing at all, so
+ * `alternatives/webflow` normalizes to `alternativeswebflow` and the two words
+ * are stored glued. Nothing about that is visible: the field accepts the
+ * keystroke, the composed path reads plausibly, and the screen ends up
+ * answering an address nobody typed. Two screens on `aglyn-marketing` carry a
+ * slug of that shape.
+ *
+ * So the separator is a question a slug field asks BEFORE normalizing, and the
+ * answer is a refusal rather than a silent repair — the caller cannot know
+ * which segment the author meant, and guessing is how the glue happened.
+ *
+ * LEADING and TRAILING slashes are not separators here. `normalizeScreenSlug`
+ * strips them deliberately (`/about/` → `about`), an author typing an address
+ * writes them without meaning a hierarchy, and refusing them would reject a
+ * value the normalizer has always handled correctly. `'/'` alone is the home
+ * page and is never a separator.
+ */
+export declare function screenSlugHasPathSeparator(input: string | null | undefined): boolean;
+/**
+ * The refusal an author reads when {@link screenSlugHasPathSeparator} answers
+ * true. One sentence, in one place, for the same reason
+ * {@link reservedScreenRouteMessage} is: the Screens page, the version view
+ * and the besigner all take a slug, and a constraint worded three ways drifts
+ * into three different constraints.
+ *
+ * It names the alternative, because a `/` in this field almost always means
+ * "nest me" and the hierarchy is what actually composes the longer path.
+ */
+export declare const SCREEN_SLUG_PATH_SEPARATOR_MESSAGE: string;
+/** Minimal screen shape the hierarchy helpers need. */
+export interface ScreenRouteNode {
+    slug?: string;
+    parentId?: ScreenUid;
+}
+/**
+ * Composes a screen's routing-map path from its own slug plus its ancestor
+ * chain: parent `company` + own `about` → `company/about`. Root (`'/'`)
+ * segments contribute nothing, so children of the home screen sit at the
+ * top level. Returns `undefined` when the screen (or any ancestor) has no
+ * slug, when the screen's own slug is `'/'` while it has a parent, or when
+ * the chain has a cycle / exceeds {@link MAX_SCREEN_DEPTH}.
+ */
+export declare function composeScreenRoutePath(screenId: ScreenUid, screensById: Record<ScreenUid, ScreenRouteNode | undefined>): string | undefined;
+/**
+ * All screens whose parent chain passes through `screenId` (children,
+ * grandchildren, …). Used to cascade routing-map rewrites when a screen's
+ * slug or parent changes.
+ */
+export declare function collectScreenDescendantIds(screenId: ScreenUid, screensById: Record<ScreenUid, ScreenRouteNode | undefined>): ScreenUid[];
+/**
+ * True when making `nextParentId` the parent of `screenId` would create a
+ * loop — i.e. the candidate parent is the screen itself or one of its own
+ * descendants.
+ */
+export declare function wouldCreateScreenCycle(screenId: ScreenUid, nextParentId: ScreenUid | undefined, screensById: Record<ScreenUid, ScreenRouteNode | undefined>): boolean;
+/** How {@link buildScreenRouteEntries} treats a screen with no entry today. */
+export interface BuildScreenRouteEntriesOptions {
+    /**
+     * Is this call a PUBLISH, or only a move? (AGL-2571)
+     *
+     * An entry in the host's `screens` map is the whole of what makes a path
+     * reachable, so writing one IS publishing — there is no second switch. This
+     * function is used for both, and until now it could not tell them apart: it
+     * wrote a path for every screen in the subtree whose slug chain happened to
+     * resolve, whether or not anyone had asked for that screen to go live.
+     *
+     * What that produced on `aglyn-marketing`: assigning a parent to an
+     * unpublished screen that merely CARRIED a slug registered it in the map,
+     * and the besigner toolbar — which reads the map — then offered `Unpublish`
+     * and an enabled `Live` link for a screen nobody had published, with no
+     * `publishedAt` on the document and a 404 at the address.
