@jini-ai/cms 0.5.2 → 0.5.4

package/dist/widgets/markers/marker.js

@@ -0,0 +1,780 @@
+/**
+ * @file The ONE parser for embed markers (`data-embed-config`), shared by every consumer.
+ *
+ * Purpose:
+ * Before the 2026-08-10 marker-spine unification there were four independent scanners, each with its
+ * own regex, matching two different attribute vocabularies that had drifted apart:
+ *
+ * - `widgets/html-embeds.ts` — `data-embed-type`/`data-embed-id` in a Page's `body_html`
+ * - `features/theme/static-render.ts` — the same pair for menu/post markers, PLUS
+ *   `data-tovu-slot`/`data-nav-current`/`data-slot-variant` for theme partials
+ * - `core/entry-refs/extractor.ts` — the same pair again, for reference extraction
+ *
+ * Four regexes for one concept is four places a fix has to land and four ways they can disagree
+ * (they already did: the widgets pipeline only matched a self-closing `<div></div>`, while the theme
+ * path deliberately matched a marker WITH authored fallback content inside it). This module is the
+ * single definition. Consumers ask it what a piece of HTML references; none of them owns a pattern.
+ *
+ * The vocabulary is one attribute carrying JSON:
+ *
+ *     <div data-embed-config='{"type":"partial","id":"nav","current":"index"}'></div>
+ *     <nav data-embed-config='{"type":"menu","id":"docs-nav","variant":"tree"}'>fallback</nav>
+ *
+ * Forms may request plain output with `{"type":"form","id":"contact","mode":"html"}`;
+ * `mode` is resolver-owned config, so the parser carries it through unchanged.
+ *
+ * `type` and `id` are ordinary keys, not separate attributes, so a new key never requires a new
+ * attribute name. The attribute is single-quoted precisely so the JSON's own double quotes need no
+ * escaping.
+ *
+ * Architectural role:
+ * PURE. No I/O, no DOM, no resolution — it reports what a marker *says*, never what it resolves to.
+ * Lives in `core/` because all three consumers above sit in different features and `core/entry-refs`
+ * (the integrity path) is already here.
+ *
+ * ## Why regex and not a parser
+ *
+ * Same reason the code it replaces used one: this runs server-side against theme-authored markup and
+ * Page bodies with no DOM available, and every consumer needs the marker's *source offsets* to
+ * substitute in place. A real HTML parse would also have to round-trip the document back to a string
+ * byte-identically, which is a much harder guarantee than locating one attribute.
+ */
+/**
+ * Matches the OPEN tag of one element carrying `data-embed-config`, capturing its tag name and its
+ * attribute text. It stops at the open tag's own `>` and makes no assumption about what closes it —
+ * {@link scanEmbedMarkers} locates the matching close tag separately, via {@link findBalancedClose}'s
+ * depth count over the masked html.
+ *
+ * Before 2026-09-23 this pattern captured inner content itself, up to a backreferenced `<\/\1>`. That
+ * was only safe for a marker with no same-named descendant — true for every marker convention in this
+ * codebase at the time, and the same assumption `injectMenuEmbed` documented before this module
+ * existed — but it meant `<div data-embed-config='...'><div>…</div></div>` (an inner element sharing
+ * the marker's own tag name) closed at the FIRST `</div>`, truncating the marker's real inner content.
+ * {@link findBalancedClose} removes that assumption by counting depth instead of matching nearest text.
+ *
+ * The tag name pattern (`[a-z][a-z0-9-]*`) accepts any HTML element name, including a custom element
+ * (`<my-card>`), not just the historical letters-only set.
+ *
+ * Carries the `d` (`hasIndices`) flag so {@link scanEmbedMarkers} can read every field back out of
+ * the ORIGINAL html at each capture group's own offsets, rather than out of whatever string was
+ * actually scanned (see {@link maskNonRenderableRegions}) — see that function's doc for why the two
+ * must never be the same string.
+ *
+ * Accepts the value single-quoted (group 3, the authored convention) OR double-quoted (group 4).
+ * Before 2026-09-26 only the single-quoted form matched, but every DOM serializer — the admin's
+ * Interactive canvas (GrapesJS), `innerHTML`, parse5 — re-emits the SAME attribute as
+ * `data-embed-config="{&quot;type&quot;:…}"`, so one canvas save silently turned a working marker into
+ * inert markup (the live /pages "That marker, live:" contact-form example rendered an empty box).
+ * A double-quoted value is HTML-decoded before parsing ({@link decodeAttributeValue}); a single-quoted
+ * one is still read verbatim, as it always was.
+ */
+const MARKER_PATTERN = /<([a-z][a-z0-9-]*)((?:\s+[^>]*?)?\sdata-embed-config=(?:'([^']*)'|"([^"]*)")(?:\s+[^>]*?)?)\s*>/gid;
+/** The `data-embed-config` attribute itself, either quoting — for the rebuild helpers that rewrite or
+ *  drop it inside a marker's {@link EmbedMarker.attrs}. Group 1 is set only for the double-quoted form. */
+const MARKER_ATTRIBUTE_PATTERN = /\s+data-embed-config=(?:'[^']*'|("[^"]*"))/;
+/** Named character references a serializer actually emits inside an attribute value. Any other named
+ *  reference is left as written rather than guessed at. */
+const NAMED_CHARACTER_REFERENCES = { quot: '"', amp: "&", apos: "'", lt: "<", gt: ">", nbsp: " " };
+/**
+ * A double-quoted attribute value as a browser reads it: `&quot;`/`&amp;`/`&apos;`/`&lt;`/`&gt;`/
+ * `&nbsp;` and decimal/hex numeric references decoded. Kept local (this module stays import-free).
+ *
+ * @complexity O(n) over `value`.
