feat: Slim dictionary (#2583)

Co-authored-by: Ryan Hitchman <hitchmanr@gmail.com>
Co-authored-by: Kurtis Grant <kurtis.b.grant@gmail.com>
Co-authored-by: kemonine <kemonine@kemonine.info>
Co-authored-by: DustinHu <hu.dustin@gmail.com>
Co-authored-by: Justin Mitchell <justin@jmitch.com>

README.md

@@ -12,7 +12,7 @@ CrossPoint is open-source e-reader firmware - community-built, fully hackable, f
 
 ## What can CrossPoint do?
 
-- **Reader engine**: EPUB 2/3 rendering with embedded-style option, image handling, hyphenation, kerning, chapter navigation, footnotes, bookmarks, go-to-percent, auto page turn, orientation control, focus reading, KOReader progress sync and more. 
+- **Reader engine**: EPUB 2/3 rendering with embedded-style option, image handling, hyphenation, kerning, chapter navigation, footnotes, bookmarks, dictionary lookups ([StarDict](docs/dictionary.md)), go-to-percent, auto page turn, orientation control, focus reading, KOReader progress sync and more. 
 
 - **Various formats**: native handling for `.epub`, `.xtc/.xtch`, `.txt`, and `.bmp`.
 
@@ -42,8 +42,6 @@ CrossPoint is open-source e-reader firmware - community-built, fully hackable, f
 
 ### Coming soon:
 
-- Dictionary lookup — inline word lookup without leaving the reader.
-
 - More themes.
 
 - Much more! stay tuned.

USER_GUIDE.md

@@ -263,6 +263,8 @@ The Settings screen allows you to configure the device's behavior. There are a f
   - "ON" - Vertical space will be added between paragraphs in Reading Mode
   - "OFF" - Paragraphs will not have vertical space added, but will have first-line indentation
 
+- **Dictionary**: Select the StarDict dictionary used for word lookups while reading, or "None" to disable lookups. *(Only shown when at least one dictionary folder exists under `/dictionaries/` on the SD card — see [docs/dictionary.md](docs/dictionary.md) for setup and usage.)*
+
 - **Text Anti-Aliasing**: Whether to show smooth grey edges (anti-aliasing) on text in reading mode. Note this slows down page turns slightly.
 
 - **Images**: Whether to display embedded images (JPG/PNG) found in EPUB files; options are "ON" (default) or "OFF".
@@ -282,6 +284,7 @@ The Settings screen allows you to configure the device's behavior. There are a f
 - **Long-press Menu**: Selects the function bound to holding the menu button (Confirm) while reading an EPUB. **Cycles through the available functions** each time the setting is selected — additional functions may be added in future releases, so this is not a binary on/off toggle. A short press of Confirm always opens the reader menu as normal:
   - "Bookmark" (default) - Hold Confirm (~0.4 second) to drop a bookmark at the current page.
   - "KOSync" - Hold Confirm (~1 second) to launch KOReader sync directly.
+  - "Dictionary" - Hold Confirm (~0.4 second) to start dictionary word selection on the current page (see [docs/dictionary.md](docs/dictionary.md)).
   - "Disabled" - Long-press is ignored; only short-press opens the reader menu.
 
 - **Short Power Button Click**: Controls the effect of a short click of the power button:
@@ -583,7 +586,7 @@ If the device goes to sleep or you close the book while viewing a footnote, the
 * **Return to Home:** Press the **Back** button to close the book and return to the **[Home](#31-home-screen)** screen.
 * **Return to Browse Files:** Press and hold the **Back** button to close the book and return to the **[Browse Files](#33-browse-files-screen)** screen.
 * **Reader Menu:** Press **Confirm** to open the **[Reader Menu](#5-reader-menu)**, which includes chapter navigation, reading options, and more.
-* **Long-press Confirm (configurable):** Holding **Confirm** runs the function chosen by the **Long-press Menu** setting in **[Controls Settings](#363-controls)** — "Bookmark" (default) drops a bookmark, "KOSync" launches KOReader Sync, "Disabled" does nothing. A short press always opens the Reader Menu.
+* **Long-press Confirm (configurable):** Holding **Confirm** runs the function chosen by the **Long-press Menu** setting in **[Controls Settings](#363-controls)** — "Bookmark" (default) drops a bookmark, "KOSync" launches KOReader Sync, "Dictionary" starts a word lookup, "Disabled" does nothing. A short press always opens the Reader Menu.
 
 ### Supported Languages
 
@@ -605,6 +608,7 @@ Available options include:
 
 - **Select Chapter** – Open the table of contents to jump to a specific chapter (see [Chapter Selection](#51-chapter-selection) below).
 - **Footnotes** – Navigate to the footnotes for the current section *(only shown in books that contain footnotes)*.
+- **Look Up** – Select a word on the current page and show its dictionary definition (see [docs/dictionary.md](docs/dictionary.md)). Requires a dictionary to be selected in **Settings → Reader → Dictionary**.
 - **Reading Orientation** – Cycle through screen orientations without leaving the reader.
 - **Auto Turn (Pages Per Minute)** – Cycle through automatic page turn speed options for hands-free reading.
 - **Go to %** – Jump to a specific position in the book by percentage.

docs/dictionary.md

@@ -0,0 +1,51 @@
+# Dictionary
+
+Look up words while reading an EPUB using an offline StarDict dictionary stored on the SD card.
+
+## Supported Format
+
+The reader supports **StarDict** dictionaries. When searching for dictionaries online, look for "StarDict format" or files with `.dict`, `.idx`, and `.ifo` extensions.
+
+A dictionary folder must contain:
+
+- `.idx` — word index (required, **must be uncompressed** — a `.idx.gz` will not work; decompress it on your computer with `gzip -d` first)
+- `.dict` or `.dict.dz` — definition data (`.dict.dz` is supported as-is; entries are decompressed on the fly during lookup)
+- `.ifo` — metadata (optional)
+
+Not supported: `.syn` synonym files (ignored), dictionaries with 64-bit index offsets (`idxoffsetbits=64` in the `.ifo` — rare, and rejected with an error), and HTML-formatted definitions render as raw markup rather than styled text.
+
+## Setting Up a Dictionary
+
+1. Copy your dictionary folder(s) to `/dictionaries/` on the SD card — one dictionary per folder, e.g. `/dictionaries/webster/webster.idx` + `webster.dict.dz`. A hidden `/.dictionaries/` folder (dot-prefixed) works the same way, for keeping it out of the file browser.
+2. Open **Settings → Reader → Dictionary** on the device.
+3. Select a dictionary from the list, or **None** to disable lookups.
+
+The Dictionary setting only appears when at least one usable dictionary folder exists. Folders containing more than one dictionary (multiple `.idx` stems) are skipped as ambiguous.
+
+## Looking Up a Word
+
+Two ways to start a lookup while reading:
+
+- Open the reader menu (**Confirm**) and choose **Look Up**.
+- Or set **Settings → Controls → Long-press Menu** to "Dictionary", then hold **Confirm** (~0.4s) on the reading page.
+
+One word on the page becomes highlighted:
+
+1. Use **Left/Right** to move between words in reading order, and the side **Up/Down** buttons to jump between lines.
+2. Press **Confirm** to look up the highlighted word.
+3. Press **Back** to return to the reader.
+
+On the very first lookup with a dictionary (and again if the dictionary file changes), the reader shows *"Indexing dictionary…"* while it builds a small `.qidx` sidecar file next to the `.idx`. This takes a few seconds for large dictionaries and makes all subsequent lookups fast. The sidecar can be deleted safely at any time — it will simply be rebuilt.
+
+### How Lookup Works
+
+1. **Direct match** — the word is found as-is (case-insensitive) in the dictionary index. Surrounding punctuation is ignored.
+2. **Stemming** — on a miss, common English word forms are retried automatically: possessives and plurals (`dogs` → `dog`, `stories` → `story`) and verb endings (`walked` → `walk`, `running` → `run`, `making` → `make`).
+3. **Not found** — a short popup appears and you return to word selection.
+
+## The Definition Screen
+
+When a word is found, the definition screen shows the matched headword at the top and the definition text below, with a page counter for long definitions.
+
+- **Left/Right** or side **Up/Down** — previous / next page
+- **Back** — return to word selection

lib/GfxRenderer/GfxRenderer.cpp

@@ -213,6 +213,61 @@ static inline void rotateCoordinates(const GfxRenderer::Orientation orientation,
   }
 }
 
+// Output of screenRectToAlignedMemRect: a rectangle in panel-memory
+// coordinates whose x and width are guaranteed to be multiples of 8 (the
+// SDK's EInkDisplay::displayWindow alignment requirement). `valid == false`
+// means the input was empty or fully outside the panel.
+struct AlignedMemRect {
+  uint16_t x = 0;
+  uint16_t y = 0;
+  uint16_t w = 0;
+  uint16_t h = 0;
+  bool valid = false;
+};
+
+// Translate a screen-coordinate rectangle (the coordinate system used by
+// fillRect / drawText / the rest of the renderer's public API) into a
+// panel-memory rectangle suitable for direct framebuffer indexing. Rotates
+// the rectangle's two opposite corners with rotateCoordinates(), takes the
+// bounding box (which naturally swaps width/height in Portrait /
+// PortraitInverted), then snaps the x extent outward to multiples of 8 and
+// clamps to panel bounds. Precondition: panel dims are multiples of 8 (true
+// for the 800x480 panel), so clamping cannot re-break alignment.
+static AlignedMemRect screenRectToAlignedMemRect(GfxRenderer::Orientation orientation, int sx, int sy, int sw, int sh,
+                                                 uint16_t panelWidth, uint16_t panelHeight) {
+  AlignedMemRect out;
+  if (sw <= 0 || sh <= 0) return out;
+
+  int x0, y0, x1, y1;
+  rotateCoordinates(orientation, sx, sy, &x0, &y0, panelWidth, panelHeight);
+  rotateCoordinates(orientation, sx + sw - 1, sy + sh - 1, &x1, &y1, panelWidth, panelHeight);
+
+  const int memXLo = std::min(x0, x1);
+  const int memYLo = std::min(y0, y1);
+  const int memXHi = std::max(x0, x1) + 1;  // exclusive upper bound
+  const int memYHi = std::max(y0, y1) + 1;
+
+  // Snap x outward to multiples of 8.
+  int alignedXLo = memXLo & ~0x7;        // round down
+  int alignedXHi = (memXHi + 7) & ~0x7;  // round up
+
+  if (alignedXLo < 0) alignedXLo = 0;
+  if (alignedXHi > panelWidth) alignedXHi = panelWidth;
+  int clampedYLo = memYLo;
+  int clampedYHi = memYHi;
+  if (clampedYLo < 0) clampedYLo = 0;
+  if (clampedYHi > panelHeight) clampedYHi = panelHeight;
+
+  if (alignedXHi <= alignedXLo || clampedYHi <= clampedYLo) return out;
+
+  out.x = static_cast<uint16_t>(alignedXLo);
+  out.y = static_cast<uint16_t>(clampedYLo);
+  out.w = static_cast<uint16_t>(alignedXHi - alignedXLo);
+  out.h = static_cast<uint16_t>(clampedYHi - clampedYLo);
+  out.valid = true;
+  return out;
+}
+
 enum class TextRotation { None, Rotated90CW };
 
 // Shared glyph rendering logic for normal and rotated text.
@@ -1451,6 +1506,39 @@ void GfxRenderer::displayBuffer(const HalDisplay::RefreshMode refreshMode) const
   display.displayBuffer(refreshMode, fadingFix);
 }
 
+size_t GfxRenderer::readFramebufferRegion(int x, int y, int w, int h, uint8_t* dst, size_t dstCapacity) const {
+  if (dst == nullptr || w <= 0 || h <= 0) return 0;
+
+  const AlignedMemRect mem = screenRectToAlignedMemRect(orientation, x, y, w, h, panelWidth, panelHeight);
+  if (!mem.valid) return 0;
+
+  const size_t rowBytes = mem.w / 8;  // exact: mem.w is a multiple of 8
+  const size_t needed = rowBytes * mem.h;
+  if (needed > dstCapacity) return 0;
+
+  for (uint16_t row = 0; row < mem.h; ++row) {
+    const uint8_t* srcRow = frameBuffer + (static_cast<uint32_t>(mem.y + row) * panelWidthBytes) + (mem.x / 8);
+    uint8_t* dstRow = dst + (static_cast<size_t>(row) * rowBytes);
+    memcpy(dstRow, srcRow, rowBytes);
+  }
+  return needed;
+}
+
+void GfxRenderer::writeFramebufferRegion(int x, int y, int w, int h, const uint8_t* src) {
+  if (src == nullptr || w <= 0 || h <= 0) return;
+
+  const AlignedMemRect mem = screenRectToAlignedMemRect(orientation, x, y, w, h, panelWidth, panelHeight);
+  if (!mem.valid) return;
+
+  const size_t rowBytes = mem.w / 8;  // exact: mem.w is a multiple of 8
+
+  for (uint16_t row = 0; row < mem.h; ++row) {
+    const uint8_t* srcRow = src + (static_cast<size_t>(row) * rowBytes);
+    uint8_t* dstRow = frameBuffer + (static_cast<uint32_t>(mem.y + row) * panelWidthBytes) + (mem.x / 8);
+    memcpy(dstRow, srcRow, rowBytes);
+  }
+}
+
 std::string GfxRenderer::truncatedText(const int fontId, const char* text, const int maxWidth,
                                        const EpdFontFamily::Style style) const {
   if (!text || maxWidth <= 0) return "";

lib/GfxRenderer/GfxRenderer.h

@@ -189,6 +189,15 @@ class GfxRenderer {
   void drawBitmap1Bit(const Bitmap& bitmap, int x, int y, int maxWidth, int maxHeight) const;
   void fillPolygon(const int* xPoints, const int* yPoints, int numPoints, bool state = true) const;
 
