Implement keep-if-fits buffer reuse to reduce heap fragmentation

Replace per-page buffer free+realloc pattern with capacity tracking that only reallocates when needed size exceeds current capacity. This prevents heap fragmentation from non-coalescing holes that occurred when each page's freed block rarely fit the next page's allocation. After a few page turns, capacities converge on the book's maximum and page turns stop touching the allocator entirely.

lib/EpdFont/SdCardFont.cpp

@@ -68,6 +68,22 @@ bool collectUniqueCodepoints(const char* text, uint32_t* codepoints, uint32_t& c
 const char* asCStr(const std::string& s) { return s.c_str(); }
 const char* asCStr(const char* s) { return s; }
 
+// Keep-if-fits buffer reuse: only reallocate when the needed size exceeds the
+// current capacity. Freeing + reallocating slightly different sizes every page
+// turn punches non-coalescing holes in the heap (the freed block rarely fits the
+// next page's need), eroding the largest contiguous block all session. With
+// reuse, capacities converge on the book's max page after a few turns and page
+// turns stop touching the allocator. Only three small instantiations exist
+// (interval/glyph/byte arrays), so template bloat is negligible.
+template <typename T, typename CapT>
+bool ensureArrayCapacity(T*& buf, CapT& capacity, const uint32_t needed) {
+  if (buf && capacity >= needed) return true;
+  delete[] buf;
+  buf = new (std::nothrow) T[needed > 0 ? needed : 1];
+  capacity = buf ? static_cast<CapT>(needed) : 0;
+  return buf != nullptr;
+}
+
 }  // namespace
 
 SdCardFont::~SdCardFont() { freeAll(); }
@@ -83,6 +99,9 @@ void SdCardFont::freeStyleMiniData(PerStyle& s) {
   s.miniBitmap = nullptr;
   s.miniIntervalCount = 0;
   s.miniGlyphCount = 0;
+  s.miniIntervalCapacity = 0;
+  s.miniGlyphCapacity = 0;
+  s.miniBitmapCapacity = 0;
   freeStyleMiniKern(s);
   memset(&s.miniData, 0, sizeof(s.miniData));
   s.epdFont.data = &s.stubData;
@@ -109,6 +128,9 @@ void SdCardFont::freeStyleMiniKern(PerStyle& s) {
   s.miniKernRightEntryCount = 0;
   s.miniKernLeftClassCount = 0;
   s.miniKernRightClassCount = 0;
+  s.miniKernLeftCapacity = 0;
+  s.miniKernRightCapacity = 0;
+  s.miniKernMatrixCapacity = 0;
 }
 
 void SdCardFont::freeStyleAll(PerStyle& s) {
@@ -311,13 +333,13 @@ bool SdCardFont::buildMiniKernMatrix(PerStyle& s, const uint32_t* codepoints, ui
     if (miniLookupKernClass(s.kernRightClasses, s.header.kernRightEntryCount, codepoints[i]) != 0) miniRightCount++;
   }
 
-  // Step 4: allocate the three mini buffers. The matrix is <1KB in practice
-  // (<30 × <30 × 1 byte) so fragmentation is a non-issue.
+  // Step 4: size the three mini buffers (reused across pages when they fit; the
+  // per-page sizes vary by a few entries, which as free+realloc churn was punching
+  // non-coalescing holes in the heap every page turn).
   const uint32_t matrixBytes = static_cast<uint32_t>(numLeft) * numRight;
-  s.miniKernLeftClasses = new (std::nothrow) EpdKernClassEntry[miniLeftCount];
-  s.miniKernRightClasses = new (std::nothrow) EpdKernClassEntry[miniRightCount];
-  s.miniKernMatrix = new (std::nothrow) int8_t[matrixBytes];
-  if (!s.miniKernLeftClasses || !s.miniKernRightClasses || !s.miniKernMatrix) {
+  if (!ensureArrayCapacity(s.miniKernLeftClasses, s.miniKernLeftCapacity, miniLeftCount) ||
+      !ensureArrayCapacity(s.miniKernRightClasses, s.miniKernRightCapacity, miniRightCount) ||
+      !ensureArrayCapacity(s.miniKernMatrix, s.miniKernMatrixCapacity, matrixBytes)) {
     LOG_ERR("SDCF", "Failed to allocate mini kern (%u+%u+%u bytes)", miniLeftCount * 3u, miniRightCount * 3u,
             matrixBytes);
     freeStyleMiniKern(s);
@@ -793,12 +815,19 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
     return missed;
   }
 
