Merge pull request #162 from spfenwick/fix-oom-crash

fix: Avoid OOM crash when using SD font with many kern pairs

lib/EpdFont/SdCardFont.cpp

@@ -52,6 +52,21 @@ void SdCardFont::freeStyleMiniData(PerStyle& s) {
   s.miniIntervalCount = 0;
   s.miniGlyphCount = 0;
   s.miniMode = PerStyle::MiniMode::NONE;
+  // Clear dangling pointers in miniData and stubData (kern data points to freed mini arrays)
+  s.miniData.kernLeftClasses = nullptr;
+  s.miniData.kernRightClasses = nullptr;
+  s.miniData.kernMatrix = nullptr;
+  s.miniData.kernLeftEntryCount = 0;
+  s.miniData.kernRightEntryCount = 0;
+  s.miniData.kernLeftClassCount = 0;
+  s.miniData.kernRightClassCount = 0;
+  s.stubData.kernLeftClasses = nullptr;
+  s.stubData.kernRightClasses = nullptr;
+  s.stubData.kernMatrix = nullptr;
+  s.stubData.kernLeftEntryCount = 0;
+  s.stubData.kernRightEntryCount = 0;
+  s.stubData.kernLeftClassCount = 0;
+  s.stubData.kernRightClassCount = 0;
   // NOTE: reportedMissCount is intentionally NOT reset here. The merge path
   // calls freeStyleMiniData() to swap mini buffers, and resetting the miss
   // tracker every paragraph would re-spam the log for the same 4 missing cps.
@@ -66,11 +81,17 @@ void SdCardFont::freeStyleMiniData(PerStyle& s) {
 void SdCardFont::freeStyleKernLigatureData(PerStyle& s) {
   delete[] s.kernLeftClasses;
   s.kernLeftClasses = nullptr;
+  s.kernClassesLoaded = false;
   delete[] s.kernRightClasses;
   s.kernRightClasses = nullptr;
   delete[] s.ligaturePairs;
   s.ligaturePairs = nullptr;
-  s.kernLigLoaded = false;
+  s.ligLoaded = false;
+  // Clear dangling pointers in EpdFontData structs
+  s.stubData.ligaturePairs = nullptr;
+  s.stubData.ligaturePairCount = 0;
+  s.miniData.ligaturePairs = nullptr;
+  s.miniData.ligaturePairCount = 0;
 }
 
 void SdCardFont::freeStyleMiniKern(PerStyle& s) {
@@ -134,14 +155,17 @@ void SdCardFont::applyKernLigaturePointers(const PerStyle& s, EpdFontData& data)
   data.ligaturePairCount = s.header.ligaturePairCount;
 }
 
-bool SdCardFont::loadStyleKernLigatureData(PerStyle& s) {
+bool SdCardFont::loadStyleKernLigatureData(PerStyle& s, bool ligatureOnly) {
-  if (s.kernLigLoaded) return true;
+  // During metadata-only (layout) prewarms, skip the kern class tables: the kern
-  bool hasKern = s.header.kernLeftEntryCount > 0;
+  // matrix is never built at layout time so getKerning() returns 0 regardless.
-  bool hasLig = s.header.ligaturePairCount > 0;
+  // Skipping them saves ~4KB per style (~17KB total for 4 styles), preventing OOM
-  if (!hasKern && !hasLig) {
+  // on low-heap devices when long paragraphs try to grow their word vector.
-    s.kernLigLoaded = true;
+  const bool wantKern = !ligatureOnly && s.header.kernLeftEntryCount > 0;
-    return true;
+  const bool wantLig = s.header.ligaturePairCount > 0;
-  }
+
+  const bool kernDone = !wantKern || s.kernClassesLoaded;
+  const bool ligDone = !wantLig || s.ligLoaded;
+  if (kernDone && ligDone) return true;
 
   FsFile file;
   if (!Storage.openFileForRead("SDCF", filePath_, file)) {
@@ -149,72 +173,79 @@ bool SdCardFont::loadStyleKernLigatureData(PerStyle& s) {
     return false;
   }
 
-  if (hasKern) {
+  if (wantKern && !s.kernClassesLoaded) {
     // Load only the small class-lookup tables (~3KB each). The full matrix
     // (~36KB contiguous for Literata) is built per-page from SD in
     // buildMiniKernMatrix().
-    s.kernLeftClasses = new (std::nothrow) EpdKernClassEntry[s.header.kernLeftEntryCount];
+    EpdKernClassEntry* newLeft = new (std::nothrow) EpdKernClassEntry[s.header.kernLeftEntryCount];
-    s.kernRightClasses = new (std::nothrow) EpdKernClassEntry[s.header.kernRightEntryCount];
+    EpdKernClassEntry* newRight = new (std::nothrow) EpdKernClassEntry[s.header.kernRightEntryCount];
 