+  // Snapshot / restore a screen-coordinate framebuffer region (byte-aligned in
+  // panel memory). readFramebufferRegion returns the bytes written to dst, or
+  // 0 when the region is empty, offscreen, or exceeds dstCapacity. Pass the
+  // same rectangle to writeFramebufferRegion to restore the saved pixels.
+  // Enables partial-repaint patterns (e.g. moving a selection highlight)
+  // without re-rendering the whole page.
+  size_t readFramebufferRegion(int x, int y, int w, int h, uint8_t* dst, size_t dstCapacity) const;
+  void writeFramebufferRegion(int x, int y, int w, int h, const uint8_t* src);
+
   // Text
   int getTextWidth(int fontId, const char* text, EpdFontFamily::Style style = EpdFontFamily::REGULAR,
                    BidiUtils::BidiBaseDir baseDir = BidiUtils::BidiBaseDir::AUTO) const;

lib/I18n/translations/english.yaml

@@ -26,6 +26,12 @@ STR_EMPTY_FILE: "Empty file"
 STR_OUT_OF_BOUNDS: "Out of bounds"
 STR_LOADING: "Loading..."
 STR_LOADING_POPUP: "Loading"
+STR_LOOKUP: "Look Up"
+STR_DICT_LOOKING_UP: "Looking up..."
+STR_DICT_INDEXING: "Indexing dictionary..."
+STR_DICT_NOT_FOUND: "Not found"
+STR_DICT_NO_DICT_SET: "No dictionary set"
+STR_DICT_ERROR: "Dictionary error"
 STR_WIFI_NETWORKS: "Wi-Fi Networks"
 STR_NO_NETWORKS: "No networks found"
 STR_NETWORKS_FOUND: "%zu networks found"
@@ -163,6 +169,7 @@ STR_PREV_NEXT: "Prev/Next"
 STR_NEXT_PREV: "Next/Prev"
 STR_KOSYNC: "KOSync"
 STR_BOOKMARK_OPTION: "Bookmark"
+STR_DICTIONARY: "Dictionary"
 STR_DISABLED: "Disabled"
 STR_NOTO_SERIF: "Noto Serif"
 STR_NOTO_SANS: "Noto Sans"

src/CrossPointSettings.h

@@ -153,6 +153,7 @@ class CrossPointSettings {
     LP_MENU_KOSYNC = 0,
     LP_MENU_DISABLED = 1,
     LP_MENU_BOOKMARK = 2,
+    LP_MENU_DICTIONARY = 3,
     LONG_PRESS_MENU_FUNCTION_COUNT
   };
 
@@ -265,6 +266,8 @@ class CrossPointSettings {
   uint8_t focusReadingEnabled = 0;
   // SD card font family name (empty = use built-in fontFamily)
   char sdFontFamilyName[32] = "";
+  // Dictionary folder name under /dictionaries (empty = no dictionary)
+  char dictionaryName[32] = "";
   // Show hidden files/directories (starting with '.') in the file browser (0 = hidden, 1 = show)
   uint8_t showHiddenFiles = 0;
   // Remove a book from the Recent Books list when its End-of-Book screen is reached (0 = off, 1 = on)

src/JsonSettingsIO.cpp

@@ -149,6 +149,10 @@ bool JsonSettingsIO::saveSettings(const CrossPointSettings& s, const char* path)
   if (s.sdFontFamilyName[0] != '\0') {
     doc["sdFontFamilyName"] = s.sdFontFamilyName;
   }
+  // Dictionary folder name — uses dynamic getter/setter in SettingsList, save manually
+  if (s.dictionaryName[0] != '\0') {
+    doc["dictionaryName"] = s.dictionaryName;
+  }
 
   // Language -- managed by LanguageSelectActivity, not in SettingsList.
   // Stored as ISO code string ("EN", "DE", ...) for stability across enum reorders.
@@ -256,6 +260,11 @@ bool JsonSettingsIO::loadSettings(CrossPointSettings& s, const char* json, bool*
     if (needsResave) *needsResave = true;
   }
 
+  // Dictionary folder name — uses dynamic getter/setter in SettingsList, load manually
+  const char* dictName = doc["dictionaryName"] | "";
+  strncpy(s.dictionaryName, dictName, sizeof(s.dictionaryName) - 1);
+  s.dictionaryName[sizeof(s.dictionaryName) - 1] = '\0';
+
   // Language -- stored as code string for stability across enum reorders.
   if (doc["language"].is<const char*>()) {
     s.language = static_cast<uint8_t>(I18n::languageFromCode(doc["language"].as<const char*>()));

src/SettingsList.h

@@ -13,6 +13,7 @@
 #include "CrossPointSettings.h"
 #include "KOReaderCredentialStore.h"
 #include "activities/settings/SettingsActivity.h"
+#include "util/DictionaryRegistry.h"
 
 // Build the font family setting dynamically. When registry is non-null, SD card fonts
 // are appended after the built-in fonts. Otherwise only built-in fonts are listed.
@@ -90,6 +91,47 @@ inline SettingInfo buildFontFamilySetting(const SdCardFontRegistry* registry) {
   return s;
 }
 
+// Build the dictionary selection setting dynamically from the folders discovered
+// under /dictionaries. "None" plus one option per dictionary; the selected folder
+// name persists in SETTINGS.dictionaryName (saved/loaded manually in
+// JsonSettingsIO — the generic loop skips dynamic entries).
+inline SettingInfo buildDictionarySetting(const std::vector<DictionaryEntry>& dictionaries) {
+  std::vector<std::string> folderNames;
+  folderNames.reserve(dictionaries.size());
+  std::transform(dictionaries.begin(), dictionaries.end(), std::back_inserter(folderNames),
+                 [](const DictionaryEntry& d) { return d.name; });
+
+  SettingInfo s;
+  s.nameId = StrId::STR_DICTIONARY;
+  s.type = SettingType::ENUM;
+  s.enumStringValues.reserve(folderNames.size() + 1);
+  s.enumStringValues.push_back(I18N.get(StrId::STR_NONE_OPT));
+  s.enumStringValues.insert(s.enumStringValues.end(), folderNames.begin(), folderNames.end());
+  s.category = StrId::STR_CAT_READER;
+
+  s.valueGetter = [folderNames]() -> uint8_t {
+    for (size_t i = 0; i < folderNames.size(); i++) {
+      // Compare within the settings field capacity: an over-long folder name is
+      // stored truncated, and must still match its list entry.
+      if (strncmp(folderNames[i].c_str(), SETTINGS.dictionaryName, sizeof(SETTINGS.dictionaryName) - 1) == 0) {
+        return static_cast<uint8_t>(i + 1);
+      }
+    }
+    return 0;  // "None", also when the stored folder no longer exists
+  };
+
+  s.valueSetter = [folderNames](uint8_t v) {
+    if (v == 0 || v > folderNames.size()) {
+      SETTINGS.dictionaryName[0] = '\0';
+      return;
+    }
+    strncpy(SETTINGS.dictionaryName, folderNames[v - 1].c_str(), sizeof(SETTINGS.dictionaryName) - 1);
+    SETTINGS.dictionaryName[sizeof(SETTINGS.dictionaryName) - 1] = '\0';
+  };
+
+  return s;
+}
+
 // Shared settings list used by both the device settings UI and the web settings API.
 // Each entry has a key (for JSON API) and category (for grouping).
 // ACTION-type entries and entries without a key are device-only.
@@ -99,7 +141,8 @@ inline SettingInfo buildFontFamilySetting(const SdCardFontRegistry* registry) {
 // SdCardFontRegistry is supplied AND has SD card fonts installed, the
 // font-family entry is replaced in a per-call copy with a registry-aware
 // version. Callers without SD fonts pay only a vector copy.
-inline std::vector<SettingInfo> getSettingsList(const SdCardFontRegistry* registry = nullptr) {
+inline std::vector<SettingInfo> getSettingsList(const SdCardFontRegistry* registry = nullptr,
+                                                const std::vector<DictionaryEntry>* dictionaries = nullptr) {
   static const std::vector<SettingInfo> baseList = [] {
     std::vector<SettingInfo> v = {
         // --- Display ---
@@ -173,8 +216,8 @@ inline std::vector<SettingInfo> getSettingsList(const SdCardFontRegistry* regist
                            StrId::STR_LONG_PRESS_BEHAVIOR_ORIENTATION},
                           "longPressButtonBehavior", StrId::STR_CAT_CONTROLS),
         SettingInfo::Enum(StrId::STR_LONG_PRESS_MENU, &CrossPointSettings::longPressMenuFunction,
-                          {StrId::STR_KOSYNC, StrId::STR_DISABLED, StrId::STR_BOOKMARK_OPTION}, "longPressMenuFunction",
+                          {StrId::STR_KOSYNC, StrId::STR_DISABLED, StrId::STR_BOOKMARK_OPTION, StrId::STR_DICTIONARY},
-                          StrId::STR_CAT_CONTROLS),
+                          "longPressMenuFunction", StrId::STR_CAT_CONTROLS),
         SettingInfo::Enum(
             StrId::STR_SHORT_PWR_BTN, &CrossPointSettings::shortPwrBtn,
             {StrId::STR_IGNORE, StrId::STR_SLEEP, StrId::STR_PAGE_TURN, StrId::STR_FORCE_REFRESH, StrId::STR_FOOTNOTES},
@@ -308,5 +351,11 @@ inline std::vector<SettingInfo> getSettingsList(const SdCardFontRegistry* regist
       *it = buildFontFamilySetting(registry);
     }
   }
+  if (dictionaries && !dictionaries->empty()) {
+    // Insert at the end of the Reader category (just before the first Controls entry).
+    auto it =
+        std::find_if(v.begin(), v.end(), [](const SettingInfo& s) { return s.category == StrId::STR_CAT_CONTROLS; });
+    v.insert(it, buildDictionarySetting(*dictionaries));
+  }
   return v;
 }