-  // Build mini intervals from sorted codepoints
-  freeStyleMiniData(s);
+  // Build mini intervals from sorted codepoints. Reset counts and fall back to the
+  // stub until the rebuild completes, but KEEP the existing buffers (keep-if-fits
+  // reuse) — the free-and-realloc-per-page pattern here was a primary fragmenter.
+  s.miniIntervalCount = 0;
+  s.miniGlyphCount = 0;
+  s.miniKernLeftEntryCount = 0;
+  s.miniKernRightEntryCount = 0;
+  s.miniKernLeftClassCount = 0;
+  s.miniKernRightClassCount = 0;
+  memset(&s.miniData, 0, sizeof(s.miniData));
+  s.epdFont.data = &s.stubData;
 
-  uint32_t intervalCapacity = validCount;
-  s.miniIntervals = new (std::nothrow) EpdUnicodeInterval[intervalCapacity];
-  if (!s.miniIntervals) {
+  if (!ensureArrayCapacity(s.miniIntervals, s.miniIntervalCapacity, validCount)) {
     LOG_ERR("SDCF", "Failed to allocate mini intervals for style %u", styleIdx);
     delete[] mappings;
     return static_cast<int>(cpCount);
@@ -816,15 +845,14 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
     }
   }
 
-  // Allocate mini glyph array
-  s.miniGlyphCount = validCount;
-  s.miniGlyphs = new (std::nothrow) EpdGlyph[s.miniGlyphCount];
-  if (!s.miniGlyphs) {
+  // Mini glyph array (reused across pages when it fits)
+  if (!ensureArrayCapacity(s.miniGlyphs, s.miniGlyphCapacity, validCount)) {
     LOG_ERR("SDCF", "Failed to allocate mini glyphs for style %u", styleIdx);
     delete[] mappings;
     freeStyleMiniData(s);
     return static_cast<int>(cpCount);
   }
+  s.miniGlyphCount = validCount;
 
   // Build sorted read order for sequential I/O
   uint32_t* readOrder = new (std::nothrow) uint32_t[validCount];
@@ -891,8 +919,7 @@ int SdCardFont::prewarmStyle(uint8_t styleIdx, const uint32_t* codepoints, uint3
       totalBitmapSize += s.miniGlyphs[i].dataLength;
     }
 
-    s.miniBitmap = new (std::nothrow) uint8_t[totalBitmapSize > 0 ? totalBitmapSize : 1];
-    if (!s.miniBitmap) {
+    if (!ensureArrayCapacity(s.miniBitmap, s.miniBitmapCapacity, totalBitmapSize)) {
       LOG_ERR("SDCF", "Failed to allocate mini bitmap (%u bytes) for style %u", totalBitmapSize, styleIdx);
       delete[] readOrder;
       delete[] mappings;

lib/EpdFont/SdCardFont.h

@@ -163,13 +163,22 @@ class SdCardFont {
     // Stub EpdFontData returned when not prewarmed
     EpdFontData stubData{};
 
-    // Mini EpdFontData built during prewarm
+    // Mini EpdFontData built during prewarm. Buffers are kept-if-fits across pages
+    // (capacities below track allocated sizes): freeing and reallocating slightly
+    // different sizes on every page turn was a primary heap fragmenter — each page's
+    // freed hole rarely fit the next page's need, so maxAlloc eroded all session.
+    // After a few pages the capacities converge on the book's max and page turns
+    // stop allocating entirely. freeStyleMiniData() still releases everything (and
+    // zeroes capacities) for style eviction / font unload.
     EpdFontData miniData{};
     EpdUnicodeInterval* miniIntervals = nullptr;
     EpdGlyph* miniGlyphs = nullptr;
     uint8_t* miniBitmap = nullptr;
     uint32_t miniIntervalCount = 0;
     uint32_t miniGlyphCount = 0;
+    uint32_t miniIntervalCapacity = 0;
+    uint32_t miniGlyphCapacity = 0;
+    uint32_t miniBitmapCapacity = 0;
 