-    if (!s.kernLeftClasses || !s.kernRightClasses) {
+    if (!newLeft || !newRight) {
+      delete[] newLeft;
+      delete[] newRight;
       LOG_ERR("SDCF", "Failed to allocate kern classes (%u+%u bytes)", s.header.kernLeftEntryCount * 3u,
               s.header.kernRightEntryCount * 3u);
-      freeStyleKernLigatureData(s);
       file.close();
       return false;
     }
 
     if (!file.seekSet(s.kernLeftFileOffset)) {
+      delete[] newLeft;
+      delete[] newRight;
       LOG_ERR("SDCF", "Failed to seek to kern data");
-      freeStyleKernLigatureData(s);
       file.close();
       return false;
     }
     size_t leftSz = s.header.kernLeftEntryCount * sizeof(EpdKernClassEntry);
     size_t rightSz = s.header.kernRightEntryCount * sizeof(EpdKernClassEntry);
-    if (file.read(reinterpret_cast<uint8_t*>(s.kernLeftClasses), leftSz) != static_cast<int>(leftSz) ||
+    if (file.read(reinterpret_cast<uint8_t*>(newLeft), leftSz) != static_cast<int>(leftSz) ||
-        file.read(reinterpret_cast<uint8_t*>(s.kernRightClasses), rightSz) != static_cast<int>(rightSz)) {
+        file.read(reinterpret_cast<uint8_t*>(newRight), rightSz) != static_cast<int>(rightSz)) {
+      delete[] newLeft;
+      delete[] newRight;
       LOG_ERR("SDCF", "Failed to read kern classes");
-      freeStyleKernLigatureData(s);
       file.close();
       return false;
     }
+    s.kernLeftClasses = newLeft;
+    s.kernRightClasses = newRight;
+    s.kernClassesLoaded = true;
   }
 
-  if (hasLig) {
+  if (wantLig && !s.ligLoaded) {
-    s.ligaturePairs = new (std::nothrow) EpdLigaturePair[s.header.ligaturePairCount];
+    EpdLigaturePair* newLig = new (std::nothrow) EpdLigaturePair[s.header.ligaturePairCount];
-    if (!s.ligaturePairs) {
+    if (!newLig) {
       LOG_ERR("SDCF", "Failed to allocate ligature pairs");
-      freeStyleKernLigatureData(s);
       file.close();
       return false;
     }
     if (!file.seekSet(s.ligatureFileOffset)) {
+      delete[] newLig;
       LOG_ERR("SDCF", "Failed to seek to ligature data");
-      freeStyleKernLigatureData(s);
       file.close();
       return false;
     }
     size_t sz = s.header.ligaturePairCount * sizeof(EpdLigaturePair);
-    if (file.read(reinterpret_cast<uint8_t*>(s.ligaturePairs), sz) != static_cast<int>(sz)) {
+    if (file.read(reinterpret_cast<uint8_t*>(newLig), sz) != static_cast<int>(sz)) {
+      delete[] newLig;
       LOG_ERR("SDCF", "Failed to read ligature pairs");
-      freeStyleKernLigatureData(s);
       file.close();
       return false;
     }
+    s.ligaturePairs = newLig;
+    s.ligLoaded = true;
+
+    // Make ligatures visible to the stub (used when no mini data built yet).
+    // Kern stays nullptr on the stub — it is only wired in miniData via
+    // applyKernLigaturePointers() after buildMiniKernMatrix() runs.
+    s.stubData.ligaturePairs = s.ligaturePairs;
+    s.stubData.ligaturePairCount = s.header.ligaturePairCount;
   }
 
   file.close();
-  s.kernLigLoaded = true;
+  LOG_DBG("SDCF", "Kern/lig loaded: kernL=%u kernR=%u ligs=%u ligOnly=%d",
-
+          s.kernClassesLoaded ? s.header.kernLeftEntryCount : 0u,
-  // Make ligatures visible to the stub (used when no mini data built yet).
+          s.kernClassesLoaded ? s.header.kernRightEntryCount : 0u, s.ligLoaded ? s.header.ligaturePairCount : 0u,
-  // Kern stays nullptr on the stub — it is only wired in miniData via
+          ligatureOnly);
-  // applyKernLigaturePointers() after buildMiniKernMatrix() runs.
-  s.stubData.ligaturePairs = s.ligaturePairs;
-  s.stubData.ligaturePairCount = s.header.ligaturePairCount;
-
-  LOG_DBG("SDCF", "Kern classes + lig loaded: kernL=%u, kernR=%u, ligs=%u", s.header.kernLeftEntryCount,
-          s.header.kernRightEntryCount, s.header.ligaturePairCount);
   return true;
 }
 
@@ -653,7 +684,7 @@ int SdCardFont::prewarm(const char* utf8Text, uint8_t styleMask, bool metadataOn
       if (!(styleMask & (1 << si)) || !styles_[si].present) continue;
       auto& s = styles_[si];
 
