@uniweb/projections 0.4.0 → 0.5.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uniweb/projections",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "Projections of a Uniweb site's content — agent index, per-page markdown, search index. Pure JS, runs anywhere.",
   "type": "module",
   "exports": {
@@ -31,8 +31,8 @@
     "node": ">=20.19"
   },
   "dependencies": {
-    "@uniweb/content-writer": "^0.3.4",
-    "@uniweb/core": "^0.13.0"
+    "@uniweb/core": "^0.13.0",
+    "@uniweb/content-writer": "^0.3.4"
   },
   "devDependencies": {
     "vitest": "^4.1.7",

package/src/markdown.js

@@ -100,6 +100,71 @@ function collectSection(section, blocks, includeChildren, ancestorAnchor) {
   }
 }
 
+/**
+ * Replace `inset_placeholder` nodes with the author's own words.
+ *
+ * ## Why this is not "restore the inset"
+ *
+ * An inset is `![Platform overview](@Diagram)` — an author's caption plus a
+ * FOUNDATION COMPONENT to render it with. The build splits them: the caption and
+ * params go to the section's `insets[]`, and the body keeps an
+ * `inset_placeholder` carrying only `{ refId, embedKind }`.
+ *
+ * ⛔ **`@Diagram` must never reach this output.** A component name is a rendering
+ * assignment, and this package's whole property is that a projection is of the
+ * SITE — identical under a swapped foundation. Emitting it here would break the
+ * same rule that keeps `type:` and params out (see the package's README/notes on
+ * why the exclusions are load-bearing rather than tidy-up).
+ *
+ * ⭐ **But the caption IS site content** — the author wrote it, and an agent
+ * retrieving this page should read it. So the placeholder becomes its title, as
+ * plain text, and nothing else.
+ *
+ * ⚠️ Before this, `proseMirrorToMarkdown` had no serializer for the node and
+ * dropped it with a warning per build — *"this is a tracked capability gap"*. It
+ * was: every inset caption was missing from every agent-facing page.
+ *
+ * @param {Object} content - the section's ProseMirror document
+ * @param {Array} insets - the section's `insets[]` (`{ refId, title }`)
+ * @returns {Object} content with placeholders resolved to text
+ */
+function resolveInsetCaptions(content, insets) {
+  if (!content?.content?.length) return content
+  // No insets array → nothing to resolve against. Dropping is then still the only
+  // option, but it is silent: the caption is genuinely not reachable from here.
+  const titleByRef = new Map(
+    (Array.isArray(insets) ? insets : [])
+      .filter((i) => i && typeof i.refId === 'string' && i.title)
+      .map((i) => [i.refId, String(i.title)])
+  )
+
+  // ⛔ THE REPLACEMENT'S SHAPE DEPENDS ON WHERE IT SITS, and getting this wrong
+  // fails SILENTLY IN THE WORSE DIRECTION: a bare text node at block level is not
+  // serializable, so the caption vanishes exactly as before — but the warning that
+  // used to announce it is gone. Measured while writing this: the first version
+  // emitted text unconditionally, removed the warning, and restored nothing.
+  const TEXTBLOCKS = new Set(['paragraph', 'heading'])
+
+  const visit = (nodes, inline) =>
+    nodes.flatMap((node) => {
+      if (!node) return []
+      if (node.type === 'inset_placeholder') {
+        const title = titleByRef.get(node.attrs?.refId)
+        // An inset with no caption contributes no author text — drop it, and do
+        // so quietly: there is nothing a reader is missing.
+        if (!title) return []
+        const text = { type: 'text', text: title }
+        return inline ? [text] : [{ type: 'paragraph', content: [text] }]
+      }
+      if (Array.isArray(node.content)) {
+        return [{ ...node, content: visit(node.content, TEXTBLOCKS.has(node.type)) }]
+      }
+      return [node]
+    })
+
+  return { ...content, content: visit(content.content, false) }
+}
+
 /**
  * @param {Object} section
  * @returns {string}
@@ -107,5 +172,5 @@ function collectSection(section, blocks, includeChildren, ancestorAnchor) {
 function serializeSectionContent(section) {
   const content = section?.content
   if (!content?.content?.length) return ''
-  return proseMirrorToMarkdown(content).trim()
+  return proseMirrorToMarkdown(resolveInsetCaptions(content, section?.insets)).trim()
 }

package/src/search/extract.js

@@ -235,10 +235,32 @@ function extractFromProseMirrorDoc(doc, options) {
     return { title, textParts }
   }
 