+     *
+     * `false` therefore restricts the write to entries that ALREADY exist:
+     * paths are rewritten, broken ones are removed, and nothing new goes live.
+     * A move is not an activation.
+     *
+     * Defaults to `true`, which is what the publish buttons mean.
+     */
+    publish?: boolean;
+}
+/**
+ * Routing-map entries for a screen plus all its descendants under a
+ * candidate screens map: a composed path sets the entry, `null` marks an
+ * existing entry whose chain no longer resolves for removal. Callers apply
+ * the result in one write so slug/parent changes cascade atomically.
+ *
+ * See {@link BuildScreenRouteEntriesOptions.publish} for the difference
+ * between publishing a screen and merely moving one.
+ */
+export declare function buildScreenRouteEntries(screenId: ScreenUid, screensById: Record<ScreenUid, ScreenRouteNode | undefined>, routingMap: Record<ScreenUid, string> | null | undefined, options?: BuildScreenRouteEntriesOptions): Record<ScreenUid, string | null>;
+/**
+ * Looks up which screen currently owns a routing path, for sibling-slug
+ * uniqueness checks before publishing.
+ */
+export declare function findScreenIdByRoutePath(screens: Record<ScreenUid, string> | null | undefined, path: string): ScreenUid | undefined;
+/** Human-facing URL for a routing-map path (`'/'` stays `/`, `about` → `/about`). */
+export declare function screenRoutePathToUrl(path: string): string;
+/**
+ * A screen's OWN slug segment, read back off its composed routing-map path
+ * (`company/about` → `about`) — the inverse of the one step
+ * {@link composeScreenRoutePath} adds for this screen (AGL-2572).
+ *
+ * The slug field holds one segment; the routing map holds the whole composed
+ * path. A field that seeds itself from the map therefore has to take the last
+ * segment, and the besigner's seeded the whole path instead. Its value then
+ * went through {@link normalizeScreenSlug}, which DELETES an interior `/`
+ * rather than separating on it, so `alternatives/webflow` came back as the
+ * glued `alternativeswebflow` and was stored as the screen's own slug —
+ * without anyone typing it. Two screens on `aglyn-marketing` carry a slug of
+ * that shape.
+ *
+ * Not a general path parser: it answers only "which segment is this screen's
+ * own", which is always the last one.
+ */
+export declare function ownScreenSlugFromRoutePath(path: string | null | undefined): string | undefined;
+/** What a screen is ACTUALLY served at, where that is not its own slug. */
+export interface LinkableScreenRouteSources {
+    /**
+     * screen id → the routing-map path the screen really answers on, for
+     * screens routed by something other than their own slug. Today that is a
+     * content collection's LIST template: `/{collectionSlug}` renders it
+     * (`composeCollectionTemplatePage`), whatever slug it was published under.
+     * Applied LAST, so it also un-drops a screen named in `unrouted`.
+     */
+    routedElsewhere?: Record<string, string> | null | undefined;
+    /**
+     * Screen ids that serve NO address of their own — a collection ENTRY
+     * template (AGL-1267), commerce's PDP/catalog templates (AGL-1270), a
+     * screen that says `kind: 'template'` (AGL-1400). The tenant router drops
+     * these before matching, so their routing-map path is a 404.
+     */
+    unrouted?: Iterable<string> | null | undefined;
+    /**
+     * collection id → collection slug, for every CONTENT collection with a slug
+     * (AGL-2799): the collections whose listing `/{collectionSlug}` serves. Each
+     * becomes a link target of its own, keyed by `formatCollectionLinkValue`,
+     * so a link can name a blog's listing whether or not a list template screen
+     * renders it.
+     */
+    collectionListings?: Record<string, string> | null | undefined;
+    /**
+     * entry key (`entry:<collectionId>/<entryId>`) → the entry's path, for the
+     * entries a page references (AGL-3118). Unlike every other source, this one
+     * is not the whole site: a renderer reads only the entries its own links
+     * name, so an entry key missing here means "not asked for", and
+     * `screenRoutesAnswerFor` judges entry links only against a map that has
+     * some.