-      loadStyleKernLigatureData(s);
+      loadStyleKernLigatureData(s, /*ligatureOnly=*/true);
       if (s.ligaturePairs && s.header.ligaturePairCount > 0) {
         for (uint8_t li = 0; li < s.header.ligaturePairCount && cpCount < MAX_PAGE_GLYPHS; li++) {
           uint32_t leftCp = s.ligaturePairs[li].pair >> 16;
@@ -736,6 +767,18 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
     // For metadata-only calls, METADATA or FULL cache both satisfy layout queries.
     // For full (bitmap) calls, only FULL satisfies — METADATA lacks bitmap data.
     if (metadataOnly || s.miniMode == PerStyle::MiniMode::FULL) {
+      // Ensure ligature metadata is wired if requested but not yet wired.
+      if (loadKernLigatureData && !s.ligLoaded) {
+        loadStyleKernLigatureData(s, /*ligatureOnly=*/true);
+      }
+      if (loadKernLigatureData && s.ligLoaded && s.miniData.ligaturePairs == nullptr) {
+        if (s.miniMode == PerStyle::MiniMode::FULL) {
+          applyKernLigaturePointers(s, s.miniData);
+        } else {
+          s.miniData.ligaturePairs = s.ligaturePairs;
+          s.miniData.ligaturePairCount = s.header.ligaturePairCount;
+        }
+      }
       // Already wired into miniData; nothing else to do.
       return 0;
     }
@@ -1143,7 +1186,7 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
       kernLigOk = buildMiniKernMatrix(s, codepoints, cpCount);
     }
   } else if (loadKernLigatureData) {
-    loadStyleKernLigatureData(s);
+    loadStyleKernLigatureData(s, /*ligatureOnly=*/true);
     // Don't set kernLigOk → mini kern matrix stays null on miniData, but
     // ligatures are still resident on stubData (set in loadStyleKernLigatureData).
   }
@@ -1161,7 +1204,7 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
   if (kernLigOk) {
     // Full prewarm: wire mini kern matrix + class tables + ligatures.
     applyKernLigaturePointers(s, s.miniData);
-  } else if (loadKernLigatureData && s.kernLigLoaded) {
+  } else if (loadKernLigatureData && s.ligLoaded) {
     // Layout-only prewarm: wire ligatures so applyLigatures() works (e.g. "fi"
     // measures correctly). Skip the kern matrix — getKerning() returns 0
     // cleanly when kernMatrix is null. Per-pair kern is applied at render time.
@@ -1173,6 +1216,8 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
 
   s.epdFont.data = &s.miniData;
   s.miniMode = metadataOnly ? PerStyle::MiniMode::METADATA : PerStyle::MiniMode::FULL;
+  LOG_DBG("SDCF", "prewarmStyle %u: mode→%s glyphs=%u bitmap=%p", styleIdx, metadataOnly ? "METADATA" : "FULL",
+          validCount, s.miniBitmap);
 
   // Accumulate stats
   stats_.sdReadTimeMs += sdTime;
@@ -1236,6 +1281,10 @@ const EpdGlyph* SdCardFont::onGlyphMiss(void* ctx, uint32_t codepoint) {
   const auto& s = self->styles_[styleIdx];
   if (!s.fullIntervals) return nullptr;
 
+  // Diagnostic: log first miss per codepoint+style to show why it bypassed prewarm
+  LOG_DBG("SDCF", "onGlyphMiss: U+%04X style %u miniMode=%u miniIntervals=%u bitmap=%p", codepoint, styleIdx,
+          (uint8_t)s.miniMode, s.miniIntervalCount, s.miniBitmap);
+
   // Check overflow cache first (matching both codepoint and style)
   for (uint32_t i = 0; i < self->overflowCount_; i++) {
     if (self->overflow_[i].codepoint == codepoint && self->overflow_[i].styleIdx == styleIdx) {
@@ -1309,6 +1358,8 @@ const EpdGlyph* SdCardFont::onGlyphMiss(void* ctx, uint32_t codepoint) {
 
   // All reads succeeded — commit to slot (evict old entry if at capacity)
   if (wasAtCapacity) {
+    LOG_DBG("SDCF", "Overflow: evicting U+%04X style %u from slot %u", self->overflow_[slot].codepoint,
+            self->overflow_[slot].styleIdx, slot);
     delete[] self->overflow_[slot].bitmap;
   }
   self->overflow_[slot].glyph = tempGlyph;

lib/EpdFont/SdCardFont.h

@@ -118,7 +118,8 @@ class SdCardFont {
     EpdKernClassEntry* kernLeftClasses = nullptr;
     EpdKernClassEntry* kernRightClasses = nullptr;
     EpdLigaturePair* ligaturePairs = nullptr;
-    bool kernLigLoaded = false;
+    bool ligLoaded = false;          ///< ligaturePairs resident
+    bool kernClassesLoaded = false;  ///< kernLeft/RightClasses resident (skipped during metadata-only prewarm)
 