src/activities/reader/DictionaryDefinitionActivity.cpp

@@ -0,0 +1,237 @@
+#include "DictionaryDefinitionActivity.h"
+
+#include <FontCacheManager.h>
+#include <GfxRenderer.h>
+#include <I18n.h>
+
+#include <algorithm>
+#include <cstdint>
+#include <cstdio>
+
+#include "CrossPointSettings.h"
+#include "components/UITheme.h"
+#include "fontIds.h"
+#include "util/HtmlToPlainText.h"
+
+namespace {
+
+// Longest measurable/drawable span. Wrapped lines stay under the screen width
+// (far below this); only pathological unbreakable tokens are split at this cap.
+constexpr size_t MAX_LINE_BYTES = 191;
+
+// Body text left/right inset, matching the reader's default feel.
+constexpr int SIDE_PADDING = 20;
+
+}  // namespace
+
+void DictionaryDefinitionActivity::onEnter() {
+  Activity::onEnter();
+  // Normalize StarDict multi-type separators so the wrap loop and the
+  // C-string font APIs below both see the whole definition.
+  std::replace(definition.begin(), definition.end(), '\0', '\n');
+  definition = htmlToPlainText(definition);
+  wrapText();
+  requestUpdate();
+}
+
+int DictionaryDefinitionActivity::measureSpan(const int fontId, const char* text, size_t len) const {
+  char buf[MAX_LINE_BYTES + 1];
+  len = std::min(len, MAX_LINE_BYTES);
+  memcpy(buf, text, len);
+  buf[len] = '\0';
+  return renderer.getTextAdvanceX(fontId, buf, EpdFontFamily::REGULAR);
+}
+
+// Greedy word-wrap of `definition` into byte spans. '\n' breaks lines (blank
+// lines survive as paragraph spacing; NULs from multi-type StarDict entries
+// were normalized to newlines in onEnter); '\r' is dropped by treating it as
+// a space at a token edge.
+void DictionaryDefinitionActivity::wrapText() {
+  lines.clear();
+  lines.reserve(definition.size() / 32 + 8);
+
+  const int fontId = SETTINGS.getReaderFontId();
+  // SD-card fonts: merge every definition codepoint into the persistent
+  // advance table up front. Otherwise each unseen codepoint measured below
+  // falls back to an on-demand glyph load from SD (8-slot overflow ring).
+  renderer.ensureSdCardFontReady(fontId, definition.c_str(), 0x01 /* REGULAR */);
+
+  const auto& metrics = UITheme::getInstance().getMetrics();
+  const auto orientation = renderer.getOrientation();
+  const bool isLandscape = orientation == GfxRenderer::Orientation::LandscapeClockwise ||
+                           orientation == GfxRenderer::Orientation::LandscapeCounterClockwise;
+  const bool isInverted = orientation == GfxRenderer::Orientation::PortraitInverted;
+  const int hintGutterWidth = isLandscape ? metrics.sideButtonHintsWidth : 0;
+  const int maxWidth = renderer.getScreenWidth() - hintGutterWidth - 2 * SIDE_PADDING;
+  const int spaceWidth = renderer.getSpaceWidth(fontId, EpdFontFamily::REGULAR);
+
+  const int lineHeight = renderer.getLineHeight(fontId);
+  const int topArea = (isInverted ? metrics.buttonHintsHeight : 0) + metrics.topPadding + metrics.headerHeight;
+  const int bottomArea = metrics.buttonHintsHeight + metrics.verticalSpacing;
+  linesPerPage = std::max(1, (renderer.getScreenHeight() - topArea - bottomArea) / lineHeight);
+
+  const char* text = definition.c_str();
+  const uint32_t n = static_cast<uint32_t>(definition.size());
+  uint32_t lineStart = 0;
+  uint32_t lineEnd = 0;  // one past the last token byte on the current line
+  int lineWidth = 0;
+
+  const auto flushLine = [&](uint32_t nextStart) {
+    lines.push_back({lineStart, static_cast<uint16_t>(lineEnd - lineStart)});
+    lineStart = nextStart;
+    lineEnd = nextStart;
+    lineWidth = 0;
+  };
+
+  uint32_t i = 0;
+  while (i < n) {
+    const char c = text[i];
+    if (c == '\n' || c == '\0') {
+      flushLine(i + 1);
+      i++;
+      continue;
+    }
+    if (c == ' ' || c == '\t' || c == '\r') {
+      i++;
+      continue;
+    }
+
+    // Token: run of non-whitespace bytes, capped at the measure buffer.
+    const uint32_t tokenStart = i;
+    while (i < n && text[i] != ' ' && text[i] != '\t' && text[i] != '\r' && text[i] != '\n' && text[i] != '\0' &&
+           i - tokenStart < MAX_LINE_BYTES) {
+      i++;
+    }
+    // If the byte cap cut the token mid-UTF-8-sequence, back off to the last
+    // complete codepoint so measure/draw never see a partial sequence. A
+    // natural stop lands on whitespace or the terminating NUL, never on a
+    // continuation byte, so this is a no-op there.
+    while (i - tokenStart > 1 && (text[i] & 0xC0) == 0x80) i--;
+    const uint32_t tokenLen = i - tokenStart;
+    const int tokenWidth = measureSpan(fontId, text + tokenStart, tokenLen);
+
+    if (lineEnd == lineStart) {
+      lineStart = tokenStart;
+      lineEnd = tokenStart + tokenLen;
+      lineWidth = tokenWidth;
+    } else if (lineWidth + spaceWidth + tokenWidth <= maxWidth &&
+               tokenStart + tokenLen - lineStart <= UINT16_MAX) {  // span len must fit Line::len
+      lineEnd = tokenStart + tokenLen;
+      lineWidth += spaceWidth + tokenWidth;
+    } else {
+      flushLine(tokenStart);
+      lineEnd = tokenStart + tokenLen;
+      lineWidth = tokenWidth;
+    }
+
+    // An unbreakable token wider than the screen is now alone on the line
+    // (any previous content was flushed above): split it at the widest
+    // fitting UTF-8 boundary and carry the remainder forward.
+    while (lineWidth > maxWidth && lineEnd - lineStart > 1) {
+      const uint32_t len = lineEnd - lineStart;
+      uint32_t lastFit = 0;
+      for (uint32_t f = 1; f <= len; f++) {
+        if (f == len || (text[lineStart + f] & 0xC0) != 0x80) {  // codepoint boundary
+          if (measureSpan(fontId, text + lineStart, f) > maxWidth) break;
+          lastFit = f;
+        }
+      }
+      if (lastFit == 0) {
+        // Even a single over-wide glyph must make progress; consume its whole
+        // UTF-8 sequence rather than splitting it into invalid fragments.
+        lastFit = 1;
+        while (lastFit < len && (text[lineStart + lastFit] & 0xC0) == 0x80) lastFit++;
+      }
+      const uint32_t rest = lineStart + lastFit;
+      lineEnd = rest;
+      flushLine(rest);
+      lineEnd = rest + (len - lastFit);
+      lineWidth = measureSpan(fontId, text + lineStart, lineEnd - lineStart);
+    }
+  }
+  if (lineEnd > lineStart) flushLine(n);
+
+  // Trim trailing blank lines so the last page is not empty padding.
+  while (!lines.empty() && lines.back().len == 0) lines.pop_back();
+
+  totalPages = std::max(1, (static_cast<int>(lines.size()) + linesPerPage - 1) / linesPerPage);
+  currentPage = 0;
+}
+
+void DictionaryDefinitionActivity::loop() {
+  if (mappedInput.wasReleased(MappedInputManager::Button::Back)) {
+    finish();
+    return;
+  }
+
+  buttonNavigator.onNext([this] {
+    if (currentPage + 1 < totalPages) {
+      currentPage++;
+      requestUpdate();
+    }
+  });
+
+  buttonNavigator.onPrevious([this] {
+    if (currentPage > 0) {
+      currentPage--;
+      requestUpdate();
+    }
+  });
+}
+
+// Draws the current page's line spans (copied into a stack buffer for NUL
+// termination). Called twice per render: once in font-cache scan mode, once
+// for the real paint.
+void DictionaryDefinitionActivity::drawBody(const int fontId, const int x, const int startY) const {
+  const int lineHeight = renderer.getLineHeight(fontId);
+  char buf[MAX_LINE_BYTES + 1];
+  const int firstLine = currentPage * linesPerPage;
+  const int lastLine = std::min(firstLine + linesPerPage, static_cast<int>(lines.size()));
+  for (int i = firstLine; i < lastLine; i++) {
+    if (lines[i].len == 0) continue;
+    const size_t len = std::min(static_cast<size_t>(lines[i].len), MAX_LINE_BYTES);
+    memcpy(buf, definition.c_str() + lines[i].start, len);
+    buf[len] = '\0';
+    renderer.drawText(fontId, x, startY + (i - firstLine) * lineHeight, buf);
+  }
+}
+
+void DictionaryDefinitionActivity::render(RenderLock&&) {
+  renderer.clearScreen();
+
+  const auto& metrics = UITheme::getInstance().getMetrics();
+  const auto orientation = renderer.getOrientation();
+  const bool isLandscapeCw = orientation == GfxRenderer::Orientation::LandscapeClockwise;
+  const bool isLandscapeCcw = orientation == GfxRenderer::Orientation::LandscapeCounterClockwise;
+  const bool isInverted = orientation == GfxRenderer::Orientation::PortraitInverted;
+  const int hintGutterWidth = (isLandscapeCw || isLandscapeCcw) ? metrics.sideButtonHintsWidth : 0;
+  const int contentX = isLandscapeCw ? hintGutterWidth : 0;
+  const int contentWidth = renderer.getScreenWidth() - hintGutterWidth;
+  const int contentY = isInverted ? metrics.buttonHintsHeight : 0;
+
+  // Header: matched headword left, page counter right.
+  const int headerY = contentY + metrics.topPadding + 10;
+  renderer.drawText(UI_12_FONT_ID, contentX + SIDE_PADDING, headerY, headword.c_str(), true, EpdFontFamily::BOLD);
+  if (totalPages > 1) {
+    char counter[16];
+    snprintf(counter, sizeof(counter), "%d/%d", currentPage + 1, totalPages);
+    const int counterWidth = renderer.getTextWidth(UI_10_FONT_ID, counter);
+    renderer.drawText(UI_10_FONT_ID, contentX + contentWidth - SIDE_PADDING - counterWidth, headerY, counter);
+  }
+
+  // Body: two-pass draw inside a prewarm scope (same pattern as the reader's
+  // renderContents) so SD-card font glyphs load from SD in one batch instead
+  // of one on-demand overflow read per character on every page turn.
+  const int fontId = SETTINGS.getReaderFontId();
+  const int bodyStartY = contentY + metrics.topPadding + metrics.headerHeight;
+  auto* fcm = renderer.getFontCacheManager();
+  auto scope = fcm->createPrewarmScope();
+  drawBody(fontId, contentX + SIDE_PADDING, bodyStartY);  // scan pass: records codepoints only
+  scope.endScanAndPrewarm();
+  drawBody(fontId, contentX + SIDE_PADDING, bodyStartY);
+
+  const auto labels =
+      mappedInput.mapLabels(tr(STR_BACK), "", (currentPage > 0 ? "<" : ""), (currentPage + 1 < totalPages ? ">" : ""));
+  GUI.drawButtonHints(renderer, labels.btn1, labels.btn2, labels.btn3, labels.btn4);
+  renderer.displayBuffer();
+}

src/activities/reader/DictionaryDefinitionActivity.h

@@ -0,0 +1,46 @@
+#pragma once
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "activities/Activity.h"
+#include "util/ButtonNavigator.h"
+
+// Paged plain-text viewer for one dictionary definition. The definition is
+// word-wrapped once on entry; each page renders spans of the original string,
+// so no per-line copies are held.
+class DictionaryDefinitionActivity final : public Activity {
+ public:
+  explicit DictionaryDefinitionActivity(GfxRenderer& renderer, MappedInputManager& mappedInput, std::string headword,
+                                        std::string definition)
+      : Activity("DictionaryDefinition", renderer, mappedInput),
+        headword(std::move(headword)),
+        definition(std::move(definition)) {}
+
+  void onEnter() override;
+  void loop() override;
+  void render(RenderLock&&) override;
+
+ private:
+  // One wrapped display line: a byte span of `definition`. Wrapping keeps
+  // lines under the screen width, so uint16_t length is ample.
+  struct Line {
+    uint32_t start;
+    uint16_t len;
+  };
+
+  void wrapText();
+  int measureSpan(int fontId, const char* text, size_t len) const;
+  void drawBody(int fontId, int x, int startY) const;
+
+  const std::string headword;
+  // Not const: onEnter() normalizes embedded NULs (StarDict multi-type
+  // separators) to newlines so C-string APIs see the whole text.
+  std::string definition;
+  std::vector<Line> lines;
+  int currentPage = 0;
+  int totalPages = 1;
+  int linesPerPage = 1;
+  ButtonNavigator buttonNavigator;
+};