+     */
+    entryRoutes?: Record<string, string> | null | undefined;
+}
+/**
+ * Where a collection's RSS feed is served, in the routing map's own format
+ * (no leading slash): `blog/rss.xml` for the collection at `/blog`.
+ */
+export declare function collectionFeedRoutePath(collectionSlug: string): string;
+/**
+ * Where a collection ENTRY is served, in the routing map's own format (no
+ * leading slash): `blog/hello` for the entry slugged `hello` in the
+ * collection at `/blog` — the `/{collection}/{entry}` shape
+ * `parseCollectionRoute` reads back.
+ */
+export declare function collectionEntryRoutePath(collectionSlug: string, entrySlug: string): string;
+/**
+ * The routing map as a LINK TARGET table (AGL-1998).
+ *
+ * The host's `screens` map is written by publishing, one entry per screen
+ * under its own composed slug, and the tenant router then edits it at serve
+ * time: template screens are dropped (see {@link LinkableScreenRouteSources})
+ * and a collection's list template is reached at `/{collectionSlug}` instead.
+ * Everything that RESOLVES a screen link — `resolveScreenHref` on the live
+ * site, the besigner's screen picker, every `Link`-typed component prop — was
+ * reading the raw map, so the two surfaces disagreed with the router in both
+ * directions at once:
+ *
+ *  - `/blog` was UNOFFERABLE. The blog's list template is published under
+ *    `blog-list-template` and serves `/blog`, so no screen link on aglyn.com
+ *    could point at the site's own blog index.
+ *  - Every template screen was offered at a path that 404s, and an author who
+ *    picked one got a dead anchor that looks exactly like a live one.
+ *
+ * So the map every linking surface reads is derived here, from the same two
+ * facts the router uses, and the picker can no longer offer a path the router
+ * would refuse.
+ *
+ * A content collection's LISTING is a target in its own right (AGL-2799),
+ * keyed `collection:<collectionId>` beside the screen ids. Its path comes from
+ * `collectionListUrl`, the builder the pager, the canonical link and the
+ * sitemap already share, so a link cannot give a listing an address the rest
+ * of the site does not; and because the key is the collection's id rather than
+ * its slug, renaming the slug moves every link to it on the next render. A
+ * collection that is gone, or that has lost its slug, has no entry — the same
+ * absence `isScreenLinkBroken` reads as a dead target.
+ *
+ * A raw map with no source given comes back unchanged; an absent map with no
+ * overrides and no listings stays absent, because `undefined` means "nothing
+ * resolves yet" to {@link ScreenLinkContext} and `{}` would mean "resolved:
+ * nowhere".
+ */
+export declare function linkableScreenRoutes(screens: Record<ScreenUid, string> | null | undefined, sources?: LinkableScreenRouteSources): Record<ScreenUid, string> | undefined;
+/**
+ * `kind` of a besigner email document (AGL-395): a screen authored on the
+ * Emails page and sent by a campaign, never served at a URL.
+ */
+export declare const SCREEN_KIND_EMAIL = "email";
+/**
+ * `kind` of a collection ENTRY template (AGL-1400): a screen that composes
+ * `/{collection}/{entry}` for every entry and has no address of its own.
+ *
+ * The value exists so that "is this screen a page?" stops being a JOIN.
+ * Four issues (AGL-1173, AGL-1383, AGL-1387, AGL-1390) were the same sentence
+ * at different depths because the answer was derived from a mutable field on a
+ * DIFFERENT document — a collection's `entryScreenId` — and every new way to
+ * edit that pointer was a new bypass of `screensPerHost`. Here the fact is a
+ * property of the screen, server-stamped by /api/hosts/screens and frozen in
+ * the rules exactly as `kind: 'email'` has been since AGL-1383.
+ *
+ * NOT a collection's LIST template. `/{collectionSlug}` renders that exact
+ * screen with its entries, so it is a designed reachable page and AGL-1387
+ * made it count; converting one would take the page off the site.