+ */
+function decodeAttributeValue(value) {
+    return value.replace(/&(?:#(\d{1,7})|#[xX]([0-9a-fA-F]{1,6})|([a-z]+));/g, (whole, dec, hex, name) => {
+        if (name !== undefined)
+            return NAMED_CHARACTER_REFERENCES[name] ?? whole;
+        const codePoint = dec !== undefined ? Number.parseInt(dec, 10) : Number.parseInt(hex, 16);
+        return codePoint > 0 && codePoint <= 0x10ffff ? String.fromCodePoint(codePoint) : whole;
+    });
+}
+/** The inverse of {@link decodeAttributeValue} for a value written back double-quoted. */
+function encodeDoubleQuotedValue(value) {
+    return value.replace(/&/g, "&amp;").replace(/"/g, "&quot;");
+}
+/** Every character of `text` replaced by a space, except newlines (left alone so a masked span
+ * cannot change how many lines the surrounding string has). Same length in, same length out. */
+function blank(text) {
+    return text.replace(/[^\n]/g, " ");
+}
+/** An HTML comment, open to close, non-greedy so two separate comments never merge into one span. */
+const HTML_COMMENT_PATTERN = /<!--[\s\S]*?-->/g;
+/** A `<script>` or `<style>` element, capturing its tag name (for the backreferenced close) and its
+ * raw-text content — the part a browser never parses as markup. */
+const RAW_TEXT_ELEMENT_PATTERN = /<(script|style)\b[^>]*>([\s\S]*?)<\/\1>/gi;
+/**
+ * Produce a same-length copy of `html` with every HTML comment, and the raw-text content of every
+ * `<script>`/`<style>` element, replaced by space filler. {@link scanEmbedMarkers} runs
+ * {@link MARKER_PATTERN} against this copy — never the original — so a well-formed
+ * `data-embed-config` marker written inside an authoring note (`<!-- ... -->`) or inside a
+ * `<style>` block's CSS comment can never be matched: neither is ever parsed as an element by a
+ * browser, so the scanner must not treat either as one either. A marker outside both is untouched
+ * here and matches exactly as before.
+ *
+ * Same length is load-bearing, not cosmetic: every offset {@link scanEmbedMarkers} reports (and
+ * {@link substituteMarkers}'s index-based splice back into the ORIGINAL html) depends on a masked
+ * span occupying exactly the same positions as what it replaces. Deleting the comment/raw-text
+ * content instead of blanking it would shift every later offset out from under those callers.
+ *
+ * EXPORTED (2026-09-09) for `features/pages/regions.ts`, which scans the SAME page bodies for
+ * `data-agent-element` region elements and needs the identical "text inside a comment or inside a
+ * <style> block is not markup" rule with the identical offset-preserving guarantee. A second copy
+ * of this masking is exactly the four-scanners-for-one-concept drift this module's own header
+ * exists to close: region scanning is a different question about the same bytes, not a second
+ * vocabulary.
+ */
+export function maskNonRenderableRegions({ html }, _optional = {}) {
+    const withoutComments = html.replace(HTML_COMMENT_PATTERN, blank);
+    return withoutComments.replace(RAW_TEXT_ELEMENT_PATTERN, (whole, tagName, content) => {
+        const closingTagLength = tagName.length + 3; // "</" + tagName + ">"
+        const openTagLength = whole.length - content.length - closingTagLength;
+        return whole.slice(0, openTagLength) + blank(content) + whole.slice(openTagLength + content.length);
+    });
+}
+/**
+ * One capture group's `[start, end)` offsets from a `d`-flagged match's `indices` array. Throws
+ * rather than returning `undefined` for a group `MARKER_PATTERN` never leaves unmatched — every
+ * group here sits on a mandatory part of the pattern, never inside an optional alternation, so a
+ * missing entry means the pattern changed underneath this function, not a normal runtime path.
+ */
+function requireGroupRange(indices, group) {
+    const range = indices[group];
+    if (!range) {
+        throw new Error(`marker.ts: expected capture group ${group} of MARKER_PATTERN to participate in the match`);
+    }
+    return range;
+}
+/**
+ * Validate one marker's raw attribute text. Split out from {@link scanEmbedMarkers} so the scan stays
+ * a plain loop: all three rejection branches live here, and the caller only chooses which list to
+ * push onto.
+ *
+ * EXPORTED (2026-09-23, Bug A / interactive-bugs plan Slice A1) so a consumer that only has the raw
+ * `data-embed-config` string in hand — the admin's canvas placeholder describer, which reads an
+ * already-authored DOM attribute rather than re-running {@link scanEmbedMarkers} — can apply the exact
+ * same parse-and-reject rule `scanEmbedMarkers` itself uses, instead of growing its own copy of it.
+ */
+export function parseEmbedMarkerConfig({ raw }, _optional = {}) {
+    let parsed;
+    try {
+        parsed = JSON.parse(raw);
+    }
+    catch (err) {
+        return { problem: { kind: "invalid-json", raw, message: err.message } };
+    }
+    if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) {
+        return { problem: { kind: "not-an-object", raw } };
+    }
+    const config = parsed;
+    if (typeof config.type !== "string" || config.type === "") {
+        return { problem: { kind: "missing-type", raw } };
+    }
+    return { config };
+}
+/**
+ * Every `<tagName`/`</tagName` token in `masked` from `fromIndex` onward, in document order. The
+ * boundary lookahead (`[\s/>]`) stops `<my-card` matching a longer name like `<my-card-extra`; each
+ * token ends at its first `>`, and the scan resumes there, so a token's own attribute text is never
+ * re-read as another token. Stops at a trailing malformed tag with no `>` at all.
+ *
+ * Runs against the MASKED copy only, so a tag written inside a comment or a `<script>`/`<style>`
+ * element's raw text (already blanked by {@link maskNonRenderableRegions}) is invisible here,
+ * exactly as it is invisible to {@link MARKER_PATTERN} itself.
+ *
+ * @complexity O(n) over the remaining length of `masked`.