src/activities/reader/DictionaryWordSelectActivity.cpp

@@ -0,0 +1,301 @@
+#include "DictionaryWordSelectActivity.h"
+
+#include <FontCacheManager.h>
+#include <GfxRenderer.h>
+#include <Memory.h>
+#include <freertos/FreeRTOS.h>
+#include <freertos/task.h>
+
+#include <cctype>
+#include <climits>
+#include <cstdlib>
+
+#include "CrossPointSettings.h"
+#include "DictionaryDefinitionActivity.h"
+#include "components/UITheme.h"
+
+namespace {
+
+constexpr unsigned long POPUP_DURATION_MS = 1500;
+
+// A token is selectable when it has an ASCII alphanumeric or a non-ASCII
+// codepoint outside U+2000-U+206F (dashes, bullets and other General
+// Punctuation that appear as standalone tokens are not words).
+bool isSelectableToken(const char* text) {
+  for (const uint8_t* p = reinterpret_cast<const uint8_t*>(text); *p != 0; p++) {
+    if (*p < 0x80) {
+      if (std::isalnum(*p)) return true;
+    } else if (*p == 0xE2 && (p[1] == 0x80 || p[1] == 0x81)) {
+      if (p[2] == 0) break;  // truncated sequence: skipping would step past the NUL
+      p += 2;                // skip the 3-byte General Punctuation codepoint
+    } else {
+      return true;
+    }
+  }
+  return false;
+}
+
+void indexBuildYield(void*) { vTaskDelay(1); }
+
+}  // namespace
+
+void DictionaryWordSelectActivity::onEnter() {
+  Activity::onEnter();
+  fontId = SETTINGS.getReaderFontId();
+  lineHeight = renderer.getLineHeight(fontId);
+  // No null check: a failed allocation just disables the differential
+  // fast path (drawHighlightWithSnapshot skips the read), keeping the
+  // full-repaint path as the fallback.
+  snapshot = makeUniqueNoThrow<uint8_t[]>(SNAPSHOT_CAPACITY);
+  extractWords();
+  // Start on the middle row's word nearest mid-screen instead of top-left:
+  // any word on the page is then at most half a page of moves away.
+  if (!words.empty()) {
+    const int initial = closestInRow(rowCount / 2, renderer.getScreenWidth() / 2);
+    if (initial >= 0) selected = initial;
+  }
+  requestUpdate();
+}
+
+void DictionaryWordSelectActivity::extractWords() {
+  words.clear();
+  words.reserve(128);
+  rowCount = 0;
+
+  // Single walk: collect the selectable words while accumulating their text
+  // and styles (~2KB transient string, freed on return). Widths are measured
+  // afterwards: merging the page's codepoints into the SD font's persistent
+  // advance table first keeps getTextAdvanceX on the in-RAM path instead of
+  // loading glyphs from SD one overflow slot at a time.
+  std::string pageText;
+  pageText.reserve(2048);
+  uint8_t styleMask = 0;
+
+  for (const auto& element : page->elements) {
+    if (element->getTag() != TAG_PageLine) continue;
+    const auto* line = static_cast<const PageLine*>(element.get());
+    const auto& block = line->getBlock();
+    if (!block || !block->valid()) continue;
+
+    bool rowHasWords = false;
+    for (uint16_t i = 0; i < block->wordCount(); i++) {
+      const char* text = block->wordText(i);
+      if (!isSelectableToken(text)) continue;
+
+      WordBox box;
+      box.x = static_cast<int16_t>(line->xPos + block->wordXpos(i) + marginLeft);
+      box.y = static_cast<int16_t>(line->yPos + marginTop);
+      box.style = block->wordStyle(i);
+      box.width = 0;  // measured below, once the advance table is ready
+      box.row = rowCount;
+      box.text = text;
+      words.push_back(box);
+      rowHasWords = true;
+
+      pageText.append(text);
+      pageText.push_back(' ');
+      styleMask |= static_cast<uint8_t>(1u << (static_cast<uint8_t>(box.style) & 0x03));
+    }
+    if (rowHasWords) rowCount++;
+  }
+
+  if (styleMask == 0) styleMask = 0x01;  // REGULAR
+  renderer.ensureSdCardFontReady(fontId, pageText.c_str(), styleMask);
+  for (auto& word : words) {
+    word.width = static_cast<int16_t>(renderer.getTextAdvanceX(fontId, word.text, word.style));
+  }
+}
+
+// Index of the word in `row` whose horizontal center is closest to centerX;
+// -1 when the row has no words.
+int DictionaryWordSelectActivity::closestInRow(const uint16_t row, const int centerX) const {
+  int best = -1;
+  int bestDistance = INT_MAX;
+  for (int i = 0; i < static_cast<int>(words.size()); i++) {
+    if (words[i].row != row) continue;
+    const int distance = std::abs(words[i].x + words[i].width / 2 - centerX);
+    if (distance < bestDistance) {
+      bestDistance = distance;
+      best = i;
+    }
+  }
+  return best;
+}
+
+void DictionaryWordSelectActivity::moveVertical(const int direction) {
+  const WordBox& current = words[selected];
+  const int targetRow = static_cast<int>(current.row) + direction;
+  if (targetRow < 0 || targetRow >= static_cast<int>(rowCount)) return;
+
+  const int best = closestInRow(static_cast<uint16_t>(targetRow), current.x + current.width / 2);
+  if (best >= 0 && best != selected) {
+    selected = best;
+    requestUpdate();
+  }
+}
+
+void DictionaryWordSelectActivity::performLookup() {
+  popup = Popup::Busy;
+  if (!dictOpenAttempted) {
+    dictOpenAttempted = true;
+    dictOpenOk = dict.open(SETTINGS.dictionaryName);
+  }
+  const bool indexing = dictOpenOk && dict.needsIndex();
+  popupMsg = indexing ? StrId::STR_DICT_INDEXING : StrId::STR_DICT_LOOKING_UP;
+  requestUpdateAndWait();  // paint the page + busy popup before blocking on SD
+
+  bool ok = dictOpenOk;
+  if (ok && indexing) ok = dict.buildIndex(&indexBuildYield);
+
+  std::string definition;
+  std::string headword;
+  const bool found = ok && dict.lookup(words[selected].text, definition, headword);
+
+  if (found) {
+    popup = Popup::None;
+    startActivityForResult(std::make_unique<DictionaryDefinitionActivity>(renderer, mappedInput, std::move(headword),
+                                                                          std::move(definition)),
+                           [this](const ActivityResult&) { requestUpdate(); });
+    return;
+  }
+  popup = ok ? Popup::NotFound : Popup::Error;
+  popupMsg = ok ? StrId::STR_DICT_NOT_FOUND : StrId::STR_DICT_ERROR;
+  popupTime = millis();
+  requestUpdate();
+}
+
+void DictionaryWordSelectActivity::loop() {
+  if (popup == Popup::NotFound || popup == Popup::Error) {
+    if (millis() - popupTime >= POPUP_DURATION_MS) {
+      popup = Popup::None;
+      requestUpdate();
+    }
+    return;
+  }
+
+  if (mappedInput.wasPressed(MappedInputManager::Button::Confirm)) confirmPressSeen = true;
+
+  if (mappedInput.wasReleased(MappedInputManager::Button::Back)) {
+    finish();
+    return;
+  }
+  if (mappedInput.wasReleased(MappedInputManager::Button::Confirm) && confirmPressSeen && !words.empty()) {
+    performLookup();
+    return;
+  }
+
+  if (words.empty()) return;
+  if (mappedInput.wasPressed(MappedInputManager::Button::Left) && selected > 0) {
+    selected--;
+    requestUpdate();
+  } else if (mappedInput.wasPressed(MappedInputManager::Button::Right) &&
+             selected + 1 < static_cast<int>(words.size())) {
+    selected++;
+    requestUpdate();
+  } else if (mappedInput.wasPressed(MappedInputManager::Button::Up)) {
+    moveVertical(-1);
+  } else if (mappedInput.wasPressed(MappedInputManager::Button::Down)) {
+    moveVertical(1);
+  }
+}
+
+// Saves the pixels under words[selected]'s highlight box, then draws the
+// highlight over them. Returns false when the pixels could not be saved
+// (no buffer / oversize box) — the highlight is drawn regardless, but the
+// next cursor move must do a full repaint.
+bool DictionaryWordSelectActivity::drawHighlightWithSnapshot() {
+  const WordBox& word = words[selected];
+  int hx = word.x - 2;
+  int hy = word.y - 2;
+  int hw = word.width + 4;
+  int hh = lineHeight + 4;
+  // Clamp to the panel so save, draw and restore all use the same box.
+  if (hx < 0) {
+    hw += hx;
+    hx = 0;
+  }
+  if (hy < 0) {
+    hh += hy;
+    hy = 0;
+  }
+
+  bool saved = false;
+  if (snapshot && hw > 0 && hh > 0) {
+    saved = renderer.readFramebufferRegion(hx, hy, hw, hh, snapshot.get(), SNAPSHOT_CAPACITY) > 0;
+  }
+  snapshotX = static_cast<int16_t>(hx);
+  snapshotY = static_cast<int16_t>(hy);
+  snapshotW = static_cast<int16_t>(hw);
+  snapshotH = static_cast<int16_t>(hh);
+  snapshotIdx = saved ? selected : -1;
+
+  renderer.fillRect(hx, hy, hw, hh, true);
+  renderer.drawText(fontId, word.x, word.y, word.text, false, word.style);
+  return saved;
+}
+
+// Front-button bar (Back/Confirm/Left/Right). Drawn last on every repaint
+// path, including the differential highlight-only path, so it always ends
+// up as the top layer even when a highlighted word's box falls under a
+// hint's screen area. No side-button hints: Up/Down row jump has no spare
+// screen area on this page (it reuses the reader's full-bleed layout), and
+// a hint box there would hide text instead of sitting in a reserved gutter.
+void DictionaryWordSelectActivity::drawHints() const {
+  // No selectable word on this page: Confirm/Left/Right are all no-ops
+  // (guarded by words.empty() in loop()/performLookup), so only Back does
+  // anything and only Back is hinted.
+  if (words.empty()) {
+    const auto labels = mappedInput.mapLabels(tr(STR_BACK), "", "", "");
+    GUI.drawButtonHints(renderer, labels.btn1, labels.btn2, labels.btn3, labels.btn4);
+    return;
+  }
+  const auto labels = mappedInput.mapLabels(tr(STR_BACK), tr(STR_LOOKUP), tr(STR_DIR_LEFT), tr(STR_DIR_RIGHT));
+  GUI.drawButtonHints(renderer, labels.btn1, labels.btn2, labels.btn3, labels.btn4);
+}
+
+void DictionaryWordSelectActivity::render(RenderLock&&) {
+  // Differential fast path: only the highlight moved and the framebuffer
+  // still holds a clean page (no popup or sub-activity since the last full
+  // repaint). Restore the pixels under the old highlight, draw the new one,
+  // and push — skipping the two-pass page render entirely.
+  if (popup == Popup::None && snapshotIdx >= 0 && !words.empty() && selected != snapshotIdx) {
+    renderer.writeFramebufferRegion(snapshotX, snapshotY, snapshotW, snapshotH, snapshot.get());
+    // The full path's PrewarmScope cleared the glyph cache on exit; batch-load
+    // just the highlighted word's glyphs before drawing them white-on-black.
+    renderer.getFontCacheManager()->prewarmCache(
+        fontId, words[selected].text, static_cast<uint8_t>(1u << (static_cast<uint8_t>(words[selected].style) & 0x03)));
+    if (drawHighlightWithSnapshot()) {
+      drawHints();
+      renderer.displayBuffer(HalDisplay::FAST_REFRESH);
+      return;
+    }
+    // Snapshot failed (oversize box) — fall through to a full repaint.
+  }
+
+  renderer.clearScreen();
+
+  // Same prewarm-scan-then-render pass the reader uses, so SD-card fonts hit
+  // the in-RAM glyph cache during the real draw.
+  auto* fcm = renderer.getFontCacheManager();
+  auto scope = fcm->createPrewarmScope();
+  page->render(renderer, fontId, marginLeft, marginTop);
+  scope.endScanAndPrewarm();
+  page->render(renderer, fontId, marginLeft, marginTop);
+
+  if (!words.empty()) {
+    drawHighlightWithSnapshot();
+  }
+
+  drawHints();
+
+  if (popup != Popup::None) {
+    // The popup overdraws the page, so the snapshot no longer matches the
+    // framebuffer — force the next render onto the full-repaint path.
+    snapshotIdx = -1;
+    // drawPopup overlays the framebuffer and refreshes the display itself.
+    // I18N.get directly: tr() only accepts literal key names.
+    GUI.drawPopup(renderer, I18N.get(popupMsg));
+    return;
+  }
+  renderer.displayBuffer(HalDisplay::FAST_REFRESH);
+}