+ */
+export declare const SCREEN_KIND_TEMPLATE = "template";
+/**
+ * `kind` of a designed ERROR screen (AGL-2092): the screen a host assigns to
+ * one of `errorScreens`' four status slots, rendered on paths that did not
+ * match — and so, like an entry template, at no address of its own.
+ *
+ * the decision on 2026-08-18 was that these must not spend the plan's screen
+ * allowance, for the reason AGL-1173 excluded an entry template: the governing
+ * question is whether the screen occupies a URL of its own, and a 404 body does
+ * not. (A collection's LIST template does — `/{collectionSlug}` renders that
+ * exact screen — which is why AGL-1387 left it counting.)
+ *
+ * ## This value does NOT get `kind: 'template'`'s trust
+ *
+ * `billableScreenIds` short-circuits on `SCREEN_KIND_TEMPLATE` BEFORE it
+ * consults the routing map, because a template is routed on purpose. An error
+ * screen is not, so it deliberately falls through to the ordinary rule and
+ * **the routing map outranks this field**: a screen still published at an
+ * address of its own counts, whatever it says about itself. That is the whole
+ * migration story for the error screens that already exist — see
+ * {@link ERROR_SCREEN_MAX_PER_HOST}.
+ *
+ * Stamped only by /api/hosts/screens, as part of binding the screen to a slot,
+ * and frozen against the client by the same rule that has frozen `kind` since
+ * AGL-1383. Nothing about that is new for this value; what IS new is that the
+ * stamp is BOUNDED, because unlike a template there is no natural limit on how
+ * many screens somebody would like to declare as error pages.
+ */
+export declare const SCREEN_KIND_ERROR = "error";
+/**
+ * How many screens one host may hold with `kind: 'error'` (AGL-2092).
+ *
+ * An exclusion the metered party can ask for is a free-screen generator unless
+ * something bounds it, and the four bindings in `HostErrorScreens` are the
+ * bound: at most one exempt screen per error slot. It is `.length` of the slot
+ * list rather than a literal `4` so the two can never disagree.
+ *
+ * Checked in TWO places, and the second is the one that makes the number true
+ * (AGL-2093). At the moment of the STAMP, against the post-state — the
+ * AGL-1390 shape, which produces the refusal worth reading because it can name
+ * the four slots. And inside `billableScreenIds` itself, where a live error
+ * screen past this bound is counted as an ordinary page: the stamp is not the
+ * only writer, and a crafted import bundle carrying `kind: 'error'` on all 200
+ * of its screens passed no bound at all until the rule owned one. An exclusion
+ * enforced only at the write paths somebody remembered is the sentence this arc
+ * has now repeated six times.
+ *
+ * The tempting alternative is to enforce "bound to a slot" as an
+ * invariant, promoting a screen back to a page whenever its slot is cleared;
+ * that is AGL-1390's refuse-the-clear bug wearing a different hat, because
+ * promotion raises the count and a host at its cap could then never unassign an
+ * error screen. So clearing a slot is always allowed and leaves the screen an
+ * error screen — unbilled, unrouted, and one deliberate click from being a page
+ * again — and what is bounded is how many of them can exist at once.
+ */
+export declare const ERROR_SCREEN_MAX_PER_HOST: 4;
+/** The two self-describing fields {@link screenClaimsToBeAPage} reads. */
+export interface ScreenPageClaim {
+    kind?: string;
+    deletedAt?: unknown;
+}
+/**
+ * Whether a screen document claims to be a page of the site at all (AGL-1383).
+ *
+ * Four things say it is not: `deletedAt` (soft-deleted — delete stamps the
+ * field rather than removing the doc), `kind: 'email'` (an Emails-page
+ * document, which has no URL and is rendered only by the campaign sender,
+ * straight off the doc), `kind: 'template'` (a collection entry template,
+ * which composes `/{collection}/{entry}` and has no address of its own —
+ * AGL-1400) and `kind: 'error'` (a screen assigned to one of the host's four
+ * error slots, rendered on paths that matched nothing — AGL-2092). All four
+ * are also the exclusions `countBillableScreens` subtracts before enforcing
+ * `screensPerHost`, and since AGL-1400 they are the WHOLE of that subtraction:
+ * the count reads no other collection.