+ */
+function* tagTokens(masked, tagName, fromIndex) {
+    const token = new RegExp(`<(/?)${tagName}(?=[\\s/>])`, "gi");
+    token.lastIndex = fromIndex;
+    for (let match = token.exec(masked); match !== null; match = token.exec(masked)) {
+        const closeAngle = masked.indexOf(">", match.index);
+        if (closeAngle === -1)
+            return;
+        yield { isClose: match[1] === "/", span: [match.index, closeAngle + 1] };
+        token.lastIndex = closeAngle + 1;
+    }
+}
+/**
+ * The depth-balanced close for the open tag whose match ended at `fromIndex`, by counting
+ * {@link tagTokens} from there: `<div><div>…</div></div>` closes at the tag that actually balances the
+ * one already open, not at the first `</div>` found. Falls back to the first close found when the
+ * region never rebalances (a genuinely unclosed inner tag) — the result the old backreferenced pattern
+ * produced, so malformed markup degrades as it did before. `null` when there is no close at all.
+ *
+ * Only {@link findBalancedClose}'s fallback for an open tag {@link indexTagTokens} tokenized
+ * differently (its JSON config holds a `>`); every other lookup is answered from that index.
+ *
+ * @complexity O(n) over the remaining length of `masked`.
+ */
+function scanForBalancedClose(masked, tagName, fromIndex) {
+    let depth = 1;
+    let firstClose = null;
+    for (const { isClose, span } of tagTokens(masked, tagName, fromIndex)) {
+        if (!isClose) {
+            depth += 1;
+            continue;
+        }
+        firstClose ??= span;
+        depth -= 1;
+        if (depth === 0)
+            return span;
+    }
+    return firstClose;
+}
+/**
+ * Pairs every `<tagName`/`</tagName` token in `masked` in ONE pass. Stack pairing gives each open the
+ * same close {@link scanForBalancedClose}'s depth count from that open would find (both match a close
+ * with the nearest still-open tag), but for every open at once — so a page of k unbalanced markers
+ * costs O(n), not the O(k·n) of rescanning to the end of the document once per marker.
+ *
+ * @complexity O(n) time, O(t) space for t tokens.
+ */
+function indexTagTokens(masked, tagName) {
+    const partners = new Map();
+    const closes = [];
+    const openStarts = [];
+    for (const { isClose, span } of tagTokens(masked, tagName, 0)) {
+        if (!isClose) {
+            partners.set(span[0], null);
+            openStarts.push(span[0]);
+            continue;
+        }
+        closes.push(span);
+        const opener = openStarts.pop();
+        if (opener !== undefined)
+            partners.set(opener, span);
+    }
+    return { partners, closes };
+}
+/** The first span in `closes` (sorted by start) starting at or after `fromIndex`, or `null`.
+ * @complexity O(log c) — binary search. */
+function firstCloseFrom(closes, fromIndex) {
+    let low = 0;
+    let high = closes.length;
+    while (low < high) {
+        const mid = (low + high) >>> 1;
+        // low < high keeps mid within the sorted closes array.
+        if (closes[mid][0] < fromIndex)
+            low = mid + 1;
+        else
+            high = mid;
+    }
+    return closes[low] ?? null;
+}
+/**
+ * The `[start, end)` span (in `masked`) of the close tag that balances the marker open tag spanning
+ * `[openStart, openEnd)` — see {@link scanForBalancedClose} for the balancing and first-close
+ * fallback rules — or `null` when no close exists at all (a void element such as `<img>`, or an
+ * unclosed marker): the caller then treats the open tag alone as the whole marker, rather than
+ * letting it swallow the rest of the document.
+ *
+ * Answers from a per-tag-name {@link indexTagTokens} built once per scan and kept in `cache`. When
+ * that index tokenized this open tag differently from the marker pattern (the index ends a tag at its
+ * first `>`, and a `>` inside the JSON config ends it early), it falls back to the exact linear count.
+ *
+ * @complexity O(log c) amortized once the tag name's O(n) index exists; O(n) on the fallback.
+ */
+function findBalancedClose(masked, tag, cache) {
+    const key = tag.name.toLowerCase();
+    let index = cache.get(key);
+    if (index === undefined) {
+        index = indexTagTokens(masked, key);
+        cache.set(key, index);
+    }
+    const partner = index.partners.get(tag.openStart);
+    const aligned = partner !== undefined && masked.indexOf(">", tag.openStart) + 1 === tag.openEnd;
+    if (!aligned)
+        return scanForBalancedClose(masked, key, tag.openEnd);
+    return partner ?? firstCloseFrom(index.closes, tag.openEnd);
+}
+/**
+ * Locate and parse every embed marker in `html`. Pure; allocates one result per marker.
+ *
+ * Scans {@link maskNonRenderableRegions}'s masked copy so a marker sitting inside an HTML comment or
+ * a `<script>`/`<style>` element's raw text is never matched, and so tags inside either are invisible
+ * to {@link findBalancedClose}'s depth count too. Every field on the resulting {@link EmbedMarker}
+ * (`whole`, `tag`, `attrs`, `inner`, the parsed config) is read back out of the ORIGINAL `html` at
+ * that match's own offsets — masking only decides which spans are eligible to be a marker or to count
+ * toward a close tag's depth, it must never change what an eligible marker's own fields report.
+ *
+ * Uses a private copy of {@link MARKER_PATTERN} (same source/flags) for the `exec` loop below, rather
+ * than mutating the shared module-level regex's `lastIndex` directly — this function is PURE, and a
+ * shared mutable scan cursor would corrupt a reentrant or recursive call.
+ */
+export function scanEmbedMarkers({ html }, _optional = {}) {
+    const markers = [];
+    const rejected = [];
+    let occurrence = 0;
+    const masked = maskNonRenderableRegions({ html: html });
+    const pattern = new RegExp(MARKER_PATTERN.source, MARKER_PATTERN.flags);
+    const closeIndexCache = new Map();
+    for (let match = pattern.exec(masked); match !== null; match = pattern.exec(masked)) {
+        occurrence += 1;
+        const indices = match.indices;
+        const [wholeStart, openTagEnd] = requireGroupRange(indices, 0);
+        const tagRange = requireGroupRange(indices, 1);
+        const attrsRange = requireGroupRange(indices, 2);
+        const doubleQuotedRange = indices[4];
+        const rawRange = doubleQuotedRange ?? requireGroupRange(indices, 3);
+        const tag = html.slice(...tagRange);
+        // No close at all (a void `<img>` marker): the open tag alone is the whole marker, inner empty.