src/activities/reader/DictionaryWordSelectActivity.h

@@ -0,0 +1,84 @@
+#pragma once
+
+#include <Epub/Page.h>
+#include <I18n.h>
+
+#include <memory>
+#include <vector>
+
+#include "activities/Activity.h"
+#include "util/Dictionary.h"
+
+// Button-driven word selection over the current reader page: Left/Right step
+// through words in reading order, Up/Down jump rows, Confirm looks the word up
+// and opens DictionaryDefinitionActivity, Back returns to the reader.
+class DictionaryWordSelectActivity final : public Activity {
+ public:
+  explicit DictionaryWordSelectActivity(GfxRenderer& renderer, MappedInputManager& mappedInput,
+                                        std::unique_ptr<Page> page, int marginLeft, int marginTop)
+      : Activity("DictionaryWordSelect", renderer, mappedInput),
+        page(std::move(page)),
+        marginLeft(marginLeft),
+        marginTop(marginTop) {}
+
+  void onEnter() override;
+  void loop() override;
+  void render(RenderLock&&) override;
+
+ private:
+  // Screen box of one selectable word. `text` points into the owned Page's
+  // TextBlock arena (NUL-terminated), valid for this activity's lifetime.
+  struct WordBox {
+    int16_t x;
+    int16_t y;
+    int16_t width;
+    uint16_t row;
+    const char* text;
+    EpdFontFamily::Style style;
+  };
+
+  enum class Popup : uint8_t { None, Busy, NotFound, Error };
+
+  void extractWords();
+  int closestInRow(uint16_t row, int centerX) const;
+  void moveVertical(int direction);
+  void performLookup();
+  bool drawHighlightWithSnapshot();
+  void drawHints() const;
+
+  std::unique_ptr<Page> page;
+  const int marginLeft;
+  const int marginTop;
+  int fontId = 0;
+  int lineHeight = 0;
+
+  std::vector<WordBox> words;
+  int selected = 0;
+  uint16_t rowCount = 0;
+
+  Dictionary dict;
+  bool dictOpenAttempted = false;
+  bool dictOpenOk = false;
+
+  Popup popup = Popup::None;
+  StrId popupMsg = StrId::STR_DICT_NOT_FOUND;
+  unsigned long popupTime = 0;
+
+  // Differential highlight repaint: the pixels under the current highlight
+  // box, so a cursor move restores them and repaints only the two affected
+  // boxes instead of re-running the full two-pass page render (which also
+  // reloads every SD-font glyph on the page). snapshotIdx is the word whose
+  // under-pixels are saved; -1 means the framebuffer no longer holds a clean
+  // page (popup drawn, sub-activity shown) and the next render must be full.
+  static constexpr size_t SNAPSHOT_CAPACITY = 4096;
+  std::unique_ptr<uint8_t[]> snapshot;
+  int16_t snapshotX = 0;
+  int16_t snapshotY = 0;
+  int16_t snapshotW = 0;
+  int16_t snapshotH = 0;
+  int snapshotIdx = -1;
+
+  // The activity is entered while Confirm is still held (long-press trigger):
+  // ignore the stale release until a fresh press is seen.
+  bool confirmPressSeen = false;
+};

src/activities/reader/EpubReaderActivity.cpp

@@ -20,6 +20,7 @@
 #include "BookmarkEntry.h"
 #include "CrossPointSettings.h"
 #include "CrossPointState.h"
+#include "DictionaryWordSelectActivity.h"
 #include "EpubReaderBookmarksActivity.h"
 #include "EpubReaderChapterSelectionActivity.h"
 #include "EpubReaderFootnotesActivity.h"
@@ -257,6 +258,29 @@ void EpubReaderActivity::openReaderMenu() {
                          });
 }
 
+void EpubReaderActivity::openDictionaryWordSelect() {
+  if (SETTINGS.dictionaryName[0] == '\0') {
+    showDictionaryMessage = true;
+    dictionaryMessageTime = millis();
+    requestUpdate();
+    return;
+  }
+  if (!section) return;
+  auto page = section->loadPage(section->currentPage);
+  if (!page) return;
+
+  // Word geometry must match render(): viewable-area margins plus screen margin.
+  int orientedMarginTop, orientedMarginRight, orientedMarginBottom, orientedMarginLeft;
+  renderer.getOrientedViewableTRBL(&orientedMarginTop, &orientedMarginRight, &orientedMarginBottom,
+                                   &orientedMarginLeft);
+  orientedMarginTop += SETTINGS.screenMargin;
+  orientedMarginLeft += SETTINGS.screenMargin;
+
+  startActivityForResult(std::make_unique<DictionaryWordSelectActivity>(renderer, mappedInput, std::move(page),
+                                                                        orientedMarginLeft, orientedMarginTop),
+                         [this](const ActivityResult&) { requestUpdate(); });
+}
+
 void EpubReaderActivity::loop() {
   if (!epub) {
     // Should never happen
@@ -379,6 +403,11 @@ void EpubReaderActivity::loop() {
     requestUpdate();
   }
 
+  if (showDictionaryMessage && (millis() - dictionaryMessageTime) >= ReaderUtils::BOOKMARK_MESSAGE_DURATION_MS) {
+    showDictionaryMessage = false;
+    requestUpdate();
+  }
+
   // While the end screen suggestion menu is showing it owns Confirm/Back/navigation
   // input. Anything it doesn't handle (e.g. long-press Back to the file browser) falls
   // through to the regular handlers below; page turns are absorbed by the end-of-book
@@ -442,6 +471,14 @@ void EpubReaderActivity::loop() {
           }
         }
         break;
+      case CrossPointSettings::LP_MENU_DICTIONARY:
+        // Hold ~0.4s starts dictionary word selection on the current page.
+        if (mappedInput.getHeldTime() >= ReaderUtils::BOOKMARK_HOLD_MS && !showDictionaryMessage) {
+          ignoreNextConfirmRelease = true;  // Prevent menu open on the release that follows
+          openDictionaryWordSelect();
+          return;
+        }
+        break;
       case CrossPointSettings::LP_MENU_DISABLED:
       default:
         break;
@@ -710,6 +747,10 @@ void EpubReaderActivity::onReaderMenuConfirm(EpubReaderMenuActivity::MenuAction
           });
       break;
     }
+    case EpubReaderMenuActivity::MenuAction::DICTIONARY: {
+      openDictionaryWordSelect();
+      break;
+    }
     case EpubReaderMenuActivity::MenuAction::DISPLAY_QR: {
       if (section && section->currentPage >= 0 && section->currentPage < section->pageCount) {
         std::string fullText = section->getTextFromSectionFile();
@@ -1300,6 +1341,10 @@ void EpubReaderActivity::render(RenderLock&& lock) {
   if (showBookmarkMessage) {
     GUI.drawPopup(renderer, bookmarkRemoved ? tr(STR_BOOKMARK_REMOVED) : tr(STR_BOOKMARK_ADDED));
   }
+
+  if (showDictionaryMessage) {
+    GUI.drawPopup(renderer, tr(STR_DICT_NO_DICT_SET));
+  }
 }
 
 bool EpubReaderActivity::applyDeferredReposition() {

src/activities/reader/EpubReaderActivity.h

@@ -39,6 +39,9 @@ class EpubReaderActivity final : public Activity {
   bool skipNextButtonCheck = false;  // Skip button processing for one frame after subactivity exit
   bool automaticPageTurnActive = false;
   bool showBookmarkMessage = false;
+  // "No dictionary set" popup, shown when a lookup is triggered without a configured dictionary.
+  bool showDictionaryMessage = false;
+  unsigned long dictionaryMessageTime = 0UL;
   bool ignoreNextConfirmRelease = false;
   bool currentPageBookmarked = false;
   bool bookmarkRemoved = false;  // true when last toggle removed (controls popup text)
@@ -119,6 +122,7 @@ class EpubReaderActivity final : public Activity {
   void onReaderMenuConfirm(EpubReaderMenuActivity::MenuAction action);
   // Opens the reader menu for the current position (short-press Confirm)
   void openReaderMenu();
+  void openDictionaryWordSelect();
   // Returns true if sync acted (launched, or surfaced a save error); false if it was a no-op
   // because no KOReader credentials are stored.
   bool launchKOReaderSync();

src/activities/reader/EpubReaderMenuActivity.cpp

@@ -22,7 +22,7 @@ EpubReaderMenuActivity::EpubReaderMenuActivity(GfxRenderer& renderer, MappedInpu
 std::vector<EpubReaderMenuActivity::MenuItem> EpubReaderMenuActivity::buildMenuItems(bool hasFootnotes,
                                                                                      bool hasBookmarks) {
   std::vector<MenuItem> items;
-  items.reserve(12);
+  items.reserve(13);
   items.push_back({MenuAction::SELECT_CHAPTER, StrId::STR_SELECT_CHAPTER});
   if (hasFootnotes) {
     items.push_back({MenuAction::FOOTNOTES, StrId::STR_FOOTNOTES});
@@ -31,6 +31,7 @@ std::vector<EpubReaderMenuActivity::MenuItem> EpubReaderMenuActivity::buildMenuI
     items.push_back({MenuAction::BOOKMARKS, StrId::STR_BOOKMARKS});
   }
   items.push_back({MenuAction::TOGGLE_BOOKMARK, StrId::STR_TOGGLE_BOOKMARK});
+  items.push_back({MenuAction::DICTIONARY, StrId::STR_LOOKUP});
   items.push_back({MenuAction::ROTATE_SCREEN, StrId::STR_ORIENTATION});
   items.push_back({MenuAction::AUTO_PAGE_TURN, StrId::STR_AUTO_TURN_PAGES_PER_MIN});
   items.push_back({MenuAction::GO_TO_PERCENT, StrId::STR_GO_TO_PERCENT});

src/activities/reader/EpubReaderMenuActivity.h

@@ -24,7 +24,8 @@ class EpubReaderMenuActivity final : public Activity {
     DISPLAY_QR,
     GO_HOME,
     SYNC,
-    DELETE_CACHE
+    DELETE_CACHE,
+    DICTIONARY
   };
 
   explicit EpubReaderMenuActivity(GfxRenderer& renderer, MappedInputManager& mappedInput, const std::string& title,

src/activities/settings/SettingsActivity.cpp

@@ -39,7 +39,12 @@ void SettingsActivity::rebuildSettingsLists() {
   // reader activity ran — otherwise the font-family picker shows stale list.
   sdFontSystem.refreshIfDirty();
 
-  for (auto& setting : getSettingsList(&sdFontSystem.registry())) {
+  // Rescan /dictionaries on every rebuild: cheap (one directory listing) and
+  // picks up dictionaries copied to the SD card since the last visit.
+  std::vector<DictionaryEntry> dictionaries;
+  DictionaryRegistry::discover(dictionaries);
+
+  for (auto& setting : getSettingsList(&sdFontSystem.registry(), &dictionaries)) {
     if (setting.category == StrId::STR_NONE_OPT) continue;
     if (setting.category == StrId::STR_CAT_DISPLAY) {
       displaySettings.push_back(setting);

