dsh-book-reader 0.3.2 → 0.3.3

package/CHANGELOG.md

@@ -37,6 +37,24 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   rejects at registration time. The exception only reached the Host log, so the
   plugin looked enabled with two tools instead of three.
 
+## [0.3.3] — 2026-10-29
+
+### Fixed
+
+- Page geometry is no longer measured while the PDF preview is scrolling. The
+  O(410) rect pass used to run on every lazy page render (ResizeObserver),
+  forcing a full preview layout against PDF.js and leaving a grey placeholder
+  for seconds at high zoom on Retina displays. Measurements now run ~260 ms
+  after scrolling stops, are skipped entirely while the saved page/zoom restore
+  is in flight, and page wrappers are cached between passes.
+- PDF text, search snippets, outline titles and document metadata are now
+  returned as well-formed Unicode. Some pages' fonts make PDF.js emit an
+  unpaired UTF-16 surrogate (for example `U+D835`); that string is valid JSON
+  but invalid for the model API, and a single such tool result in a Session log
+  made every later DeepSeek Messages request fail with `400 INVALID_REQUEST`.
+  Unpaired surrogates are replaced with `U+FFFD` (length-preserving, so search
+  offsets stay aligned).
+
 ## [0.3.2] — 2026-10-23
 
 ### Fixed
@@ -123,6 +141,7 @@ First public release.
 - `pdfjs-dist` is pinned to `6.3.289` to match the host process.
 - MIT licensed.
 
+[0.3.3]: https://github.com/Strepsiades2/dsh-book-reader/releases/tag/v0.3.3
 [0.3.2]: https://github.com/Strepsiades2/dsh-book-reader/releases/tag/v0.3.2
 [0.3.1]: https://github.com/Strepsiades2/dsh-book-reader/releases/tag/v0.3.1
 [0.3.0]: https://github.com/Strepsiades2/dsh-book-reader/releases/tag/v0.3.0

package/lib/client.js

@@ -1250,6 +1250,13 @@ window.__ModuleLoader__.load({
       }
       /** Latest sampled page, readable from the debounced view save. */
       const pageRef = useRef(1)
+      /**
+       * True while the saved page/zoom restore is in flight. Geometry
+       * measurements must stay out of that window: measuring 410 wrappers
+       * forces a full preview layout and would land exactly when PDF.js is
+       * painting the first restored page.
+       */
+      const restoringViewRef = useRef(false)
       /** Installed by the view effect; `null` when this preview keeps no view. */
       const requestViewSaveRef = useRef(null)
 
@@ -1521,6 +1528,8 @@ window.__ModuleLoader__.load({
         let retryTimer
         let retryCount = 0
         let geometry = []
+        let cachedPages = []
+        let scrollIdleTimer
         let lastPage
         const view = typeof window !== 'undefined' ? window : globalThis
 
@@ -1567,6 +1576,10 @@ window.__ModuleLoader__.load({
 
         const measure = () => {
           if (disposed) return
+          if (restoringViewRef.current) {
+            scheduleMeasureOnIdle()
+            return
+          }
           const currentPane = resolvePane(host)
           if (currentPane === undefined) {
             scheduleRetry()
@@ -1578,7 +1591,13 @@ window.__ModuleLoader__.load({
             bind(currentPane, currentScrollport, currentSection)
             return
           }
-          const pages = pageElements(currentPane)
+          let pages = cachedPages
+          if (pages.length === 0
+            || pages[0].element.isConnected === false
+            || pages[pages.length - 1].element.isConnected === false) {
+            pages = pageElements(currentPane)
+            cachedPages = pages
+          }
           const nextGeometry = []
           try {
             const portRect = currentScrollport === undefined ? { top: 0 } : currentScrollport.getBoundingClientRect()
@@ -1598,12 +1617,46 @@ window.__ModuleLoader__.load({
           sample()
         }
 
+        /**
+         * Measuring every page's rect forces a full layout of the preview's
+         * page/text-layer DOM. Doing that while pages are being rendered (the
+         * scroll-active window) serializes layout against PDF.js and can delay
+         * a page's canvas paint by seconds. Record geometry only once the
+         * scrollport has been quiet for a moment.
+         */
+        const clearScrollIdle = () => {
+          if (scrollIdleTimer !== undefined) {
+            globalThis.clearTimeout(scrollIdleTimer)
+            scrollIdleTimer = undefined
+          }
+        }
+
+        const scheduleMeasureOnIdle = () => {
+          clearScrollIdle()
+          scrollIdleTimer = globalThis.setTimeout(() => {
+            scrollIdleTimer = undefined
+            if (disposed) return
+            if (restoringViewRef.current) {
+              scheduleMeasureOnIdle()
+              return
+            }
+            measureDebouncer.schedule(() => {
+              measure()
+              sample()
+            })
+          }, 260)
+        }
+
         const onScroll = () => {
+          scheduleMeasureOnIdle()
           throttle.schedule(sample)
         }
 
         const onResize = () => {
           if (disposed) return
+          // Page renders change the section height while scrolling; defer that
+          // measurement to the scroll-idle pass instead of thrashing layout now.
+          if (scrollIdleTimer !== undefined || restoringViewRef.current) return
           measureDebouncer.schedule(() => {
             measure()
             sample()
@@ -1645,6 +1698,7 @@ window.__ModuleLoader__.load({
         }
 
         const detach = () => {
+          clearScrollIdle()
           scrollport?.removeEventListener('scroll', onScroll)
           pane?.removeEventListener('scroll', onScroll, true)
           view.removeEventListener('resize', onResize)
@@ -1661,6 +1715,7 @@ window.__ModuleLoader__.load({
           scrollport = nextScrollport
           section = nextSection
           geometry = []
+          cachedPages = nextPane === undefined ? [] : pageElements(nextPane)
           lastPage = undefined
           if (pane !== undefined) {
             pane.addEventListener('scroll', onScroll, true)
@@ -1675,6 +1730,13 @@ window.__ModuleLoader__.load({
           if (nextScrollport !== undefined) {
             nextScrollport.addEventListener('scroll', onScroll, { passive: true })
           }
+          // Publish the page total from the wrapper count, then do the initial
+          // geometry pass. This runs before the saved-view restore starts, on
+          // placeholder layout only; later page-render driven re-measurement is
+          // the expensive case and stays deferred.
+          if (cachedPages.length > 0) {
+            setDomTotal(cachedPages[cachedPages.length - 1].number)
+          }
           measure()
         }
 