+        const [closeStart, closeEnd] = findBalancedClose(masked, { name: tag, openStart: wholeStart, openEnd: openTagEnd }, closeIndexCache) ?? [
+            openTagEnd,
+            openTagEnd,
+        ];
+        pattern.lastIndex = closeEnd; // resume past this marker's whole balanced span, never re-entering it
+        const index = wholeStart;
+        const whole = html.slice(wholeStart, closeEnd);
+        const attrs = html.slice(...attrsRange);
+        const raw = doubleQuotedRange ? decodeAttributeValue(html.slice(...rawRange)) : html.slice(...rawRange);
+        const inner = html.slice(openTagEnd, closeStart);
+        const result = parseEmbedMarkerConfig({ raw: raw });
+        if ("problem" in result) {
+            rejected.push({ problem: result.problem, occurrence, index });
+            continue;
+        }
+        const { config } = result;
+        markers.push({
+            type: config.type,
+            id: typeof config.id === "string" ? config.id : undefined,
+            config,
+            tag,
+            attrs,
+            whole,
+            inner,
+            index,
+            occurrence,
+        });
+    }
+    return { markers, rejected };
+}
+/**
+ * The JSON text of a decoded marker value, made safe to sit verbatim inside a single-quoted
+ * attribute: `'` and `&` become the JSON escapes `\u0027`/`\u0026`. Both can only occur inside a
+ * JSON string, where those escapes mean the same character, so `JSON.parse` returns the identical
+ * value. Deliberately not `&#39;`/`&amp;`: {@link scanEmbedMarkers} reads a single-quoted value
+ * verbatim, so an entity there would reach `JSON.parse` undecoded and change the config.
+ *
+ * @complexity O(n) over `json`.
+ */
+function toSingleQuotedValue(json) {
+    return json.replace(/'/g, "\\u0027").replace(/&/g, "\\u0026");
+}
+/**
+ * Rewrites every double-quoted, entity-encoded `data-embed-config="{&quot;type&quot;:…}"` (what a DOM
+ * serializer such as the admin's Interactive canvas emits) back to the authored, readable form
+ * `data-embed-config='{"type":…}'`, so the stored page source reads the way it was written. Every
+ * Page HTML write runs through this (`PagesHtmlDocumentStore`), and the admin applies it to what the
+ * canvas hands back.
+ *
+ * Exact at the config level: a rewritten marker scans to the same `config`, `attrs` and `inner` it did
+ * before. Only the quoted value changes; the tag, other attributes and content are untouched. Left
+ * as written: a single-quoted marker, a marker inside a comment or `<script>`/`<style>` (never matched,
+ * see {@link maskNonRenderableRegions}), and a value that is not valid JSON, so a broken marker still
+ * reaches {@link scanEmbedMarkers}'s rejection with its original text.
+ *
+ * @complexity O(n) over `html`.
+ */
+export function normalizeEmbedMarkerQuoting({ html }, _optional = {}) {
+    const masked = maskNonRenderableRegions({ html: html });
+    const pattern = new RegExp(MARKER_PATTERN.source, MARKER_PATTERN.flags);
+    let result = "";
+    let copiedUpTo = 0;
+    for (let match = pattern.exec(masked); match !== null; match = pattern.exec(masked)) {
+        const valueRange = match.indices[4];
+        if (!valueRange)
+            continue;
+        const decoded = decodeAttributeValue(html.slice(...valueRange));
+        try {
+            JSON.parse(decoded);
+        }
+        catch {
+            continue;
+        }
+        // The value's quotes sit one character outside its capture group on either side.
+        result += `${html.slice(copiedUpTo, valueRange[0] - 1)}'${toSingleQuotedValue(decoded)}'`;
+        copiedUpTo = valueRange[1] + 1;
+    }
+    return result + html.slice(copiedUpTo);
+}
+/** Every marker of one type, in document order. The common "which menus does this theme name" shape. */
+export function markersOfType({ html, type }, _optional = {}) {
+    return scanEmbedMarkers({ html: html }).markers.filter((m) => m.type === type);
+}
+/**
+ * The two marker `type` values this module parses but never itself resolves — owned end-to-end by
+ * `features/theme/static-render.ts` (`injectMenuEmbeds` for {@link MENU_MARKER_TYPE}, `resolveSlots`
+ * for {@link PARTIAL_MARKER_TYPE}), which run AFTER `widgets/resolver-service.ts`'s page-embed stage
+ * on every static-tier page render.
+ *
+ * Hoisted here (2026-08-17) so that ownership fact has exactly one spelling. Before this, it existed
+ * only as inline string-literal comparisons inside `static-render.ts` itself
+ * (`marker.type !== "menu"`, `marker.type !== "partial"`), and `resolver-service.ts`'s own
+ * `THEME_OWNED_MARKER_TYPES` carried an independently-typed copy of the same two strings as a
+ * documented stopgap pending this exact hoist (`ADS-memory/reports/
+ * 2026-08-16-embed-placeholder-gap.md`, "Reuse decision") — two spellings of one fact is exactly the
+ * kind of drift risk this module exists to close for marker *parsing*; these two constants close it
+ * for marker *type naming* the same way.
+ *
+ * Deliberately NOT the same question `isPageEmbedType()` (`widgets/resolver-service.ts`) answers —
+ * that function tests registry membership (`Object.hasOwn(HTML_EMBED_RESOLVERS, type)`), which is
+ * `false` for these two AND for a genuine unregistered/typo type alike. These constants name the
+ * "owned elsewhere, not a typo" fact directly instead; see `isPageEmbedType`'s own doc for why
+ * conflating the two would silence a real warning.