src/util/DictZip.cpp

@@ -0,0 +1,180 @@
+#include "DictZip.h"
+
+#include <InflateReader.h>
+#include <Memory.h>
+
+namespace DictZip {
+namespace {
+
+// Caps the chunk table at 32KB of heap (8192 * 4 bytes); at the typical ~58KB
+// chunk length that still allows ~460MB of uncompressed dictionary data.
+constexpr uint16_t MAX_CHUNK_COUNT = 8192;
+
+bool readLe16(HalFile& file, uint16_t* out) {
+  uint8_t raw[2];
+  if (file.read(raw, 2) != 2) return false;
+  *out = static_cast<uint16_t>(raw[0] | (static_cast<uint16_t>(raw[1]) << 8));
+  return true;
+}
+
+bool extractChunkSlice(HalFile& file, uint32_t compressedOffset, uint32_t compressedSize, uint32_t discardSize,
+                       uint32_t extractSize, HalFile& outFile) {
+  if (extractSize == 0) return true;
+  auto compBuf = makeUniqueNoThrow<uint8_t[]>(compressedSize);
+  if (!compBuf) return false;
+
+  file.seekSet(compressedOffset);
+  if (file.read(compBuf.get(), static_cast<int>(compressedSize)) != static_cast<int>(compressedSize)) return false;
+
+  InflateReader reader;
+  if (!reader.init(true)) return false;
+  reader.setSource(compBuf.get(), compressedSize);
+
+  auto buf = makeUniqueNoThrow<uint8_t[]>(512);
+  if (!buf) return false;
+
+  uint32_t batch;
+  while (discardSize > 0) {
+    batch = discardSize < 512 ? discardSize : 512;
+    if (!reader.read(buf.get(), batch)) return false;
+    discardSize -= batch;
+  }
+
+  while (extractSize > 0) {
+    batch = extractSize < 512 ? extractSize : 512;
+    if (!reader.read(buf.get(), batch)) return false;
+    if (outFile.write(buf.get(), batch) != batch) return false;
+    extractSize -= batch;
+  }
+
+  return true;
+}
+
+}  // namespace
+
+bool parse(HalFile& file, Info* info) {
+  if (!info) return false;
+  *info = {};
+
+  uint8_t header[10];
+  if (file.read(header, sizeof(header)) != static_cast<int>(sizeof(header))) return false;
+  if (header[0] != 0x1f || header[1] != 0x8b || header[2] != 8) return false;
+
+  const uint8_t flags = header[3];
+  if ((flags & 0x04) == 0) return false;  // dictzip requires FEXTRA
+
+  uint16_t xlen = 0;
+  if (!readLe16(file, &xlen)) return false;
+
+  uint32_t extraRead = 0;
+  bool foundRa = false;
+  while (extraRead + 4 <= xlen) {
+    uint8_t subHeader[4];
+    if (file.read(subHeader, sizeof(subHeader)) != static_cast<int>(sizeof(subHeader))) return false;
+    extraRead += 4;
+    const uint16_t subLen = static_cast<uint16_t>(subHeader[2] | (static_cast<uint16_t>(subHeader[3]) << 8));
+    if (extraRead + subLen > xlen) return false;
+
+    if (subHeader[0] == 'R' && subHeader[1] == 'A') {
+      if (subLen < 6) return false;
+
+      uint16_t version = 0;
+      uint16_t chunkLen = 0;
+      uint16_t chunkCount = 0;
+      if (!readLe16(file, &version) || !readLe16(file, &chunkLen) || !readLe16(file, &chunkCount)) return false;
+      extraRead += 6;
+      if (version != 1 || chunkLen == 0 || chunkCount == 0 || chunkCount > MAX_CHUNK_COUNT) return false;
+      if (subLen != static_cast<uint16_t>(6 + chunkCount * 2)) return false;
+
+      info->chunkLength = chunkLen;
+      info->chunkOffsets.reserve(static_cast<size_t>(chunkCount) + 1);
+      info->chunkOffsets.push_back(0);
+      uint32_t cumulative = 0;
+      for (uint16_t i = 0; i < chunkCount; i++) {
+        uint16_t compLen = 0;
+        if (!readLe16(file, &compLen)) return false;
+        extraRead += 2;
+        cumulative += compLen;
+        info->chunkOffsets.push_back(cumulative);
+      }
+      foundRa = true;
+    } else {
+      file.seekSet(file.position() + subLen);
+      extraRead += subLen;
+    }
+  }
+  if (extraRead != xlen || !foundRa) return false;
+
+  if (flags & 0x08) {  // FNAME
+    int b;
+    do {
+      b = file.read();
+      if (b < 0) return false;
+    } while (b != 0);
+  }
+  if (flags & 0x10) {  // FCOMMENT
+    int b;
+    do {
+      b = file.read();
+      if (b < 0) return false;
+    } while (b != 0);
+  }
+  if (flags & 0x02) {  // FHCRC
+    uint8_t crc[2];
+    if (file.read(crc, 2) != 2) return false;
+  }
+
+  info->dataOffset = static_cast<uint32_t>(file.position());
+  const uint32_t fileSize = static_cast<uint32_t>(file.fileSize());
+  if (fileSize < 4) return false;
+  file.seekSet(fileSize - 4);
+  uint8_t isizeRaw[4];
+  if (file.read(isizeRaw, 4) != 4) return false;
+  info->totalSize = static_cast<uint32_t>(isizeRaw[0]) | (static_cast<uint32_t>(isizeRaw[1]) << 8) |
+                    (static_cast<uint32_t>(isizeRaw[2]) << 16) | (static_cast<uint32_t>(isizeRaw[3]) << 24);
+  if (info->totalSize == 0) return false;
+  info->valid = true;
+  return true;
+}
+
+bool extractEntry(const char* path, uint32_t offset, uint32_t size, HalFile& outFile) {
+  if (size == 0) return true;
+
+  HalFile file;
+  if (!Storage.openFileForRead("DICTZIP", path, file)) return false;
+
+  Info info;
+  if (!parse(file, &info)) return false;
+
+  // Reject ranges outside the uncompressed data (offset/size come from the
+  // untrusted .idx). Subtraction form avoids uint32 overflow in offset + size
+  // and guarantees localOffset < chunkOutSize in the loop below.
+  if (offset > info.totalSize || size > info.totalSize - offset) return false;
+
+  const uint32_t startChunk = offset / info.chunkLength;
+  const uint32_t endChunk = (offset + size - 1) / info.chunkLength;
+  if (endChunk + 1 >= info.chunkOffsets.size()) return false;
+
+  uint32_t remaining = size;
+  const uint32_t lastChunk = static_cast<uint32_t>(info.chunkOffsets.size() - 2);
+  for (uint32_t chunk = startChunk; chunk <= endChunk; chunk++) {
+    uint32_t chunkOutSize = info.chunkLength;
+    if (chunk == lastChunk) chunkOutSize = info.totalSize - chunk * info.chunkLength;
+    if (chunkOutSize == 0 || chunkOutSize > info.chunkLength) chunkOutSize = info.chunkLength;
+
+    const uint32_t localOffset = (chunk == startChunk) ? (offset % info.chunkLength) : 0;
+    const uint32_t available = chunkOutSize - localOffset;
+    const uint32_t take = remaining < available ? remaining : available;
+
+    const uint32_t compOffset = info.dataOffset + info.chunkOffsets[chunk];
+    const uint32_t compSize = info.chunkOffsets[chunk + 1] - info.chunkOffsets[chunk];
+    if (!extractChunkSlice(file, compOffset, compSize, localOffset, take, outFile)) return false;
+
+    remaining -= take;
+    if (remaining == 0) break;
+  }
+
+  return remaining == 0;
+}
+
+}  // namespace DictZip

src/util/DictZip.h

@@ -0,0 +1,26 @@
+#pragma once
+
+#include <HalStorage.h>
+
+#include <cstdint>
+#include <vector>
+
+// Random-access reader for dictzip (.dict.dz) files: gzip with an extra "RA"
+// field holding a chunk table, so any byte range can be decompressed without
+// inflating the whole file. Format: https://linux.die.net/man/1/dictzip
+namespace DictZip {
+
+struct Info {
+  uint32_t dataOffset = 0;             // file offset where compressed chunk data starts
+  uint32_t totalSize = 0;              // uncompressed size (gzip ISIZE trailer)
+  uint16_t chunkLength = 0;            // uncompressed bytes per chunk
+  std::vector<uint32_t> chunkOffsets;  // cumulative compressed offsets, chunkCount+1 entries
+  bool valid = false;
+};
+
+bool parse(HalFile& file, Info* info);
+
+// Decompress the uncompressed byte range [offset, offset+size) into outFile.
+bool extractEntry(const char* path, uint32_t offset, uint32_t size, HalFile& outFile);
+
+}  // namespace DictZip