     // Stub EpdFontData returned when not prewarmed
     EpdFontData stubData{};
@@ -204,7 +205,7 @@ class SdCardFont {
   void freeStyleAll(PerStyle& s);
   void freeStyleKernLigatureData(PerStyle& s);
   void freeStyleMiniKern(PerStyle& s);
-  bool loadStyleKernLigatureData(PerStyle& s);
+  bool loadStyleKernLigatureData(PerStyle& s, bool ligatureOnly = false);
   bool buildMiniKernMatrix(PerStyle& s, const uint32_t* codepoints, uint32_t cpCount);
   void applyKernLigaturePointers(const PerStyle& s, EpdFontData& data) const;
   void applyGlyphMissCallback(uint8_t styleIdx);

lib/Epub/Epub/ParsedText.cpp

@@ -215,7 +215,14 @@ void ParsedText::layoutAndExtractLines(
   }
 
   // Apply fixed transforms before any per-line layout work.
-  applyParagraphIndent();
+  // Paragraph indent only applies to the first layout pass; skip on continuations.
+  if (!isContinuation_) {
+    applyParagraphIndent();
+  }
+  // Bionic transform is incremental: applyBionicReadingTransform() is a no-op
+  // for already-transformed words (bionicTransformedUpTo_ == words.size()) and
+  // only processes raw words appended since the last flush, so it is always safe
+  // to call regardless of isContinuation_.
   if (bionicReadingEnabled) {
     applyBionicReadingTransform();
   }
@@ -242,6 +249,15 @@ void ParsedText::layoutAndExtractLines(
   }
 
   const int pageWidth = viewportWidth;
+
+  // Compute firstLineIndent once here so all layout helpers use the same value.
+  // On a continuation flush the remaining words are mid-paragraph, so no indent.
+  const int firstLineIndent =
+      !isContinuation_ && blockStyle.textIndentDefined &&
+              (blockStyle.alignment == CssTextAlign::Justify || blockStyle.alignment == CssTextAlign::Left)
+          ? std::min(std::max<int>(static_cast<int>(blockStyle.textIndent), -(pageWidth - 1)), pageWidth - 1)
+          : 0;
+
   auto wordWidths = calculateWordWidths(renderer, fontId);
 
   std::vector<size_t> lineBreakIndices;
@@ -252,9 +268,9 @@ void ParsedText::layoutAndExtractLines(
     // Use greedy layout that can split words mid-loop when a hyphenated prefix fits.
     lineBreakIndices =
         computeHyphenatedLineBreaks(renderer, fontId, pageWidth, wordWidths, wordContinues, lineEndsWithHyphenatedWord,
-                                    splitPrefixWordIndexes, splitInsertedHyphen);
+                                    splitPrefixWordIndexes, splitInsertedHyphen, firstLineIndent);
   } else {
-    lineBreakIndices = computeLineBreaks(renderer, fontId, pageWidth, wordWidths, wordContinues);
+    lineBreakIndices = computeLineBreaks(renderer, fontId, pageWidth, wordWidths, wordContinues, firstLineIndent);
     lineEndsWithHyphenatedWord.assign(lineBreakIndices.size(), false);
     splitPrefixWordIndexes.assign(lineBreakIndices.size(), -1);
     splitInsertedHyphen.assign(lineBreakIndices.size(), false);
@@ -264,7 +280,7 @@ void ParsedText::layoutAndExtractLines(
   for (size_t i = 0; i < lineCount; ++i) {
     const bool lineEndedWithHyphenation = i < lineEndsWithHyphenatedWord.size() ? lineEndsWithHyphenatedWord[i] : false;
     const auto result = extractLine(i, pageWidth, wordWidths, wordContinues, lineBreakIndices, processLine, renderer,
-                                    fontId, lineEndedWithHyphenation, false);
+                                    fontId, lineEndedWithHyphenation, false, firstLineIndent);
 
     if (result == LineProcessResult::RetryWithoutHyphenation && lineEndedWithHyphenation) {
       const size_t lineStart = i > 0 ? lineBreakIndices[i - 1] : 0;
@@ -309,8 +325,8 @@ void ParsedText::layoutAndExtractLines(
 
       // Keep previous lines fixed; recompute only this specific line without hyphenation.
       // Suppression is intentionally line-local.
-      const size_t retryBreak =
+      const size_t retryBreak = computeSingleLineBreakNoHyphen(renderer, fontId, pageWidth, wordWidths, wordContinues,
-          computeSingleLineBreakNoHyphen(renderer, fontId, pageWidth, wordWidths, wordContinues, lineStart);
+                                                               lineStart, firstLineIndent);
 
       lineBreakIndices.resize(i + 1);
       lineEndsWithHyphenatedWord.resize(i + 1);
@@ -330,7 +346,7 @@ void ParsedText::layoutAndExtractLines(
         LOG_DBG("PTX", "Rerendering line %u with hyphenation suppressed, retry attempt: %s", static_cast<unsigned>(i),
                 retryPreview.c_str());
         extractLine(i, pageWidth, wordWidths, wordContinues, lineBreakIndices, processLine, renderer, fontId, false,
-                    true);
+                    true, firstLineIndent);
 
         // Resume regular hyphenation from the first word after the retried line.
         const size_t resumeIndex = lineBreakIndices[i];
@@ -355,13 +371,23 @@ void ParsedText::layoutAndExtractLines(
     }
   }
 