+ */
+export const MENU_MARKER_TYPE = "menu";
+export const PARTIAL_MARKER_TYPE = "partial";
+/**
+ * The post-previews marker type (2026-09-03) — a theme/page marker that renders a bounded list of
+ * published post previews wherever it appears, so any authored Page (a static theme's own marketing
+ * page, or a Page/Post rendered through a `theme.json` `templates` entry) can host a post listing.
+ * Same ownership shape as {@link MENU_MARKER_TYPE}/{@link PARTIAL_MARKER_TYPE} immediately above:
+ * resolved end-to-end by `features/theme/static-render.ts`'s `injectPostPreviewsEmbeds`, never
+ * registered in `widgets/resolver-service.ts`'s `HTML_EMBED_RESOLVERS` — see that file's
+ * `THEME_OWNED_MARKER_TYPES`, which carries a matching literal entry for the same disclosed
+ * "duplicated, not imported" reason its own doc already states for the other two.
+ *
+ * Config shape: `{"type":"post-previews","limit":6}` — `limit` is optional (defaults to
+ * `DEFAULT_POST_PREVIEWS_LIMIT`, `static-render.ts`) and is always clamped, never trusted verbatim,
+ * before it reaches the bounded repo query that feeds this marker.
+ */
+export const POST_PREVIEWS_MARKER_TYPE = "post-previews";
+/**
+ * The collection marker type (2026-09-23) — a theme/page marker that renders a bounded, filtered,
+ * sorted list of entries from one custom content type wherever it appears. Same ownership shape as
+ * {@link MENU_MARKER_TYPE}/{@link PARTIAL_MARKER_TYPE}/{@link POST_PREVIEWS_MARKER_TYPE} immediately
+ * above: resolved end-to-end by `features/theme/static-render.ts`'s `injectCollectionEmbeds`, never
+ * registered in `widgets/resolver-service.ts`'s `HTML_EMBED_RESOLVERS` — see that file's
+ * `THEME_OWNED_MARKER_TYPES`, which carries a matching literal entry for the same disclosed
+ * "duplicated, not imported" reason its own doc gives for the other three.
+ *
+ * Config shape (validated by `entries/public-list.ts`'s `parseCollectionListConfig`, not by this
+ * module): `{"type":"collection","id":"recipe", …}` (`typeKey` accepted as an alias) plus the type's own filter/sort/limit/layout
+ * keys. Unlike {@link POST_PREVIEWS_MARKER_TYPE}, a resolved collection marker's wrapper is stripped
+ * of `data-embed-config` on the hit path ({@link withInnerContentFinal}) so a later re-scan of the
+ * same output can never rediscover and re-resolve it.
+ */
+export const COLLECTION_MARKER_TYPE = "collection";
+/**
+ * The featured-image marker type (2026-10-05) — a post template's slot for the current post's
+ * featured image. A static-tier theme has no template language to read `post.featuredImage` (the
+ * Liquid/Handlebars tiers' theme data), so before this marker a static theme had no way to show the
+ * image at all; it only reached `og:image`. Same ownership shape as {@link COLLECTION_MARKER_TYPE}:
+ * resolved end-to-end by `features/theme/static-render.ts`'s `injectFeaturedImage`, listed in
+ * `resolver-service.ts`'s `THEME_OWNED_MARKER_TYPES`.
+ *
+ * Config shape: `{"type":"featured-image"}` — no target; it always means the entity the route is
+ * rendering. A hit fills the authored element with the `<img>` and strips the marker; no image
+ * (none set, unresolved, or a page with no current post) removes the WHOLE element, wrapper and
+ * fallback included, so a theme's figure/margins never render around nothing.
+ */
+export const FEATURED_IMAGE_MARKER_TYPE = "featured-image";
+/**
+ * The longest a target value ({@link embedMarkerTarget}'s `id`/`slug`/`typeKey`) may be before it
+ * counts as absent rather than a usable reference. A duplicate of `features/widgets/html-embeds.ts`'s
+ * `MAX_EMBED_ID_LENGTH` (same number, same reasoning: a bound too long to safely carry into a
+ * `Map`/`Set` key or an `entry_refs` row), not a shared import — this module stays import-free (see
+ * {@link maskNonRenderableRegions}'s file header and `marker-target-parity.unit.test.ts`'s boundary
+ * assertion), the same "duplicated fact, disclosed in a comment" tradeoff {@link MENU_MARKER_TYPE} and
+ * {@link PARTIAL_MARKER_TYPE} already document for their own server-side twins.
+ */
+const MAX_TARGET_VALUE_LENGTH = 200;
+/**
+ * One marker type's `data-embed-config` keys that can name its resolution target, in the exact
+ * precedence order the server tries them — `embedMarkerTarget` returns the FIRST of these keys with a
+ * usable string value. This is the "which key names the target" rule Bug A's root cause report
+ * (2026-09-23 interactive-bugs plan) found spread across four different resolvers
+ * (`resolver-service.ts`'s widget/media/post/content, `routes/site/pages.ts`'s collection) with no
+ * single shared table — the admin's own placeholder describer could not reuse it because nothing here
+ * exposed it. Hoisting it as data (not a switch statement) is what lets both the server's own resolvers
+ * and the admin's UI-only describer read the identical precedence rule.
+ *
+ * - `widget`/`form`/`media`/`post`/`content`: `id`, else `slug` — registered in
+ *   `resolver-service.ts`'s `HTML_EMBED_RESOLVERS` (see `marker-target-parity.unit.test.ts`, which fails
+ *   loudly if a registered type has no matching row or incorrect target keys).
+ * - `taxonomy`: `id` only — `widgets/resolvers/taxonomy.ts` reads `ref.id`, accepting a real
+ *   taxonomy id or a unique taxonomy name through that key; `ref.slug` is never consulted.