+ *
+ * Adding the fourth needed no edit to `billableScreenIds` and no edit to
+ * `nonPageScreenIds`, which is the property AGL-1439 wrote them for: the
+ * routing-map override applies to a new non-page `kind` by default, and the
+ * flat infrastructure cap is the complement, so there is no second list to
+ * remember. Only `kind: 'template'` is named anywhere else, because only it
+ * opts OUT of the routing-map override.
+ *
+ * That is exactly why this is ONE function with TWO callers rather than two
+ * matching filters. Both fields are ordinary client-writable fields on
+ * `hosts/{hostId}/screens/{screenId}`, and until AGL-1383 only the count read
+ * them: an editor on a Free site could `updateDoc(screenRef, {kind: 'email'})`
+ * and the screen stopped counting against the plan while the routing map still
+ * pointed at it and the runtime still served it — a live page, for free. An
+ * exclusion is only sound if an excluded screen genuinely is not a page, so the
+ * serve path asks this question too, and flipping either field now costs the
+ * page instead of the plan.
+ *
+ * A CLAIM, not the answer, for the two fields a client can still reach: the
+ * routing map decides reachability, and `countBillableScreens` trusts
+ * `deletedAt` / `kind: 'email'` only for screens the map does not route.
+ * `kind: 'error'` is in the client-can-still-reach group for counting purposes
+ * even though no client writes it either: an error screen that is ALSO
+ * published at an address is a page somebody is using as a page, so the map
+ * outranks the stamp and it counts. Serving is the other way round — see
+ * `getScreen`, which returns error screens rather than 404ing them, because
+ * refusing would break the customer's own designed error page without closing
+ * anything (they are already paying for it if it is routed).
+ *
+ * `kind: 'template'` is the exception, and the reason it can be one is that no
+ * client writes it — /api/hosts/screens stamps it, demotion lowers the count
+ * and promotion is checked exactly like a create, so there is no toggle to
+ * launder. A template is ROUTED on purpose (publishing is how the compose
+ * pipeline picks it up), which is why the map cannot be the arbiter for it.
+ */
+export declare function screenClaimsToBeAPage(screen: ScreenPageClaim | null | undefined): boolean;
+/**
+ * One screen, reduced to the two things {@link billableScreenIds} asks about.
+ *
+ * Rows rather than a query (AGL-1390) because the enforcement points ask about
+ * a state that does not exist yet: a promotion is checked against the count it
+ * WOULD leave behind. A function that does its own reads can only answer for
+ * the present. `readScreenSources` in /api/hosts/resources is the reader.
+ */
+export interface BillableScreenSource {
+    id: string;
+    kind?: unknown;
+    deletedAt?: unknown;
+}
+/**
+ * A host's `screens` routing map (screen id → route path) as
+ * `publishScreenRoute` writes it. Only membership is read, never the path.
+ */
+export type ScreenRoutingMap = Record<string, unknown> | null | undefined;
+/**
+ * WHICH screens spend the plan's screen allowance (AGL-1173).
+ *
+ * The ids rather than the count, because an enforcement point that cannot name
+ * the screen it is refusing over is a refusal the person reading it cannot act
+ * on.
+ *
+ * A plain `screens.count()` charged for three things the subscriber never
+ * chose to author, and the screens list — which filters them out — then
+ * disagreed with the server about how much of the plan was used:
+ *
+ *  - **Soft-deleted screens.** Delete stamps `deletedAt` rather than
+ *    removing the doc, so deleting a screen never freed a slot. On the free
+ *    plan (5 screens) that was a dead end with no way out from the UI.
+ *  - **Email screens** (`kind: 'email'`), which live on the Emails page and
+ *    were already excluded from the list count but not from enforcement.