-  // Remove consumed words so size() reflects only remaining words
+  // Remove consumed words so size() reflects only remaining words, then
+  // release excess capacity.  Without shrink_to_fit the vector retains a
+  // large allocation from before the flush; the next paragraph fills it
+  // back up and eventually needs an even larger contiguous realloc.
   if (lineCount > 0) {
     const size_t consumed = lineBreakIndices[lineCount - 1];
     words.erase(words.begin(), words.begin() + consumed);
     wordStyles.erase(wordStyles.begin(), wordStyles.begin() + consumed);
     wordContinues.erase(wordContinues.begin(), wordContinues.begin() + consumed);
+    words.shrink_to_fit();
+    wordStyles.shrink_to_fit();
+    wordContinues.shrink_to_fit();
+    // All remaining words were already transformed before the flush; reset the
+    // watermark so that words appended by addWord() are processed next time.
+    bionicTransformedUpTo_ = words.size();
   }
+  isContinuation_ = !includeLastLine;
 }
 
 std::vector<uint16_t> ParsedText::calculateWordWidths(const GfxRenderer& renderer, const int fontId) {
@@ -376,20 +402,12 @@ std::vector<uint16_t> ParsedText::calculateWordWidths(const GfxRenderer& rendere
 }
 
 std::vector<size_t> ParsedText::computeLineBreaks(const GfxRenderer& renderer, const int fontId, const int pageWidth,
-                                                  std::vector<uint16_t>& wordWidths, std::vector<bool>& continuesVec) {
+                                                  std::vector<uint16_t>& wordWidths, std::vector<bool>& continuesVec,
+                                                  const int firstLineIndent) {
   if (words.empty()) {
     return {};
   }
 
-  // Calculate first line indent (only for left/justified text).
-  // Explicit CSS text-indent always applies — author intent overrides the extraParagraphSpacing
-  // toggle. Only the implicit EmSpace fallback in applyParagraphIndent() is gated on it.
-  const int firstLineIndent =
-      blockStyle.textIndentDefined &&
-              (blockStyle.alignment == CssTextAlign::Justify || blockStyle.alignment == CssTextAlign::Left)
-          ? std::min(std::max<int>(static_cast<int>(blockStyle.textIndent), -(pageWidth - 1)), pageWidth - 1)
-          : 0;
-
   // Ensure any word that would overflow even as the first entry on a line is split using fallback hyphenation.
   for (size_t i = 0; i < wordWidths.size(); ++i) {
     // First word needs to fit in reduced width if there's an indent
@@ -510,19 +528,14 @@ std::vector<size_t> ParsedText::computeLineBreaks(const GfxRenderer& renderer, c
 
 size_t ParsedText::computeSingleLineBreakNoHyphen(const GfxRenderer& renderer, const int fontId, const int pageWidth,
                                                   const std::vector<uint16_t>& wordWidths,
-                                                  const std::vector<bool>& continuesVec,
+                                                  const std::vector<bool>& continuesVec, const size_t lineStartIndex,
-                                                  const size_t lineStartIndex) const {
+                                                  const int firstLineIndent) const {
   // One-line non-hyphenating breaker used by the page-boundary retry path.
   if (lineStartIndex >= wordWidths.size()) {
     return lineStartIndex;
   }
 
-  const int firstLineIndent =
+  const int effectivePageWidth = pageWidth - (lineStartIndex == 0 ? firstLineIndent : 0);
-      lineStartIndex == 0 && blockStyle.textIndentDefined &&
-              (blockStyle.alignment == CssTextAlign::Justify || blockStyle.alignment == CssTextAlign::Left)
-          ? std::min(std::max<int>(static_cast<int>(blockStyle.textIndent), -(pageWidth - 1)), pageWidth - 1)
-          : 0;
-  const int effectivePageWidth = pageWidth - firstLineIndent;
 
   size_t currentIndex = lineStartIndex;
   int lineWidth = 0;
@@ -577,22 +590,26 @@ void ParsedText::applyParagraphIndent() {
 }
 
 void ParsedText::applyBionicReadingTransform() {
-  if (words.empty()) {
+  // Only transform words that haven't been processed yet.  On a fresh block
+  // bionicTransformedUpTo_ == 0 so all words are processed.  After an
+  // intermediate flush, only the new raw words appended since the last flush
+  // (indices bionicTransformedUpTo_..words.size()-1) need transformation.
+  if (words.empty() || bionicTransformedUpTo_ >= words.size()) {
     return;
   }
 
-  std::vector<std::string> transformedWords;
+  const size_t suffixStart = bionicTransformedUpTo_;
-  std::vector<EpdFontFamily::Style> transformedStyles;
+  std::vector<std::string> transformedSuffix;
-  std::vector<bool> transformedContinues;
+  std::vector<EpdFontFamily::Style> transformedSuffixStyles;
-  transformedWords.reserve(words.size() * 2);
+  std::vector<bool> transformedSuffixContinues;
-  transformedStyles.reserve(wordStyles.size() * 2);
+  transformedSuffix.reserve((words.size() - suffixStart) * 2);
-  transformedContinues.reserve(wordContinues.size() * 2);
+  transformedSuffixStyles.reserve(transformedSuffix.capacity());
+  transformedSuffixContinues.reserve(transformedSuffix.capacity());
 