@@ -1723,6 +1785,7 @@ window.__ModuleLoader__.load({
         const host = hostRef.current
         if (host === null) return undefined
         const pdfPath = absolutePath
+        restoringViewRef.current = true
 
         let disposed = false
         let finished = false
@@ -1808,6 +1871,7 @@ window.__ModuleLoader__.load({
         const finish = () => {
           if (disposed || finished) return
           finished = true
+          restoringViewRef.current = false
           requestSave()
         }
 
@@ -2046,6 +2110,7 @@ window.__ModuleLoader__.load({
 
         return () => {
           disposed = true
+          restoringViewRef.current = false
           restoreController.abort()
           saveController.abort()
           saveDebouncer.cancel()

package/lib/core/README.md

@@ -115,8 +115,14 @@ Hits rank **body pages first** (numeric `printed`) in page/offset order, then
 front matter (roman labels) in order, with at most 3 front-matter hits while the
 overall `limit` is applied. `pages` counts every matched page in the active pass
 before ranking/limit. `indexOnPage` is a UTF-16 offset into the original
-`getPageText()` text, and snippets are cut from that original text (surrogate
-pairs preserved).
+`getPageText()` text, and snippets are cut from that original text.
+
+All returned text and metadata is **well-formed Unicode**. PDF.js can emit an
+unpaired UTF-16 surrogate from a broken font `ToUnicode` map (for example
+`U+D835` in LADR's math fonts); such strings are valid JSON but invalid for a
+model API and one tool result containing one made every later DeepSeek Messages
+request fail with `400 INVALID_REQUEST`. Unpaired surrogates are replaced with
+`U+FFFD` (one code unit, so search offsets stay aligned).
 
 * `snippet` is at most about `2 * contextChars + query.length` characters, with
   leading/trailing `…` when the hit is mid-page.

package/lib/core/pdf.js

@@ -599,8 +599,9 @@ async function flattenOutline(doc, items, meta) {
     for (const node of nodes) {
       if (!node || typeof node !== 'object') continue
 
-      const title =
-        typeof node.title === 'string' ? node.title.replace(/\s+/g, ' ').trim() : ''
+      const title = wellFormedText(
+        typeof node.title === 'string' ? node.title.replace(/\s+/g, ' ').trim() : '',
+      )
       if (title !== '') {
         const destination = await resolveDestination(doc, node.dest, namedCache)
         const pdfPage = await resolvePdfPage(doc, destination, pageCache)
@@ -637,7 +638,9 @@ async function readDocumentMetadata(doc) {
   try {
     const result = await doc.getMetadata()
     const info = result?.info ?? {}
-    const str = (value) => (typeof value === 'string' && value.trim() !== '' ? value.trim() : null)
+    const str = (value) => (typeof value === 'string' && value.trim() !== ''
+      ? wellFormedText(value.trim())
+      : null)
     return {
       title: str(info.Title),
       author: str(info.Author),
@@ -795,6 +798,30 @@ function normalizePageText(raw) {
   return raw.replace(/\s+/g, ' ').trim()
 }
 
+/**
+ * Replace unpaired UTF-16 surrogates with U+FFFD.
+ *
+ * PDF.js text items can contain a lone high surrogate when a font's ToUnicode
+ * map is broken (typical for math fonts split across runs). Such strings are
+ * valid JSON (`"\ud835"`) but invalid Unicode for the model API; one such tool
+ * result in a Session log makes every later DeepSeek Messages request fail with
+ * 400 INVALID_REQUEST. Each plugin string that reaches a Tool result or route
+ * response must be well formed.
+ *
+ * `toWellFormed()` preserves the string length (one code unit -> one), which
+ * keeps search offset maps aligned.
+ * @param {unknown} value
+ * @returns {string}
+ */
+function wellFormedText(value) {
+  if (typeof value !== 'string') return ''
+  if (typeof value.toWellFormed === 'function') return value.toWellFormed()
+  return value.replace(
+    /[\uD800-\uDBFF](?![\uDC00-\uDFFF])|(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/g,
+    '\uFFFD',
+  )
+}
+
 function normalizePdfPageNumber(value, pageCount) {
   const number = typeof value === 'string' && value.trim() !== '' ? Number(value) : value
   if (!Number.isInteger(number) || number < 1 || number > pageCount) {
@@ -815,7 +842,7 @@ async function extractPageText(entry, pdfPage) {
     const raw = content.items
       .map((item) => (typeof item?.str === 'string' ? item.str : ''))
       .join(' ')
-    return normalizePageText(raw)
+    return wellFormedText(normalizePageText(raw))
   } finally {
     try {
       page.cleanup()

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-book-reader",
-  "version": "0.3.2",
+  "version": "0.3.3",
   "description": "Read books and papers in DSH: a reading toolbar on the official PDF preview that remembers your page and zoom, plus PDF tools for the model (page map, full-text search, page text and images).",
   "keywords": [
     "dsh",