+ * - {@link COLLECTION_MARKER_TYPE}: `id`, else `typeKey` — `routes/site/pages.ts`'s
+ *   `collectionMarkerTypeKey` tries `config.id` before `config.typeKey`; `slug` is never consulted.
+ * - {@link MENU_MARKER_TYPE} / {@link PARTIAL_MARKER_TYPE}: `id` only — `static-render.ts` resolves
+ *   both purely by id.
+ * - {@link POST_PREVIEWS_MARKER_TYPE}: no keys at all — this marker needs no target
+ *   ({@link embedMarkerTarget} reports that as its own `{ key: "none" }` outcome, never `undefined`,
+ *   which means "target missing").
+ *
+ * An unknown type (not a key of this table) is not a member of it; {@link embedMarkerTarget} falls back
+ * to `["id"]` for that case rather than treating an unrecognized type as needing no target — a display
+ * must never claim a slug resolves for a type no resolver has ever registered.
+ */
+export const EMBED_MARKER_TARGET_KEYS = {
+    widget: ["id", "slug"],
+    form: ["id", "slug"],
+    taxonomy: ["id"],
+    media: ["id", "slug"],
+    post: ["id", "slug"],
+    content: ["id", "slug"],
+    [COLLECTION_MARKER_TYPE]: ["id", "typeKey"],
+    [MENU_MARKER_TYPE]: ["id"],
+    [PARTIAL_MARKER_TYPE]: ["id"],
+    [POST_PREVIEWS_MARKER_TYPE]: [],
+};
+/** The default target-key list for a type absent from {@link EMBED_MARKER_TARGET_KEYS} — an unknown
+ *  or typo'd type is assumed id-addressable like most types, but never slug-addressable: no resolver
+ *  has ever registered to read a slug for it, so claiming one would resolve is always wrong. */
+const UNKNOWN_TYPE_TARGET_KEYS = ["id"];
+/** `value` trimmed down to a usable target string, or `undefined` if it is not a non-empty string of
+ *  at most {@link MAX_TARGET_VALUE_LENGTH} characters — the same absent-vs-present rule
+ *  `features/widgets/html-embeds.ts`'s `normalizeEmbedId`/`normalizeEmbedSlug` apply to their own
+ *  config reads. @complexity O(1). */
+function normalizeTargetValue(value) {
+    if (typeof value !== "string" || value.length === 0 || value.length > MAX_TARGET_VALUE_LENGTH)
+        return undefined;
+    return value;
+}
+/**
+ * Which `data-embed-config` key names `type`'s resolution target, and that key's value, per
+ * {@link EMBED_MARKER_TARGET_KEYS}'s precedence order. Three distinct outcomes:
+ *
+ * - `{ key, value }` — the first target key present with a usable string value.
+ * - `{ key: "none" }` — this type ({@link POST_PREVIEWS_MARKER_TYPE}) needs no target; a caller must
+ *   never read this as "target missing".
+ * - `undefined` — the type has one or more target keys, but the marker's config supplies none of them
+ *   (or only with an unusable value per {@link normalizeTargetValue}) — "target missing".
+ *
+ * `type` is compared lower-cased, matching `toEmbedRef`'s own `marker.type.toLowerCase()` in
+ * `resolver-service.ts` — a hand-authored `"Widget"` must resolve exactly as `"widget"` does.
+ *
+ * @complexity O(k) in the number of candidate keys for `type` (at most 2 today) — independent of the
+ * surrounding document's size.
+ */
+export function embedMarkerTarget({ type, config }, _optional = {}) {
+    const keys = EMBED_MARKER_TARGET_KEYS[type.toLowerCase()] ?? UNKNOWN_TYPE_TARGET_KEYS;
+    if (keys.length === 0)
+        return { key: "none" };
+    for (const key of keys) {
+        const value = normalizeTargetValue(config[key]);
+        if (value !== undefined)
+            return { key, value };
+    }
+    return undefined;
+}
+/**
+ * The "Copy embed code" snippet for any marker type: the same `<div data-embed-config='{"type":…,
+ * "<key>":"<value>"}'></div>` wrapper {@link scanEmbedMarkers}/{@link parseEmbedMarkerConfig} already
+ * parse back — generalised (readable-slugs S5b) out of `apps/admin/src/features/collections/rules.ts`'s
+ * `collectionEmbedSnippet`, the one admin-generated marker that existed before this: that function is
+ * now a one-line call onto this (`embedMarkerSnippet("collection", "id", key)`), byte-for-byte
+ * unchanged. `features/media`'s own "Copy embed code" button (S5a) is the second caller, with
+ * `("media", "slug", asset.slug)` — see {@link EMBED_MARKER_TARGET_KEYS}'s `media` row for why `slug`
+ * is a real, resolvable target key for this type.
+ *
+ * No escaping beyond the plain template: `key`/`value` are expected to already be validated,
+ * quote-free identifiers (a slug, a type key, an id) by the time a caller reaches for a copy button —
+ * the same assumption `collectionEmbedSnippet` always made.
+ *
+ * @complexity O(1) — one template string.
+ */
+export function embedMarkerSnippet({ type, key, value }, _optional = {}) {
+    return `<div data-embed-config='{"type":"${type}","${key}":"${value}"}'></div>`;
+}
+/**
+ * Rebuild a marker's element around new inner content, keeping its own tag and every authored
+ * attribute (`class`, `aria-label`, …). The counterpart to a wholesale replace: a menu marker keeps
+ * its `<nav class="docs-nav">` wrapper and only swaps what's inside, whereas a partial slot marker
+ * disappears entirely and is replaced by the partial.
+ *
+ * Synthesizing the open tag from `config` alone instead of using this is a silent bug — it drops the
+ * theme's styling hooks and accessible names with nothing failing.