-  for (size_t i = 0; i < words.size(); ++i) {
+  for (size_t i = suffixStart; i < words.size(); ++i) {
     std::string source = std::move(words[i]);
     const auto originalStyle = wordStyles[i];
     const bool originalAttachToPrevious = wordContinues[i];
-    const char* raw = source.c_str();
 
     const auto spans = tokenizeBionicWord(source);
     if (spans.empty()) {
@@ -603,12 +620,7 @@ void ParsedText::applyBionicReadingTransform() {
     for (size_t spanIndex = 0; spanIndex < spans.size(); ++spanIndex) {
       const TokenSpan span = spans[spanIndex];
       const size_t spanLength = span.end - span.start;
-      std::string token;
+      std::string token = source.substr(span.start, spanLength);
-      if (spans.size() == 1 && spanIndex == 0) {
-        token = std::move(source);
-      } else {
-        token.assign(raw + span.start, spanLength);
-      }
 
       if (span.isWord) {
         const unsigned char* ptr = reinterpret_cast<const unsigned char*>(token.c_str());
@@ -630,47 +642,42 @@ void ParsedText::applyBionicReadingTransform() {
             std::string suffix(reinterpret_cast<const char*>(prefixEnd), token.size() - prefixByteCount);
             token.resize(prefixByteCount);
             const auto boldStyle = static_cast<EpdFontFamily::Style>(originalStyle | EpdFontFamily::BOLD);
-            transformedWords.push_back(std::move(token));
+            transformedSuffix.push_back(std::move(token));
-            transformedStyles.push_back(boldStyle);
+            transformedSuffixStyles.push_back(boldStyle);
-            transformedContinues.push_back(attachToPrevious);
+            transformedSuffixContinues.push_back(attachToPrevious);
 
-            transformedWords.push_back(std::move(suffix));
+            transformedSuffix.push_back(std::move(suffix));
-            transformedStyles.push_back(originalStyle);
+            transformedSuffixStyles.push_back(originalStyle);
-            transformedContinues.push_back(true);
+            transformedSuffixContinues.push_back(true);
             attachToPrevious = true;
             continue;
           }
         }
       }
 
-      transformedWords.push_back(std::move(token));
+      transformedSuffix.push_back(std::move(token));
-      transformedStyles.push_back(originalStyle);
+      transformedSuffixStyles.push_back(originalStyle);
-      transformedContinues.push_back(attachToPrevious);
+      transformedSuffixContinues.push_back(attachToPrevious);
       attachToPrevious = true;
     }
   }
 
-  words = std::move(transformedWords);
+  // Replace the (now move-emptied) suffix with the transformed version.
-  wordStyles = std::move(transformedStyles);
+  words.resize(suffixStart);
-  wordContinues = std::move(transformedContinues);
+  wordStyles.resize(suffixStart);
+  wordContinues.resize(suffixStart);
+  words.insert(words.end(), std::make_move_iterator(transformedSuffix.begin()),
+               std::make_move_iterator(transformedSuffix.end()));
+  wordStyles.insert(wordStyles.end(), transformedSuffixStyles.begin(), transformedSuffixStyles.end());
+  wordContinues.insert(wordContinues.end(), transformedSuffixContinues.begin(), transformedSuffixContinues.end());
+  bionicTransformedUpTo_ = words.size();
 }
 
 // Builds break indices while opportunistically splitting the word that would overflow the current line.
-std::vector<size_t> ParsedText::computeHyphenatedLineBreaks(const GfxRenderer& renderer, const int fontId,
+std::vector<size_t> ParsedText::computeHyphenatedLineBreaks(
-                                                            const int pageWidth, std::vector<uint16_t>& wordWidths,
+    const GfxRenderer& renderer, const int fontId, const int pageWidth, std::vector<uint16_t>& wordWidths,
-                                                            std::vector<bool>& continuesVec,
+    std::vector<bool>& continuesVec, std::vector<bool>& lineEndsWithHyphenatedWord,
-                                                            std::vector<bool>& lineEndsWithHyphenatedWord,
+    std::vector<int>& splitPrefixWordIndexes, std::vector<bool>& splitInsertedHyphen, const int firstLineIndent) {
-                                                            std::vector<int>& splitPrefixWordIndexes,
-                                                            std::vector<bool>& splitInsertedHyphen) {
-  // Calculate first line indent (only for left/justified text).
-  // Explicit CSS text-indent always applies — author intent overrides the extraParagraphSpacing
-  // toggle. Only the implicit EmSpace fallback in applyParagraphIndent() is gated on it.
-  const int firstLineIndent =
-      blockStyle.textIndentDefined &&
-              (blockStyle.alignment == CssTextAlign::Justify || blockStyle.alignment == CssTextAlign::Left)
-          ? std::min(std::max<int>(static_cast<int>(blockStyle.textIndent), -(pageWidth - 1)), pageWidth - 1)
-          : 0;
-
   // Pre-compute inter-word gaps to avoid repeated codepoint scanning and renderer
   // calls in the inner loop. When hyphenateWordAtIndex inserts a new word, we insert
   // a placeholder gap (0) at that position to keep the vector in sync; the remainder
@@ -949,20 +956,14 @@ ParsedText::LineProcessResult ParsedText::extractLine(
     const std::vector<bool>& continuesVec, const std::vector<size_t>& lineBreakIndices,
     const std::function<LineProcessResult(std::shared_ptr<TextBlock>, bool, bool)>& processLine,
     const GfxRenderer& renderer, const int fontId, const bool lineEndsWithHyphenatedWord,
-    const bool suppressHyphenationRetry) {
+    const bool suppressHyphenationRetry, const int firstLineIndent) {
   const size_t lineBreak = lineBreakIndices[breakIndex];
   const size_t lastBreakAt = breakIndex > 0 ? lineBreakIndices[breakIndex - 1] : 0;
   const size_t lineWordCount = lineBreak - lastBreakAt;
 