src/util/Dictionary.cpp

@@ -0,0 +1,378 @@
+#include "Dictionary.h"
+
+#include <Arduino.h>
+#include <Logging.h>
+#include <Memory.h>
+
+#include <algorithm>
+#include <cctype>
+#include <cstring>
+
+#include "DictZip.h"
+#include "DictionaryRegistry.h"
+#include "StringUtils.h"
+
+namespace {
+
+// Shared temp file for entries lazily extracted from .dict.dz.
+constexpr const char* DICT_TMP_FILE = "/.crosspoint/dict.tmp";
+
+// .qidx sidecar header: magic, version, sample interval, sample count, and the
+// .idx file size the sidecar was built from (staleness check).
+constexpr uint32_t QIDX_MAGIC = 0x58444951;  // "QIDX" little-endian
+constexpr uint32_t QIDX_VERSION = 1;
+constexpr size_t QIDX_HEADER_BYTES = 5 * sizeof(uint32_t);
+
+struct QidxHeader {
+  uint32_t sampleCount = 0;
+  uint32_t idxFileSize = 0;
+  bool valid = false;
+};
+
+QidxHeader readQidxHeader(HalFile& qidx, uint32_t sampleInterval) {
+  QidxHeader header;
+  uint32_t raw[5];
+  if (!qidx.seekSet(0) || qidx.read(raw, sizeof(raw)) != static_cast<int>(sizeof(raw))) return header;
+  if (raw[0] != QIDX_MAGIC || raw[1] != QIDX_VERSION || raw[2] != sampleInterval) return header;
+  header.sampleCount = raw[3];
+  header.idxFileSize = raw[4];
+  header.valid = true;
+  return header;
+}
+
+bool readSampleOffset(HalFile& qidx, uint32_t sampleIndex, uint32_t* out) {
+  if (!qidx.seekSet(QIDX_HEADER_BYTES + static_cast<size_t>(sampleIndex) * sizeof(uint32_t))) return false;
+  return qidx.read(out, sizeof(*out)) == static_cast<int>(sizeof(*out));
+}
+
+uint32_t readBe32(const uint8_t* p) {
+  return (static_cast<uint32_t>(p[0]) << 24) | (static_cast<uint32_t>(p[1]) << 16) |
+         (static_cast<uint32_t>(p[2]) << 8) | static_cast<uint32_t>(p[3]);
+}
+
+// Word characters for cleaning: ASCII alphanumerics plus any UTF-8
+// continuation/lead byte, so accented words keep their edges.
+bool isWordByte(unsigned char c) { return c >= 0x80 || std::isalnum(c) != 0; }
+
+// True when the .ifo declares 64-bit index offsets, which this reader does not
+// support (only scans the first 2KB — idxoffsetbits always appears early).
+bool ifoDeclares64BitOffsets(const std::string& ifoPath) {
+  HalFile ifo;
+  if (!Storage.openFileForRead("DICT", ifoPath, ifo)) return false;
+  char buf[2048];
+  const int n = ifo.read(buf, sizeof(buf) - 1);
+  if (n <= 0) return false;
+  buf[n] = '\0';
+  const char* line = strstr(buf, "idxoffsetbits");
+  if (!line) return false;
+  const char* eq = strchr(line, '=');
+  return eq && strtol(eq + 1, nullptr, 10) == 64;
+}
+
+}  // namespace
+
+bool Dictionary::open(const char* folderName) {
+  basePath.clear();
+  std::string resolved;
+  if (!DictionaryRegistry::resolveBasePath(folderName, resolved)) {
+    LOG_ERR("DICT", "No dictionary found in folder '%s'", folderName ? folderName : "");
+    return false;
+  }
+
+  if (!Storage.exists((resolved + ".idx").c_str())) {
+    LOG_ERR("DICT", "%s.idx missing (compressed .idx.gz is not supported)", resolved.c_str());
+    return false;
+  }
+  hasPlainDict = Storage.exists((resolved + ".dict").c_str());
+  if (!hasPlainDict && !Storage.exists((resolved + ".dict.dz").c_str())) {
+    LOG_ERR("DICT", "%s has no .dict or .dict.dz", resolved.c_str());
+    return false;
+  }
+  if (ifoDeclares64BitOffsets(resolved + ".ifo")) {
+    LOG_ERR("DICT", "%s uses 64-bit index offsets (unsupported)", resolved.c_str());
+    return false;
+  }
+
+  basePath = std::move(resolved);
+  return true;
+}
+
+bool Dictionary::needsIndex() {
+  if (!isOpen()) return false;
+
+  HalFile idx;
+  if (!Storage.openFileForRead("DICT", basePath + ".idx", idx)) return false;
+  const uint32_t idxSize = static_cast<uint32_t>(idx.fileSize());
+
+  HalFile qidx;
+  if (!Storage.openFileForRead("DICT", basePath + ".qidx", qidx)) return true;
+  const QidxHeader header = readQidxHeader(qidx, SAMPLE_INTERVAL);
+  return !header.valid || header.idxFileSize != idxSize;
+}
+
+bool Dictionary::buildIndex(void (*yieldFn)(void*), void* ctx) {
+  if (!isOpen()) return false;
+
+  HalFile idx;
+  if (!Storage.openFileForRead("DICT", basePath + ".idx", idx)) return false;
+  const uint32_t idxSize = static_cast<uint32_t>(idx.fileSize());
+
+  constexpr size_t CHUNK_BYTES = 4096;
+  auto buf = makeUniqueNoThrow<uint8_t[]>(CHUNK_BYTES);
+  if (!buf) {
+    LOG_ERR("DICT", "OOM: %u byte index scan buffer", CHUNK_BYTES);
+    return false;
+  }
+
+  // Stream each sample offset straight to the sidecar instead of accumulating
+  // them in RAM: a large .idx would otherwise cost tens of KB of vector heap,
+  // and vector growth aborts on OOM under -fno-exceptions. The header slot is
+  // zero-filled until the scan succeeds, so an interrupted build leaves a file
+  // readQidxHeader rejects (magic mismatch) and needsIndex() triggers a rebuild.
+  const std::string qidxPath = basePath + ".qidx";
+  HalFile out;
+  if (!Storage.openFileForWrite("DICT", qidxPath, out)) return false;
+  const auto writeU32 = [&out](uint32_t v) { return out.write(&v, sizeof(v)) == static_cast<int>(sizeof(v)); };
+  const uint32_t placeholder[5] = {};
+  bool ok = out.write(placeholder, sizeof(placeholder)) == sizeof(placeholder);
+  uint32_t sampleCount = 0;
+  if (ok) {
+    ok = writeU32(0);  // entry 0 always starts at byte 0
+    sampleCount = 1;
+  }
+
+  const unsigned long startMs = millis();
+  uint32_t entryCount = 0;
+  uint32_t pos = 0;
+  uint32_t suffixLeft = 0;  // 0 while scanning a headword, else suffix bytes remaining
+  uint32_t sinceYield = 0;
+  while (ok && pos < idxSize) {
+    const int n = idx.read(buf.get(), CHUNK_BYTES);
+    if (n <= 0) {
+      LOG_ERR("DICT", "Index scan read failed at %lu", static_cast<unsigned long>(pos));
+      ok = false;
+      break;
+    }
+    for (int i = 0; ok && i < n; i++) {
+      if (suffixLeft == 0) {
+        if (buf[i] == 0) suffixLeft = 8;
+      } else if (--suffixLeft == 0) {
+        entryCount++;
+        const uint32_t nextEntryStart = pos + i + 1;
+        if (entryCount % SAMPLE_INTERVAL == 0 && nextEntryStart < idxSize) {
+          ok = writeU32(nextEntryStart);
+          sampleCount++;
+        }
+      }
+    }
+    pos += n;
+    sinceYield += n;
+    if (yieldFn && sinceYield >= 64 * 1024) {
+      sinceYield = 0;
+      yieldFn(ctx);
+    }
+  }
+
+  if (ok) {
+    // Backpatch the now-valid header over the placeholder.
+    const uint32_t header[5] = {QIDX_MAGIC, QIDX_VERSION, SAMPLE_INTERVAL, sampleCount, idxSize};
+    ok = out.seekSet(0) && out.write(header, sizeof(header)) == sizeof(header);
+  }
+  if (!ok) {
+    LOG_ERR("DICT", "Index build failed, removing %s", qidxPath.c_str());
+    out.close();  // close before remove of the same path
+    Storage.remove(qidxPath.c_str());
+    return false;
+  }
+
+  LOG_INF("DICT", "Indexed %lu entries (%lu samples) in %lu ms", static_cast<unsigned long>(entryCount),
+          static_cast<unsigned long>(sampleCount), millis() - startMs);
+  return true;
+}
+
+int Dictionary::readWordInto(HalFile& file, char* buf, size_t bufSize) {
+  size_t i = 0;
+  while (i < bufSize - 1) {
+    const int ch = file.read();
+    if (ch < 0) return -1;  // EOF or I/O error
+    if (ch == 0) {
+      buf[i] = '\0';
+      return static_cast<int>(i);
+    }
+    buf[i++] = static_cast<char>(ch);
+  }
+  // Word too long for buffer — consume remaining bytes to stay in sync
+  buf[bufSize - 1] = '\0';
+  int ch;
+  do {
+    ch = file.read();
+  } while (ch > 0);
+  return static_cast<int>(bufSize - 1);
+}
+
+DictLocation Dictionary::locate(const char* target, std::string* matchedHeadwordOut) {
+  DictLocation result;
+  if (!isOpen()) return result;
+
+  HalFile idx;
+  if (!Storage.openFileForRead("DICT", basePath + ".idx", idx)) return result;
+  const uint32_t idxSize = static_cast<uint32_t>(idx.fileSize());
+
+  // Bisect the sampled offsets to the last sample whose headword <= target.
+  // Falls back to a full scan from byte 0 when the sidecar is unusable.
+  uint32_t startByte = 0;
+  HalFile qidx;
+  if (Storage.openFileForRead("DICT", basePath + ".qidx", qidx)) {
+    const QidxHeader header = readQidxHeader(qidx, SAMPLE_INTERVAL);
+    if (header.valid && header.idxFileSize == idxSize && header.sampleCount > 0) {
+      uint32_t lo = 0;
+      uint32_t hi = header.sampleCount - 1;
+      while (lo < hi) {
+        const uint32_t mid = (lo + hi + 1) / 2;
+        uint32_t offset = 0;
+        if (!readSampleOffset(qidx, mid, &offset) || !idx.seekSet(offset) ||
+            readWordInto(idx, wordBuf, sizeof(wordBuf)) < 0) {
+          lo = 0;
+          break;
+        }
+        if (StringUtils::asciiCaseCmp(wordBuf, target) <= 0) {
+          lo = mid;
+        } else {
+          hi = mid - 1;
+        }
+      }
+      readSampleOffset(qidx, lo, &startByte);
+    }
+  }
+
+  // Linear scan of at most SAMPLE_INTERVAL entries: headword NUL, BE32 offset,
+  // BE32 size. The index is sorted, so stop at the first headword > target.
+  idx.seekSet(startByte);
+  while (static_cast<uint32_t>(idx.position()) < idxSize) {
+    if (readWordInto(idx, wordBuf, sizeof(wordBuf)) < 0) break;
+    uint8_t suffix[8];
+    if (idx.read(suffix, 8) != 8) break;
+
+    const int cmp = StringUtils::asciiCaseCmp(wordBuf, target);
+    if (cmp == 0) {
+      result.offset = readBe32(suffix);
+      result.size = readBe32(suffix + 4);
+      result.found = true;
+      if (matchedHeadwordOut) *matchedHeadwordOut = wordBuf;
+      return result;
+    }
+    if (cmp > 0) break;
+  }
+  return result;
+}
+
+bool Dictionary::readDefinition(const DictLocation& location, std::string& out) {
+  if (!location.found) return false;
+  const uint32_t size = std::min(location.size, MAX_DEFINITION_BYTES);
+
+  std::string path;
+  uint32_t offset = 0;
+  if (hasPlainDict) {
+    path = basePath + ".dict";
+    offset = location.offset;
+  } else {
+    HalFile tmp = Storage.open(DICT_TMP_FILE, O_WRITE | O_CREAT | O_TRUNC);
+    if (!tmp) {
+      LOG_ERR("DICT", "Failed to open %s", DICT_TMP_FILE);
+      return false;
+    }
+    if (!DictZip::extractEntry((basePath + ".dict.dz").c_str(), location.offset, size, tmp)) {
+      LOG_ERR("DICT", "dictzip extraction failed for %s", basePath.c_str());
+      return false;
+    }
+    tmp.close();  // close before reopening the same path for read
+    path = DICT_TMP_FILE;
+  }
+
+  HalFile dict;
+  if (!Storage.openFileForRead("DICT", path, dict)) return false;
+  const uint32_t dictSize = static_cast<uint32_t>(dict.fileSize());
+  if (offset > dictSize || size > dictSize - offset) {
+    LOG_ERR("DICT", "Definition out of bounds (%lu+%lu > %lu)", static_cast<unsigned long>(offset),
+            static_cast<unsigned long>(size), static_cast<unsigned long>(dictSize));
+    return false;
+  }
+
+  // std::string growth aborts on OOM (-fno-exceptions); refuse up front unless
+  // the allocation fits comfortably in the largest free block.
+  if (ESP.getMaxAllocHeap() < size + 8 * 1024) {
+    LOG_ERR("DICT", "Low heap for %lu byte definition", static_cast<unsigned long>(size));
+    return false;
+  }
+
+  dict.seekSet(offset);
+  out.assign(size, '\0');
+  const int bytesRead = dict.read(&out[0], size);
+  if (bytesRead < 0) {
+    out.clear();
+    return false;
+  }
+  if (static_cast<uint32_t>(bytesRead) < size) out.resize(bytesRead);
+  return true;
+}
+
+std::string Dictionary::cleanWord(const char* word) {
+  if (!word) return "";
+  size_t start = 0;
+  size_t end = strlen(word);
+  while (start < end && !isWordByte(static_cast<unsigned char>(word[start]))) start++;
+  while (end > start && !isWordByte(static_cast<unsigned char>(word[end - 1]))) end--;
+  if (start >= end) return "";
+
+  std::string result(word + start, end - start);
+  std::transform(result.begin(), result.end(), result.begin(),
+                 [](unsigned char c) { return c >= 0x80 ? c : static_cast<unsigned char>(std::tolower(c)); });
+  return result;
+}
+
+void Dictionary::stemVariants(const std::string& word, std::vector<std::string>& out) {
+  out.clear();
+  out.reserve(6);
+  const size_t n = word.size();
+  const auto add = [&out](std::string v) {
+    if (std::find(out.begin(), out.end(), v) == out.end()) out.push_back(std::move(v));
+  };
+  // endsWith requires a non-empty remainder so variants never come out empty.
+  const auto endsWith = [&word, n](const char* suffix) {
+    const size_t len = strlen(suffix);
+    return n > len && word.compare(n - len, len, suffix) == 0;
+  };
+
+  if (endsWith("'s")) add(word.substr(0, n - 2));
+  if (endsWith("\xE2\x80\x99s")) add(word.substr(0, n - 4));  // U+2019 apostrophe
+  if (endsWith("ies")) add(word.substr(0, n - 3) + "y");      // stories -> story
+  if (endsWith("es")) add(word.substr(0, n - 2));             // boxes -> box
+  if (endsWith("s")) add(word.substr(0, n - 1));              // dogs -> dog
+  if (endsWith("ed")) {
+    add(word.substr(0, n - 2));                                            // walked -> walk
+    add(word.substr(0, n - 1));                                            // loved -> love
+    if (n >= 4 && word[n - 3] == word[n - 4]) add(word.substr(0, n - 3));  // stopped -> stop
+  }
+  if (endsWith("ing")) {
+    add(word.substr(0, n - 3));                                            // walking -> walk
+    add(word.substr(0, n - 3) + "e");                                      // making -> make
+    if (n >= 5 && word[n - 4] == word[n - 5]) add(word.substr(0, n - 4));  // running -> run
+  }
+}
+
+bool Dictionary::lookup(const char* word, std::string& definitionOut, std::string& matchedHeadwordOut) {
+  const std::string cleaned = cleanWord(word);
+  if (cleaned.empty() || !isOpen()) return false;
+
+  DictLocation location = locate(cleaned.c_str(), &matchedHeadwordOut);
+  if (!location.found) {
+    std::vector<std::string> variants;
+    stemVariants(cleaned, variants);
+    for (const auto& variant : variants) {
+      location = locate(variant.c_str(), &matchedHeadwordOut);
+      if (location.found) break;
+    }
+  }
+  if (!location.found) return false;
+  return readDefinition(location, definitionOut);
+}

src/util/Dictionary.h

@@ -0,0 +1,65 @@
+#pragma once
+
+#include <HalStorage.h>
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+// Result of an index search — file location of a definition without reading it.
+struct DictLocation {
+  uint32_t offset = 0;  // byte offset in .dict data
+  uint32_t size = 0;    // byte length in .dict data
+  bool found = false;
+};
+
+// Slim StarDict reader: exact-match lookup with a mini stemming fallback.
+//
+// Expects /dictionaries/<folder>/<stem>.idx (uncompressed) plus <stem>.dict or
+// <stem>.dict.dz. Lookups binary-search a lazily built sampled-offset sidecar
+// (<stem>.qidx, byte offset of every SAMPLE_INTERVAL-th .idx entry), then
+// linear-scan at most SAMPLE_INTERVAL entries. Everything streams from SD; no
+// index is held in RAM.
+class Dictionary {
+ public:
+  // Resolve the dictionary folder and validate its files. Rejects
+  // dictionaries with 64-bit index offsets (idxoffsetbits=64 in .ifo).
+  bool open(const char* folderName);
+  bool isOpen() const { return !basePath.empty(); }
+
+  // True when the .qidx sidecar is missing or stale — call buildIndex() first
+  // so the UI can show an "Indexing…" message for the slow first pass.
+  bool needsIndex();
+
+  // One streaming pass over .idx writing the .qidx sidecar. yieldFn (optional)
+  // is called every ~64KB consumed to feed the watchdog / repaint the UI.
+  bool buildIndex(void (*yieldFn)(void*) = nullptr, void* ctx = nullptr);
+
+  // Clean the word, look it up, and on a miss retry mini stem variants
+  // (-'s/-s/-es/-ies/-ed/-ing). On a hit fills the definition text (capped at
+  // MAX_DEFINITION_BYTES) and the headword as stored in the index.
+  bool lookup(const char* word, std::string& definitionOut, std::string& matchedHeadwordOut);
+
+  static std::string cleanWord(const char* word);
+
+  static constexpr uint32_t MAX_DEFINITION_BYTES = 64 * 1024;
+
+ private:
+  static constexpr uint32_t SAMPLE_INTERVAL = 256;
+
+  DictLocation locate(const char* target, std::string* matchedHeadwordOut);
+  bool readDefinition(const DictLocation& location, std::string& out);
+  static void stemVariants(const std::string& word, std::vector<std::string>& out);
+
+  // Read a null-terminated word from an open file into buf (max bufSize-1
+  // chars). Returns the number of characters read (excluding null), or -1 on
+  // EOF/error. Over-long words are truncated but the stream stays in sync.
+  static int readWordInto(HalFile& file, char* buf, size_t bufSize);
+
+  std::string basePath;  // "/dictionaries/<folder>/<stem>", empty when not open
+  bool hasPlainDict = false;
+
+  // Shared scan buffer: lookups are single-threaded and this avoids a
+  // 256-byte array on the stack of every locate() call.
+  char wordBuf[256] = {};
+};