+ */
+export function withInnerContent({ marker, inner }, _optional = {}) {
+    return `<${marker.tag}${marker.attrs}>${inner}</${marker.tag}>`;
+}
+/**
+ * Rebuild a marker's element with `id` added to its `data-embed-config`, keeping every other
+ * authored config key, the marker's own tag, its other attributes, and its authored inner content
+ * ({@link EmbedMarker.inner}) completely unchanged. The counterpart to {@link withInnerContent} for a
+ * caller that needs to SUPPLY a marker's target rather than its content — e.g. the unified `content`
+ * marker (2026-08-11): a theme-authored `{"type":"content"}` with no `id` means "the entity the route
+ * already resolved", and the route fills that id in here before the marker is ever resolved, so every
+ * downstream consumer (the scanner, the resolver, a later re-scan) sees an ordinary id-carrying marker
+ * rather than needing its own "no id means current entity" special case.
+ *
+ * Callers must check `marker.id === undefined` themselves before calling this — an author's own
+ * explicit id always wins over a caller-supplied default, and this function does not re-check that
+ * (it would silently overwrite an explicit reference otherwise, which is the one thing the unified
+ * marker's id-vs-no-id contract must never do).
+ *
+ * @complexity O(n) over the marker's own attrs length (one JSON stringify of a small object, one
+ * regex replace over `attrs`) — independent of the surrounding document's size.
+ */
+export function withAddedId({ marker, id }, _optional = {}) {
+    const config = JSON.stringify({ ...marker.config, id });
+    // Keeps the marker's own quoting: a double-quoted value is re-encoded, a single-quoted one is not.
+    const attrsWithId = marker.attrs.replace(MARKER_ATTRIBUTE_PATTERN, (_whole, doubleQuoted) => doubleQuoted === undefined ? ` data-embed-config='${config}'` : ` data-embed-config="${encodeDoubleQuotedValue(config)}"`);
+    return `<${marker.tag}${attrsWithId}>${marker.inner}</${marker.tag}>`;
+}
+/**
+ * Rebuild a marker's element around new inner content, like {@link withInnerContent}, but ALSO strips
+ * the marker's own `data-embed-config` attribute from the rebuilt tag — every other authored attribute
+ * (`class`, `aria-label`, …) survives, only the marker-ness is removed.
+ *
+ * For a caller that performs its OWN final resolution outside the shared registry/render pipeline
+ * (`pages.ts`'s recursive `content`-marker pre-splice, which fetches and inlines an `"html"`-format
+ * entity's own body ahead of the normal `resolveHtmlPageEmbeds`/`renderHtmlPageBody` pass) and must
+ * guarantee the result is not rediscovered and re-resolved by a LATER marker scan over the same output
+ * — an ordinary {@link withInnerContent} splice would leave `data-embed-config` intact, and the next
+ * scan would see what looks like a fresh, unresolved marker and try to resolve it again.
+ *
+ * @complexity O(n) over the marker's own attrs length (one regex replace) — independent of the
+ * surrounding document's size.
+ */
+export function withInnerContentFinal({ marker, inner }, _optional = {}) {
+    const attrsWithoutMarker = marker.attrs.replace(MARKER_ATTRIBUTE_PATTERN, "");
+    return `<${marker.tag}${attrsWithoutMarker}>${inner}</${marker.tag}>`;
+}
+/**
+ * Replace each marker's WHOLE element with what `resolve` returns for it.
+ *
+ * Whole-element rather than inner-content because the two consumers genuinely differ: a partial slot
+ * marker is scaffolding that vanishes once the partial is spliced in, while a menu marker is real
+ * theme markup that must survive with only its contents swapped. Callers in the second camp wrap
+ * their output in {@link withInnerContent}; making that explicit at the call site is better than a
+ * mode flag, because the choice is a real per-type decision and not a preference.
+ *
+ * `resolve` returning `undefined` means "leave this marker exactly as authored", and that default is
+ * load-bearing: a menu that does not exist, a target that was deleted, or a render that produces
+ * nothing must fall back to the theme's authored content rather than blanking a nav. Every consumer
+ * wants that rule, so it lives here rather than being re-implemented per call site.
+ *
+ * Applies right-to-left so each splice leaves the earlier markers' offsets valid.
+ */
+export function substituteMarkers({ html, resolve }, _optional = {}) {
+    const { markers } = scanEmbedMarkers({ html: html });
+    let out = html;
+    for (let i = markers.length - 1; i >= 0; i -= 1) {
+        // The reverse loop keeps i within the scanner result.
+        const m = markers[i];
+        const replacement = resolve(m);
+        if (replacement === undefined)
+            continue;
+        out = out.slice(0, m.index) + replacement + out.slice(m.index + m.whole.length);
+    }
+    return out;
+}
+/**
+ * The shape every attribute name {@link parseMarkerAttributes} accepts must have — deliberately
+ * narrower than the full HTML spec's own name production, sized for real attribute names (`class`,
+ * `data-*`, `aria-*`, `autoplay`, …), the same "narrower on purpose" discipline
+ * `@jini-ai/cms/media`'s `html-attributes.ts` applies to its own allowlisted names. A token whose name
+ * fails this is skipped, not thrown — this parser runs over already-authored page HTML (page trust,
+ * no allowlist; see the 2026-09-16 embed-attributes plan §0), so "not a well-formed attribute name" is
+ * a shape problem this parser degrades past, not a security boundary.
+ */
+const MARKER_ATTRIBUTE_NAME_PATTERN = /^[a-z_:][-a-z0-9_:.]*$/;
+/**
+ * Quote-aware attribute-token scanner: matches one `name`, or one `name=unquoted`, `name="value"`, or
+ * `name='value'` pair. The identical loose shape `@jini-ai/cms/media`'s `html-attributes.ts` uses for
+ * its own `HTML_ATTRIBUTE_TOKEN` — quote-aware so a `data-embed-config='{"type":"media"}'` token
+ * (double quotes living inside a single-quoted value) is consumed as ONE token rather than corrupting
+ * the scan at its first interior `"`.