-  // Calculate first line indent (only for left/justified text).
+  // Apply indent only to line 0 of the layout pass; firstLineIndent is already
-  // Explicit CSS text-indent always applies — author intent overrides the extraParagraphSpacing
+  // 0 for continuation flushes (computed once in layoutAndExtractLines).
-  // toggle. Only the implicit EmSpace fallback in applyParagraphIndent() is gated on it.
+  const int lineIndent = (breakIndex == 0) ? firstLineIndent : 0;
-  const bool isFirstLine = breakIndex == 0;
-  const int firstLineIndent =
-      isFirstLine && blockStyle.textIndentDefined &&
-              (blockStyle.alignment == CssTextAlign::Justify || blockStyle.alignment == CssTextAlign::Left)
-          ? std::min(std::max<int>(static_cast<int>(blockStyle.textIndent), -(pageWidth - 1)), pageWidth - 1)
-          : 0;
 
   // Calculate total word width for this line, count actual word gaps,
   // and accumulate total natural gap widths (including space kerning adjustments).
@@ -989,7 +990,7 @@ ParsedText::LineProcessResult ParsedText::extractLine(
   }
 
   // Calculate spacing (account for indent reducing effective page width on first line)
-  const int effectivePageWidth = pageWidth - firstLineIndent;
+  const int effectivePageWidth = pageWidth - lineIndent;
   // A line is only truly last when it consumes all paragraph words.
   // During single-line retry we may temporarily pass a truncated break vector,
   // so relying only on breakIndex would incorrectly disable justification.
@@ -1003,7 +1004,7 @@ ParsedText::LineProcessResult ParsedText::extractLine(
 
   // Calculate initial x position (first line starts at indent for left/justified text;
   // may be negative for hanging indents, e.g. margin-left:3em; text-indent:-1em).
-  auto xpos = static_cast<int16_t>(firstLineIndent);
+  auto xpos = static_cast<int16_t>(lineIndent);
   if (blockStyle.alignment == CssTextAlign::Right) {
     xpos = effectivePageWidth - lineWordWidthSum - totalNaturalGaps;
   } else if (blockStyle.alignment == CssTextAlign::Center) {

lib/Epub/Epub/ParsedText.h

@@ -27,16 +27,19 @@ class ParsedText {
   bool extraParagraphSpacing;
   bool hyphenationEnabled;
   bool bionicReadingEnabled;
+  bool isContinuation_ = false;       ///< true after an intermediate flush; suppresses re-applying paragraph indent
+  size_t bionicTransformedUpTo_ = 0;  ///< words[0..bionicTransformedUpTo_) have already been bionic-transformed
 
   void applyParagraphIndent();
   void applyBionicReadingTransform();
   std::vector<size_t> computeLineBreaks(const GfxRenderer& renderer, int fontId, int pageWidth,
-                                        std::vector<uint16_t>& wordWidths, std::vector<bool>& continuesVec);
+                                        std::vector<uint16_t>& wordWidths, std::vector<bool>& continuesVec,
+                                        int firstLineIndent);
   std::vector<size_t> computeHyphenatedLineBreaks(const GfxRenderer& renderer, int fontId, int pageWidth,
                                                   std::vector<uint16_t>& wordWidths, std::vector<bool>& continuesVec,
                                                   std::vector<bool>& lineEndsWithHyphenatedWord,
                                                   std::vector<int>& splitPrefixWordIndexes,
-                                                  std::vector<bool>& splitInsertedHyphen);
+                                                  std::vector<bool>& splitInsertedHyphen, int firstLineIndent);
   // Recompute hyphenated breaks for a suffix that starts at startIndex.
   // Used after a single-line retry so later lines keep normal hyphenation.
   std::vector<size_t> computeHyphenatedLineBreaksFromIndex(const GfxRenderer& renderer, int fontId, int pageWidth,
@@ -49,7 +52,7 @@ class ParsedText {
   // Used only for the page-boundary retry line.
   size_t computeSingleLineBreakNoHyphen(const GfxRenderer& renderer, int fontId, int pageWidth,
                                         const std::vector<uint16_t>& wordWidths, const std::vector<bool>& continuesVec,
-                                        size_t lineStartIndex) const;
+                                        size_t lineStartIndex, int firstLineIndent) const;
   bool hyphenateWordAtIndex(size_t wordIndex, int availableWidth, const GfxRenderer& renderer, int fontId,
                             std::vector<uint16_t>& wordWidths, bool allowFallbackBreaks,
                             bool* outInsertedHyphen = nullptr);
@@ -57,7 +60,8 @@ class ParsedText {
       size_t breakIndex, int pageWidth, const std::vector<uint16_t>& wordWidths, const std::vector<bool>& continuesVec,
       const std::vector<size_t>& lineBreakIndices,
       const std::function<LineProcessResult(std::shared_ptr<TextBlock>, bool, bool)>& processLine,
-      const GfxRenderer& renderer, int fontId, bool lineEndsWithHyphenatedWord, bool suppressHyphenationRetry);
+      const GfxRenderer& renderer, int fontId, bool lineEndsWithHyphenatedWord, bool suppressHyphenationRetry,
+      int firstLineIndent);
   std::vector<uint16_t> calculateWordWidths(const GfxRenderer& renderer, int fontId);
 