+ *  - **A collection's ENTRY template** — one screen serving every entry at no
+ *    URL of its own. Adding a blog cost two of the free plan's five screens
+ *    before the first page existed.
+ *
+ * ## Why it lives HERE, next to the predicate (AGL-2093)
+ *
+ * It was in /api/hosts/resources, and the console's Screens page — which runs
+ * the same rule as a precheck before it lets somebody add a page — could not
+ * import it from there (that module reaches `@aglyn/aglyn/server`, which pulls
+ * `node:stream` in through the API adapter). So the page restated it, and the
+ * restatement drifted: it never learned the error-screen bound below, and on a
+ * host holding five `kind: 'error'` screens the console offered room the API
+ * then refused. A precheck that warns on a different number than the API
+ * enforces is worse than no precheck at all, and a rule that has to be
+ * restated to be reused will be restated wrongly eventually. One function, two
+ * callers.
+ *
+ * ## ONE document answers it now (AGL-1400)
+ *
+ * The third exclusion used to be a JOIN: the count read the `collections`
+ * collection and subtracted whatever screen ids the template pointers named.
+ * That shape produced four issues in one arc, because the other side of the
+ * join is editable and each new way to edit it was a new bypass — AGL-1173
+ * created the exclusion, AGL-1383 froze the two fields the screen itself
+ * carried, AGL-1387 found the list template was a page after all, and AGL-1390
+ * found the pointer could be toggled to mint permanent slots and had to move
+ * the write server-side and evaluate the cap against the post-state.
+ *
+ * Since AGL-1400 an entry template says so on its own document
+ * (`kind: 'template'`), stamped by /api/hosts/screens and frozen in the rules
+ * exactly as `kind: 'email'` is. So this reads the `screens` collection and
+ * nothing else: the pointer is a pointer again, freely writable, and it excuses
+ * nothing. `screenClaimsToBeAPage` is the whole rule, which is the property
+ * that keeps the count and the serve path (`getScreen`) from ever disagreeing.
+ *
+ * What did NOT change is what anybody pays. A collection's LIST template stays
+ * a page — `/{collectionSlug}` renders that exact screen (AGL-1387) — so it is
+ * never stamped and still counts, and an entry template stays excluded
+ * (AGL-1173's charge is not reinstated).
+ *
+ * ## The routing map outranks the document, and ONLY ADDS (AGL-1383, AGL-1445)
+ *
+ * `deletedAt` and `kind: 'email'` are ordinary client-writable fields on the
+ * screen's OWN document, and the party they are subtracted on behalf of is the
+ * party being metered. `updateDoc(screenRef, {kind: 'email'})` — one write, no
+ * route change — used to take a live page off a Free plan's five and leave it
+ * serving, because nothing else read either field. So for those two the
+ * document's account of itself is consulted only for screens the routing map
+ * does not route: **a routed screen counts, whatever it says about itself**.
+ *
+ * The converse is NOT true and must not become true, which is AGL-1445's
+ * settlement. An unrouted screen with no slug — a draft, or an orphan left
+ * behind by an unpublish — serves nothing, and it still counts:
+ *
+ *  - `screensPerHost` is a CREATE-TIME gate. /api/hosts/resources counts and
+ *    refuses; `report-usage` re-measures monthly but RECORDS rather than
+ *    enforces. The create is the only door.
+ *  - A screen is born unrouted, and publishing is a separate later act — a
+ *    client `updateDoc` on the host's `screens` map (`publishScreenRoute`),
+ *    gated on `canPublishHostContent` and on nothing else.
+ *
+ * So subtracting unrouted screens would not make the count fairer, it would
+ * make every create free — the new document is unrouted by construction — and
+ * leave the publishes that follow ungated. A Free site would hold unlimited
+ * pages. AGL-1445's option 2 (server-stamping a non-page `kind` on orphans by
+ * a sweep) lands in the same place unless the promotion BACK passes a gate,
+ * and for an orphan the promotion is the publish. `kind: 'template'` can be
+ * trusted precisely because its promotion is `convertScreenKind`, checked
+ * exactly like a create; there is no equivalent door in front of the routing
+ * map. The price is one slot until somebody deletes the orphan, which is one
+ * click and which the Screens page lists like any other row —
+ * `unrouted-screen-counts.spec.ts` pins it at the enforcement point.