+ */
+const MARKER_ATTRIBUTE_TOKEN = /([^\s="']+)(?:=(?:"([^"]*)"|'([^']*)'|([^\s"']+)))?/g;
+/**
+ * Parses a marker's verbatim {@link EmbedMarker.attrs} text into structured attributes, in authored
+ * order, with `data-embed-config` itself removed (it is the marker, not an attribute a caller would
+ * ever want to forward) — the read half of the 2026-09-16 embed-attributes contract:
+ * {@link formatMarkerAttributes} is its inverse.
+ *
+ * - Names are lowercased (HTML attribute names are case-insensitive).
+ * - A token whose (lowercased) name is not a valid attribute name per
+ *   {@link MARKER_ATTRIBUTE_NAME_PATTERN} is skipped, never thrown.
+ * - A duplicate name keeps its FIRST occurrence — the same rule a real HTML parser applies, so this
+ *   parser's result matches what a browser would actually expose on the element.
+ * - Every value is kept as SOURCE TEXT: never entity-decoded. Re-decoding `&amp;` here and
+ *   re-escaping on output (see {@link formatMarkerAttributes}) would double-escape a value like
+ *   `title="a &amp; b"` into `&amp;amp;` — this module never round-trips through a decoded form.
+ *
+ * @complexity O(n) in the marker's own attrs length (one regex pass) — independent of the surrounding
+ * document's size.
+ */
+export function parseMarkerAttributes({ marker }, _optional = {}) {
+    const seenNames = new Set();
+    const attributes = [];
+    MARKER_ATTRIBUTE_TOKEN.lastIndex = 0;
+    let match;
+    while ((match = MARKER_ATTRIBUTE_TOKEN.exec(marker.attrs)) !== null) {
+        const name = match[1].toLowerCase();
+        if (name === "data-embed-config")
+            continue;
+        if (!MARKER_ATTRIBUTE_NAME_PATTERN.test(name))
+            continue;
+        if (seenNames.has(name))
+            continue;
+        seenNames.add(name);
+        const value = match[2] ?? match[3] ?? match[4] ?? null;
+        attributes.push({ name, value });
+    }
+    return attributes;
+}
+/**
+ * Formats parsed attributes back into attribute-string text, the inverse of
+ * {@link parseMarkerAttributes}: a valueless attribute is emitted as a bare ` name`; a valued one as
+ * ` name="value"`. The ONLY transform applied to a value is `"` -> `&quot;` (a value may itself
+ * contain a literal `"`, e.g. carried over from a single-quoted authored attribute) — `&` is never
+ * touched, so source text that already reads `&amp;` is emitted unchanged rather than
+ * double-escaping into `&amp;amp;`.
+ *
+ * @complexity O(n) in the total length of every attribute's name and value.
+ */
+export function formatMarkerAttributes({ attrs }, _optional = {}) {
+    return attrs.map((attr) => (attr.value === null ? ` ${attr.name}` : ` ${attr.name}="${attr.value.replace(/"/g, "&quot;")}"`)).join("");
+}
+/** Whether a marker carries any authored attribute besides `data-embed-config` itself — the "bare
+ *  marker" test {@link withElementKeptIfAttributed} uses to decide whether an element survives a
+ *  whole-marker substitution.
+ *
+ * @complexity O(n) in the marker's own attrs length (delegates to {@link parseMarkerAttributes}).
+ */
+export function hasAuthoredAttributes({ marker }, _optional = {}) {
+    return parseMarkerAttributes({ marker: marker }).length > 0;
+}
+/**
+ * Rebuild a marker's element around new inner content using an EXPLICIT attribute list (rather than
+ * the marker's own verbatim {@link EmbedMarker.attrs}) — for a caller that has partitioned a marker's
+ * authored attributes and needs only a subset on the rebuilt wrapper (e.g. {@link substituteHtmlEmbeds}
+ * moving media-element attribute names off a non-media wrapper onto the resolved `<video>`/`<img>`,
+ * keeping the rest on the wrapper). The counterpart to {@link withInnerContent}, which always keeps
+ * every authored attribute verbatim.
+ */
+export function withInnerContentAndAttributes({ marker, attrs, inner }, _optional = {}) {
+    return `<${marker.tag}${formatMarkerAttributes({ attrs: attrs })}>${inner}</${marker.tag}>`;
+}
+/**
+ * Whole-element replace UNLESS the marker carries authored attributes, in which case its element
+ * survives (minus `data-embed-config`) around the resolved content — the 2026-09-16 rule for marker
+ * types (`widget`, `partial`) that default to disappearing entirely: a bare marker keeps disappearing,
+ * byte-identical to before, while an attributed one keeps whatever styling/accessibility hook the
+ * author put on it. Delegates to {@link withInnerContentFinal} for the attributed case so the kept
+ * wrapper's attribute text is the marker's own verbatim {@link EmbedMarker.attrs} (not a
+ * parse-then-reformat round trip) — the same byte-identical-when-nothing-moves discipline
+ * {@link withInnerContentFinal} itself already provides.
+ */
+export function withElementKeptIfAttributed({ marker, inner }, _optional = {}) {
+    return hasAuthoredAttributes({ marker: marker }) ? withInnerContentFinal({ marker: marker, inner: inner }) : inner;
+}
+/**
+ * Human-readable one-liner for a rejection, for a warning line or a write-time validation error.
+ * Deliberately quotes the offending JSON: the author needs to see what they typed, and the raw text
+ * is theme- or admin-authored, never end-user input.
+ */
+export function describeRejection({ rejection }, _optional = {}) {
+    const { problem, occurrence } = rejection;
+    const where = `embed marker #${occurrence}`;
+    if (problem.kind === "invalid-json")
+        return `${where}: data-embed-config is not valid JSON (${problem.message}) — ${problem.raw}`;
+    if (problem.kind === "not-an-object")
+        return `${where}: data-embed-config must be a JSON object — ${problem.raw}`;
+    return `${where}: data-embed-config is missing a "type" — ${problem.raw}`;
+}
+//# sourceMappingURL=marker.js.map