  public:
@@ -74,6 +78,7 @@ class ParsedText {
   BlockStyle& getBlockStyle() { return blockStyle; }
   size_t size() const { return words.size(); }
   bool isEmpty() const { return words.empty(); }
+  bool isContinuation() const { return isContinuation_; }
   void layoutAndExtractLines(
       const GfxRenderer& renderer, int fontId, uint16_t viewportWidth,
       const std::function<LineProcessResult(std::shared_ptr<TextBlock>, bool, bool)>& processLine,

lib/Epub/Epub/parsers/ChapterHtmlSlimParser.cpp

@@ -189,6 +189,20 @@ void ChapterHtmlSlimParser::flushPartWordBuffer() {
   currentTextBlock->addWord(partWordBuffer, fontStyle, false, nextWordContinues);
   partWordBufferIndex = 0;
   nextWordContinues = false;
+
+  if (currentTextBlock->size() > 96) {
+    LOG_DBG("EHP", "Text block too long, splitting into multiple pages");
+    const int horizontalInset = currentTextBlock->getBlockStyle().totalHorizontalInset();
+    const uint16_t effectiveWidth =
+        (horizontalInset < viewportWidth) ? static_cast<uint16_t>(viewportWidth - horizontalInset) : viewportWidth;
+    currentTextBlock->layoutAndExtractLines(
+        renderer, fontId, effectiveWidth,
+        [this](const std::shared_ptr<TextBlock>& textBlock, const bool lineEndsWithHyphenatedWord,
+               const bool suppressHyphenationRetry) {
+          return addLineToPage(textBlock, lineEndsWithHyphenatedWord, suppressHyphenationRetry);
+        },
+        false);
+  }
 }
 
 // Emit the current page, keeping paragraphLutPerPage and completedPageCount in lockstep.
@@ -1206,25 +1220,6 @@ void XMLCALL ChapterHtmlSlimParser::characterData(void* userData, const XML_Char
 
     self->partWordBuffer[self->partWordBufferIndex++] = s[i];
   }
-
-  // If we have > 750 words buffered up, perform the layout and consume out all but the last line
-  // There should be enough here to build out 1-2 full pages and doing this will free up a lot of
-  // memory.
-  // Spotted when reading Intermezzo, there are some really long text blocks in there.
-  if (self->currentTextBlock->size() > 750) {
-    LOG_DBG("EHP", "Text block too long, splitting into multiple pages");
-    const int horizontalInset = self->currentTextBlock->getBlockStyle().totalHorizontalInset();
-    const uint16_t effectiveWidth = (horizontalInset < self->viewportWidth)
-                                        ? static_cast<uint16_t>(self->viewportWidth - horizontalInset)
-                                        : self->viewportWidth;
-    self->currentTextBlock->layoutAndExtractLines(
-        self->renderer, self->fontId, effectiveWidth,
-        [self](const std::shared_ptr<TextBlock>& textBlock, const bool lineEndsWithHyphenatedWord,
-               const bool suppressHyphenationRetry) {
-          return self->addLineToPage(textBlock, lineEndsWithHyphenatedWord, suppressHyphenationRetry);
-        },
-        false);
-  }
 }
 
 void XMLCALL ChapterHtmlSlimParser::defaultHandlerExpand(void* userData, const XML_Char* s, const int len) {
@@ -1591,13 +1586,17 @@ void ChapterHtmlSlimParser::makePages() {
 
   const int lineHeight = renderer.getLineHeight(fontId) * lineCompression;
 
-  // Apply top spacing before the paragraph (stored in pixels)
+  // Apply top spacing before the paragraph — skip for continuation fragments
+  // (words left over after an intermediate flush): the top margin was already
+  // applied before the first set of lines from this logical paragraph.
   const BlockStyle& blockStyle = currentTextBlock->getBlockStyle();
-  if (blockStyle.marginTop > 0) {
+  if (!currentTextBlock->isContinuation()) {
-    currentPageNextY += blockStyle.marginTop;
+    if (blockStyle.marginTop > 0) {
-  }
+      currentPageNextY += blockStyle.marginTop;
-  if (blockStyle.paddingTop > 0) {
+    }
-    currentPageNextY += blockStyle.paddingTop;
+    if (blockStyle.paddingTop > 0) {
+      currentPageNextY += blockStyle.paddingTop;
+    }
   }
 
   // Calculate effective width accounting for horizontal margins/padding