+ *
+ * `kind: 'template'` is deliberately outside the map's override, and it is the
+ * one value that can be. A template is routed ON PURPOSE — publishing is how
+ * the compose pipeline picks it up, and AGL-1267 drops it from routing when the
+ * request arrives — so the map cannot be the arbiter for it. What makes
+ * trusting it safe is that no client writes it: the demotion that sets it
+ * always succeeds (it lowers the count) and the promotion that clears it is
+ * checked exactly like a create (it raises it), so there is no toggle to run a
+ * loop through.
+ */
+export declare function billableScreenIds(screens: ReadonlyArray<BillableScreenSource>, routingMap?: ScreenRoutingMap): Set<string>;
+/**
+ * The live `kind: 'error'` screens whose billing exemption actually HOLDS —
+ * at most {@link ERROR_SCREEN_MAX_PER_HOST} of them (AGL-2093).
+ *
+ * ## Why the bound lives here and not at one endpoint
+ *
+ * AGL-2092 introduced the exemption and bounded it at four — one per
+ * `HostErrorScreens` slot — by checking the post-state at the moment
+ * /api/hosts/screens STAMPS the kind. That made the claim true of the assign
+ * route and false end to end, because the stamp is not the only writer:
+ * `SITE_EXPORT_FIELDS.screens` carries `kind`, so a hand-edited bundle could
+ * declare `kind: 'error'` on all 200 of its screens and /api/hosts/import
+ * excluded every one of them from `screensPerHost` — never passing the bound,
+ * never binding a slot. N unbilled screens, one file.
+ *
+ * That is the arc's fifth repetition of one sentence (AGL-1173, AGL-1383,
+ * AGL-1387, AGL-1390, AGL-1400): *an exclusion enforced at a write path is
+ * enforced only at the write paths somebody remembered.* So the bound moves
+ * into the rule the enforcement points SHARE. However a screen came to say
+ * `kind: 'error'` — assign, import, or the next writer — a host gets four
+ * unbilled ones and the rest spend the plan's allowance exactly as pages do.
+ *
+ * ## Which four
+ *
+ * The lowest ids, sorted. The choice is arbitrary and the DETERMINISM is not:
+ * every enforcement point models a post-state by re-running this over a
+ * different row set, and a rule that picked "the first four encountered" would
+ * hand two callers different answers for the same host. The SIZE of the exempt
+ * set never depends on the order — it is `min(live errors, 4)` — so only which
+ * ids a refusal message names can move, never whether one is refused.
+ *
+ * Tombstones are excluded for `nonPageScreenIds`' reason: a cap counting
+ * soft-deleted rows is AGL-1173's bug one cap over, where deleting an error
+ * screen never frees its slot.
+ *
+ * The stamp-time bound in /api/hosts/screens is deliberately KEPT. It is the
+ * better refusal — it names the four slots and arrives before the write — and
+ * this is the backstop under it, so no writer can mint a fifth unbilled error
+ * screen even by a path that never consults it.
+ */
+export declare function exemptErrorScreenIds(screens: ReadonlyArray<BillableScreenSource>): Set<string>;
+/**
+ * The screens the flat platform cap bounds (AGL-1399, AGL-1439): LIVE documents
+ * that are not billable pages — the exact complement of {@link billableScreenIds}.
+ *
+ * ## Why the complement, and not `kind === 'email' || kind === 'template'`
+ *
+ * Both issues are one sentence — *a `kind` value that excludes a document from
+ * billing, declarable by the metered party* — and a cap that enumerated the two
+ * values would be correct until the third arrived, which is how this arc has
+ * already gone four times (AGL-1173, AGL-1383, AGL-1387, AGL-1390). Written as
+ * the complement, a new non-page `kind` is bounded by the act of adding it to
+ * `screenClaimsToBeAPage`: there is no second list to remember.