src/util/DictionaryRegistry.cpp

@@ -0,0 +1,113 @@
+#include "DictionaryRegistry.h"
+
+#include <HalStorage.h>
+#include <Logging.h>
+
+#include <algorithm>
+#include <cstring>
+
+#include "StringUtils.h"
+
+namespace DictionaryRegistry {
+namespace {
+
+// Dictionaries are looked up in both roots, in order. The hidden variant
+// lets users keep the folder out of the file browser (hidden by default,
+// see FileBrowserActivity's showHiddenFiles check).
+constexpr const char* DICT_ROOTS[] = {"/dictionaries", "/.dictionaries"};
+
+// Find the single .idx stem inside one dictionary folder. Returns false when
+// the folder holds no .idx or more than one distinct stem (ambiguous).
+bool findStem(const char* folderPath, std::string& stemOut) {
+  auto dir = Storage.open(folderPath);
+  if (!dir || !dir.isDirectory()) return false;
+
+  dir.rewindDirectory();
+  char name[128];
+  char foundStem[128];
+  foundStem[0] = '\0';
+  for (auto entry = dir.openNextFile(); entry; entry = dir.openNextFile()) {
+    entry.getName(name, sizeof(name));
+    // Skip macOS metadata files (AppleDouble resource forks)
+    if (entry.isDirectory() || strncmp(name, "._", 2) == 0) continue;
+
+    const size_t len = strlen(name);
+    if (len <= 4 || strcmp(name + len - 4, ".idx") != 0) continue;
+
+    name[len - 4] = '\0';
+    if (foundStem[0] != '\0' && strcmp(foundStem, name) != 0) {
+      LOG_DBG("DREG", "Skipping %s: multiple index stems found", folderPath);
+      return false;
+    }
+    strncpy(foundStem, name, sizeof(foundStem) - 1);
+    foundStem[sizeof(foundStem) - 1] = '\0';
+  }
+
+  if (foundStem[0] == '\0') return false;
+
+  // Require dictionary data next to the index, so folders holding only an
+  // .idx never surface as selectable dictionaries that fail at lookup time.
+  const std::string base = std::string(folderPath) + "/" + foundStem;
+  if (!Storage.exists((base + ".dict").c_str()) && !Storage.exists((base + ".dict.dz").c_str())) {
+    LOG_DBG("DREG", "Skipping %s: no .dict or .dict.dz", folderPath);
+    return false;
+  }
+
+  stemOut = foundStem;
+  return true;
+}
+
+}  // namespace
+
+void discover(std::vector<DictionaryEntry>& out) {
+  out.clear();
+  out.reserve(8);
+
+  for (const char* dictRoot : DICT_ROOTS) {
+    auto rootDir = Storage.open(dictRoot);
+    if (!rootDir || !rootDir.isDirectory()) {
+      LOG_DBG("DREG", "No %s directory on SD card", dictRoot);
+      continue;
+    }
+
+    rootDir.rewindDirectory();
+    char name[128];
+    for (auto entry = rootDir.openNextFile(); entry; entry = rootDir.openNextFile()) {
+      entry.getName(name, sizeof(name));
+      if (!entry.isDirectory() || name[0] == '.') continue;
+
+      std::string folderPath = std::string(dictRoot) + "/" + name;
+      std::string stem;
+      if (!findStem(folderPath.c_str(), stem)) continue;
+
+      DictionaryEntry e;
+      e.name = name;
+      e.stem = std::move(stem);
+      out.push_back(std::move(e));
+      LOG_DBG("DREG", "Found dictionary: %s", name);
+    }
+  }
+
+  // Case-insensitive sort by folder name (matches FileBrowserActivity ordering).
+  std::sort(out.begin(), out.end(), [](const DictionaryEntry& a, const DictionaryEntry& b) {
+    return StringUtils::asciiCaseCmp(a.name.c_str(), b.name.c_str()) < 0;
+  });
+}
+
+bool resolveBasePath(const char* folderName, std::string& basePathOut) {
+  if (!folderName || folderName[0] == '\0') return false;
+  // folderName is persisted in the settings JSON: reject separators and dot
+  // prefixes so a crafted value cannot escape the dictionary roots.
+  if (folderName[0] == '.' || strpbrk(folderName, "/\\") != nullptr) return false;
+
+  for (const char* dictRoot : DICT_ROOTS) {
+    std::string folderPath = std::string(dictRoot) + "/" + folderName;
+    std::string stem;
+    if (!findStem(folderPath.c_str(), stem)) continue;
+    basePathOut = folderPath + "/" + stem;
+    return true;
+  }
+  return false;
+}
+
+}  // namespace DictionaryRegistry

src/util/DictionaryRegistry.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#include <string>
+#include <vector>
+
+// One StarDict dictionary found under /dictionaries or /.dictionaries: a
+// subfolder holding <stem>.idx plus <stem>.dict or <stem>.dict.dz.
+struct DictionaryEntry {
+  std::string name;  // subfolder name (shown to the user, stored in settings)
+  std::string stem;  // index basename without .idx
+};
+
+namespace DictionaryRegistry {
+
+// Scan /dictionaries/*/ and /.dictionaries/*/ for dictionaries. Folders with
+// multiple index stems are ambiguous and skipped. Result is sorted
+// case-insensitively by name.
+void discover(std::vector<DictionaryEntry>& out);
+
+// Resolve a folder name to its extensionless base path
+// ("/dictionaries/<folder>/<stem>" or "/.dictionaries/<folder>/<stem>").
+// Returns false if the folder holds no usable dictionary in either root.
+bool resolveBasePath(const char* folderName, std::string& basePathOut);
+
+}  // namespace DictionaryRegistry

src/util/HtmlToPlainText.cpp

@@ -0,0 +1,150 @@
+#include "HtmlToPlainText.h"
+
+#include <Epub/htmlEntities.h>
+
+#include <cctype>
+#include <cstdint>
+#include <cstring>
+
+namespace {
+
+bool isTagStart(const std::string& input, size_t pos) {
+  if (pos + 1 >= input.size()) return false;
+  const unsigned char next = input[pos + 1];
+  return next == '/' || next == '!' || next == '?' || std::isalpha(next);
+}
+
+enum class TagBreak : uint8_t { None, Line, Paragraph };
+
+TagBreak tagBreak(const std::string& input, size_t start, size_t end) {
+  while (start < end && (input[start] == '/' || std::isspace(static_cast<unsigned char>(input[start])))) start++;
+  const size_t nameStart = start;
+  while (start < end && std::isalpha(static_cast<unsigned char>(input[start]))) start++;
+  const size_t len = start - nameStart;
+  if (len == 0) return TagBreak::None;
+
+  const auto equals = [&](const char* name) {
+    const size_t nameLen = strlen(name);
+    if (len != nameLen) return false;
+    for (size_t i = 0; i < len; i++) {
+      if (std::tolower(static_cast<unsigned char>(input[nameStart + i])) != name[i]) return false;
+    }
+    return true;
+  };
+
+  if (equals("p") || equals("h1") || equals("h2") || equals("h3") || equals("h4") || equals("h5") || equals("h6") ||
+      equals("hr")) {
+    return TagBreak::Paragraph;
+  }
+  if (equals("br") || equals("div") || equals("li") || equals("tr")) return TagBreak::Line;
+  return TagBreak::None;
+}
+
+void appendCodepoint(std::string& output, uint32_t codepoint) {
+  if (codepoint == 0 || codepoint > 0x10FFFF || (codepoint >= 0xD800 && codepoint <= 0xDFFF)) return;
+  if (codepoint <= 0x7F) {
+    output.push_back(static_cast<char>(codepoint));
+  } else if (codepoint <= 0x7FF) {
+    output.push_back(static_cast<char>(0xC0 | (codepoint >> 6)));
+    output.push_back(static_cast<char>(0x80 | (codepoint & 0x3F)));
+  } else if (codepoint <= 0xFFFF) {
+    output.push_back(static_cast<char>(0xE0 | (codepoint >> 12)));
+    output.push_back(static_cast<char>(0x80 | ((codepoint >> 6) & 0x3F)));
+    output.push_back(static_cast<char>(0x80 | (codepoint & 0x3F)));
+  } else {
+    output.push_back(static_cast<char>(0xF0 | (codepoint >> 18)));
+    output.push_back(static_cast<char>(0x80 | ((codepoint >> 12) & 0x3F)));
+    output.push_back(static_cast<char>(0x80 | ((codepoint >> 6) & 0x3F)));
+    output.push_back(static_cast<char>(0x80 | (codepoint & 0x3F)));
+  }
+}
+
+bool appendNumericEntity(std::string& output, const char* entity, size_t len) {
+  if (len < 4 || entity[0] != '&' || entity[1] != '#' || entity[len - 1] != ';') return false;
+  size_t pos = 2;
+  uint32_t base = 10;
+  if (entity[pos] == 'x' || entity[pos] == 'X') {
+    base = 16;
+    pos++;
+  }
+  if (pos == len - 1) return false;
+
+  uint32_t value = 0;
+  for (; pos < len - 1; pos++) {
+    const unsigned char c = entity[pos];
+    uint32_t digit;
+    if (c >= '0' && c <= '9')
+      digit = c - '0';
+    else if (base == 16 && c >= 'a' && c <= 'f')
+      digit = c - 'a' + 10;
+    else if (base == 16 && c >= 'A' && c <= 'F')
+      digit = c - 'A' + 10;
+    else
+      return false;
+    if (value > (0x10FFFF - digit) / base) return false;
+    value = value * base + digit;
+  }
+  appendCodepoint(output, value);
+  return true;
+}
+
+void appendBreak(std::string& output, size_t count = 1) {
+  while (!output.empty() && output.back() == ' ') output.pop_back();
+  if (output.empty()) return;
+  size_t existing = 0;
+  while (existing < output.size() && output[output.size() - existing - 1] == '\n') existing++;
+  while (existing++ < count) output.push_back('\n');
+}
+
+}  // namespace
+
+std::string htmlToPlainText(const std::string& html) {
+  std::string output;
+  output.reserve(html.size());
+
+  for (size_t i = 0; i < html.size();) {
+    if (html[i] == '<' && isTagStart(html, i)) {
+      const size_t close = html.find('>', i + 1);
+      if (close == std::string::npos) {
+        output.push_back(html[i++]);
+        continue;
+      }
+      const TagBreak separator = tagBreak(html, i + 1, close);
+      if (separator == TagBreak::Line)
+        appendBreak(output);
+      else if (separator == TagBreak::Paragraph)
+        appendBreak(output, 2);
+      i = close + 1;
+      continue;
+    }
+
+    if (html[i] == '&') {
+      const size_t semicolon = html.find(';', i + 1);
+      if (semicolon != std::string::npos && semicolon - i <= 16) {
+        const size_t len = semicolon - i + 1;
+        const char* value = lookupHtmlEntity(html.data() + i, len);
+        if (value != nullptr) {
+          output.append(value);
+          i = semicolon + 1;
+          continue;
+        }
+        if (appendNumericEntity(output, html.data() + i, len)) {
+          i = semicolon + 1;
+          continue;
+        }
+      }
+    }
+
+    const unsigned char c = html[i++];
+    if (c == '\r' || c == '\t') {
+      if (!output.empty() && output.back() != ' ' && output.back() != '\n') output.push_back(' ');
+    } else if (c == '\n') {
+      appendBreak(output);
+    } else {
+      output.push_back(static_cast<char>(c));
+    }
+  }
+
+  while (!output.empty() && (output.back() == ' ' || output.back() == '\n')) output.pop_back();
+  return output;
+}

src/util/HtmlToPlainText.h

@@ -0,0 +1,7 @@
+#pragma once
+
+#include <string>
+
+// Convert an HTML fragment to readable plain text. This intentionally ignores
+// styling; block elements become line breaks and HTML entities are decoded.
+std::string htmlToPlainText(const std::string& html);

src/util/StringUtils.h

@@ -1,9 +1,32 @@
 #pragma once
 
+#include <cctype>
 #include <string>
 
 namespace StringUtils {
 
+/**
+ * Case-insensitive ASCII strcmp. Returns <0, 0, or >0 like strcmp, comparing
+ * each byte by its lowercased value.
+ *
+ * Used wherever data is sorted case-insensitively: StarDict indexes (including
+ * wiktionary-derived dictionaries) and the on-disk dictionary folder list.
+ * Plain strcmp would land a binary search on the wrong page for any word whose
+ * alphabetic neighbourhood contains mixed-case boundaries.
+ *
+ * Inline (header) because Dictionary's binary search calls it per comparison
+ * step; a cross-TU call here would defeat inlining on a hot path.
+ */
+inline int asciiCaseCmp(const char* a, const char* b) {
+  while (*a && *b) {
+    int diff = std::tolower(static_cast<unsigned char>(*a)) - std::tolower(static_cast<unsigned char>(*b));
+    if (diff != 0) return diff;
+    ++a;
+    ++b;
+  }
+  return std::tolower(static_cast<unsigned char>(*a)) - std::tolower(static_cast<unsigned char>(*b));
+}
+
 /**
  * Sanitize a string for use as a filename.
  * Replaces invalid characters with underscores, trims spaces/dots,