-  for (const node of doc.content) {
+  // ⛔ CONTAINERS WHOSE CHILDREN ARE ORDINARY AUTHOR PROSE. Until 2026-08-27 this
+  // walk was FLAT over `doc.content`, so anything nested one level down was
+  // invisible to search — measured: a blockquote's prose and every table cell were
+  // lost outright, on a site that looked perfectly indexed.
+  //
+  // ⭐ An ALLOWLIST rather than a blind recursion, deliberately: descending into
+  // everything would pull in `codeBlock` and `math_display`, which are not prose —
+  // they inflate the index and match on tokens nobody searches for. Adding a
+  // container here is a decision, not a default.
+  const DESCEND = new Set(['blockquote', 'table', 'tableRow', 'tableCell'])
+
+  const walk = (nodes) => {
+    for (const node of nodes || []) {
+      if (!node) continue
+      if (DESCEND.has(node.type)) {
+        walk(node.content)
+        continue
+      }
+      visit(node)
+    }
+  }
+
+  const visit = (node) => {
     if (node.type === 'heading') {
       const text = extractTextFromNode(node)
-      if (!text) continue
+      if (!text) return
 
       // First H1 becomes the title
       if (!foundFirstHeading && node.attrs?.level === 1) {
@@ -269,9 +291,16 @@ function extractFromProseMirrorDoc(doc, options) {
     } else if ((node.type === 'bulletList' || node.type === 'orderedList') && includeLists) {
       const listTexts = extractFromList(node)
       textParts.push(...listTexts)
+    } else if (node.type === 'image') {
+      // ⭐ `alt` is the author describing their own image — the only words an
+      // image contributes, and what a reader searching for it would type.
+      const alt = typeof node.attrs?.alt === 'string' ? node.attrs.alt.trim() : ''
+      if (alt) textParts.push(alt)
     }
   }
 
+  walk(doc.content)
+
   return { title, textParts }
 }
 

package/src/search/records.js

@@ -79,11 +79,12 @@ const DISPLAY_VALUE_MAX = 200
  *    as an index OF collections or a collection's own listing. It is neither: it
  *    is a search index derived FROM records.
  *
- * ⚠️ The RESULT still carries the old vocabulary — `type: 'collection'`,
- * `collection: name`, and `id: "collection:<name>:<slug>"`. That is a data shape
- * with live consumers (`kit`'s endpoint search provider and hosting's search both
- * read `entry.collection`), so it is a separate, larger decision than this rename
- * and is deliberately NOT bundled into it.
+ * ✅ The RESULT was renamed to match, 2026-08-27 — `type: 'record'`, `group: name`,
+ * `id: "record:<group>:<slug>"`. That was a real data break against two live
+ * consumers (kit's endpoint search provider and hosting's search), taken
+ * deliberately rather than left to rot: doing it while the shape was already
+ * being discussed cost one coordinated change; leaving it would have made the
+ * entry the last place `collection` survived as a lane-crossing word.
  */
 export function generateRecordSearchIndex(name, config, collectionData, locale) {
   // ⛔ NO DEFAULT FIELD LIST. This was `|| ['title']` — a claim about someone
@@ -113,9 +114,18 @@ export function generateRecordSearchIndex(name, config, collectionData, locale)
     // correctly, looks plausible, and 404s on click.
     const route = item.route || composeRoute(config.route, slug)
     return {
-      id: `collection:${name}:${slug}`,
-      type: 'collection',
-      collection: name,
+      // ⛔ RENAMED 2026-08-27 — `collection` is FRAMEWORK'S build concept (a named
+      // set our build compiles to one file) and the live lane has no such thing:
+      // a host calls this with records fetched from a folder. Same category error
+      // the function name carried until it became `generateRecordSearchIndex`.
+      //
+      // ⭐ `record` is symmetric with the page entry's `type: 'page'` / `id:
+      // "page:<route>"`, and true on both lanes. `group` names what a result UI
+      // actually does with it — label or group results by the set they came from —
+      // without borrowing either lane's word for that set.
+      id: `record:${name}:${slug}`,
+      type: 'record',
+      group: name,
       ...(route ? { route } : {}),
       title: item.title || item.name || slug,
       content,
@@ -130,8 +140,8 @@ export function generateRecordSearchIndex(name, config, collectionData, locale)
   // No `generated` timestamp — see the note in `generate.js`. A clock defeats
   // content-addressing and byte-parity between publishers.
   return {
-    type: 'collection',
-    collection: name,
+    type: 'record',
+    group: name,
     locale,
     entries,
   }