+ *
+ * Two properties fall out of taking the complement rather than negating the
+ * predicate directly:
+ *
+ *  - **A soft-deleted screen is in NEITHER set.** Delete stamps `deletedAt`
+ *    rather than removing the doc, and a cap that counted tombstones would be
+ *    the AGL-1173 bug one cap over — deleting an email document would never free
+ *    a slot, and a host at the cap could never author another.
+ *  - **A ROUTED email document counts once, against the plan.** The routing map
+ *    outranks the document (AGL-1383), so such a screen is a page somebody is
+ *    already paying for; charging it to the infrastructure cap as well would be
+ *    counting one document twice. The two sets partition the live screens.
+ *
+ * Taking the complement is also what makes AGL-2093's bound arrive here for
+ * free: an error screen past `ERROR_SCREEN_MAX_PER_HOST` becomes billable
+ * above, so it leaves this set in the same act. The partition holds — every
+ * live screen is in exactly one of the two — and the fifth error screen is
+ * bounded by the PLAN rather than by the flat 5,000, which is the stronger of
+ * the two numbers on every finite tier.
+ */
+export declare function nonPageScreenIds(screens: ReadonlyArray<BillableScreenSource>, routingMap?: ScreenRoutingMap): Set<string>;
+/**
+ * How many LIVE screen documents a host may hold that are NOT billable pages —
+ * email documents, entry templates, and whatever non-page `kind` comes next
+ * (AGL-1399, AGL-1439).
+ *
+ * ## Why a flat number and not a plan dimension
+ *
+ * Every exclusion `screenClaimsToBeAPage` names is declared by the party being
+ * metered: `kind: 'email'` is on the create allow-list because the email
+ * composer must send it, and AGL-1400 gave `kind: 'template'` the same
+ * billing-excluding meaning. So `POST /api/hosts/resources` with
+ * `kind: 'email'` created a document that no cap counted, without limit, on a
+ * free plan — unbounded Firestore documents rather than a bypass of anything we
+ * sell. AGL-1439 is the same sentence one value over, and a cap enumerating the
+ * two values would leave the hole open for the third.
+ *
+ * The shape is `WEBHOOK_MAX_PER_HOST`'s (AGL-1360): no `OrgEntitlements` key, no
+ * variation by plan, counted from a server read at the create path. An
+ * `emailDocumentsPerHost` would have been a **pricing** decision — a number the
+ * price list has to explain and support has to defend — for a limit whose only
+ * job is to bound infrastructure. This is not the charge AGL-1173 removed and it
+ * does not reinstate it: nothing here is billable, and `screensPerHost` counts
+ * exactly what it counted yesterday.
+ *
+ * ## Why 5,000
+ *
+ * Sized to the heaviest library anybody could plausibly author, not to what
+ * exists: production's busiest host holds FOUR non-page screens (one email,
+ * three entry templates) across five hosts, so any bar sized to today's data
+ * would be meaningless. The real quantity is campaigns, which accumulate and are
+ * never deleted — a DAILY newsletter reaches 5,000 after thirteen years, and a
+ * thrice-weekly one after thirty-two. Entry templates contribute one per
+ * collection, so tens at the outside. Nothing a real customer does approaches
+ * this, which is the property that matters: the failure mode of a too-low flat
+ * cap is blocking legitimate work with an error the price list cannot explain.
+ *
+ * A cheaper bound is not worth much either. What the number defends against is a
+ * script in a loop, and a loop stopped at 5,000 documents is stopped. It also
+ * bounds the cost of the count itself — `countBillableScreens` scans the screens
+ * collection on every create at one billed read per document (AGL-1440), and
+ * before this cap that scan had no ceiling at all.
+ */
+export declare const NON_PAGE_SCREEN_MAX_PER_HOST = 5000;