     // Per-page mini kern matrix (built by buildMiniKernMatrix on each full
     // prewarm). miniKernLeftClasses/miniKernRightClasses map ONLY the codepoints
@@ -184,6 +193,10 @@ class SdCardFont {
     uint8_t miniKernLeftClassCount = 0;
     uint8_t miniKernRightClassCount = 0;
     int8_t* miniKernMatrix = nullptr;
+    // Kept-if-fits capacities, same rationale as the mini glyph buffers above.
+    uint16_t miniKernLeftCapacity = 0;
+    uint16_t miniKernRightCapacity = 0;
+    uint32_t miniKernMatrixCapacity = 0;
 
     // The EpdFont whose data pointer we manage
     EpdFont epdFont{&stubData};

src/activities/reader/EpubReaderActivity.cpp

@@ -257,6 +257,27 @@ void EpubReaderActivity::openReaderMenu() {
                          });
 }
 
+bool EpubReaderActivity::buildTickHeapGate() {
+  const size_t freeHeap = ESP.getFreeHeap();
+  const size_t maxBlock = ESP.getMaxAllocHeap();
+  if (freeHeap >= BACKGROUND_BUILD_MIN_FREE_HEAP && maxBlock >= BACKGROUND_BUILD_MIN_MAX_ALLOC) {
+    return true;
+  }
+  // Below the floors. If the BLE stack is what's squeezing the heap, shed it — the
+  // established policy on this branch is that builds and resident BLE don't coexist,
+  // and this was the one build path without that protection (field crash: a tick's
+  // parse allocation aborted at maxAlloc ~11 KB with BLE resident). The lifecycle's
+  // build-pending deferral keeps BLE down until the window is caught up, then
+  // restarts it behind the start floor. Without BLE resident, just wait: page-turn
+  // transients free up between turns and the tick retries every loop pass.
+  if (BleHid.isRunning()) {
+    LOG_INF("ERS", "Background build needs heap (free=%u maxAlloc=%u); freeing BLE RAM", (unsigned)freeHeap,
+            (unsigned)maxBlock);
+    bleinput::stop();
+  }
+  return false;
+}
+
 void EpubReaderActivity::loop() {
   if (!epub) {
     // Should never happen
@@ -269,12 +290,10 @@ void EpubReaderActivity::loop() {
   // RenderLock and locked out page turns. The build follows the reader instead, and instant
   // reopen comes from suspendBuild() persisting the laid-out pages as a partial on exit.
   // Skip while the render mutex is busy so we never delay a pending render; re-check
-  // isBuilding() under the lock since render() may have just finished it. Also skip
-  // while free heap is below the floor — the tick is deferrable, and parsing into a
-  // starved heap abort()s (see BACKGROUND_BUILD_MIN_FREE_HEAP).
+  // isBuilding() under the lock since render() may have just finished it.
   if (section && section->isBuilding() && !RenderLock::peek() &&
       static_cast<int>(section->pageCount) < section->currentPage + BUILD_WINDOW_AHEAD &&
-      ESP.getFreeHeap() >= BACKGROUND_BUILD_MIN_FREE_HEAP) {
+      buildTickHeapGate()) {
     RenderLock lock;
     // Re-check under the lock: render() (which also holds the RenderLock) may have finalized the
     // build between the outer isBuilding() check and acquiring the lock here, in which case
@@ -1273,6 +1292,10 @@ void EpubReaderActivity::render(RenderLock&& lock) {
     const auto start = millis();
     renderContents(std::move(p), orientedMarginTop, orientedMarginRight, orientedMarginBottom, orientedMarginLeft);
     LOG_DBG("ERS", "Rendered page in %dms", millis() - start);
+    // Fragmentation tracker: free vs largest block after every page. A falling
+    // maxAlloc/free ratio across pages points at whichever allocation pattern the
+    // preceding lines show (mini rebuilds, kern reloads, BLE churn).
+    LOG_DBG("MEM", "post-render: free=%u maxAlloc=%u", (unsigned)ESP.getFreeHeap(), (unsigned)ESP.getMaxAllocHeap());
   }
   saveProgress(currentSpineIndex, section->currentPage, section->estimatedTotalPages());
 

src/activities/reader/EpubReaderActivity.h

@@ -90,6 +90,17 @@ class EpubReaderActivity final : public Activity {
   // page-turn transients free up between turns and the build resumes; the render
   // path still builds the page it actually needs regardless of this floor.
   static constexpr size_t BACKGROUND_BUILD_MIN_FREE_HEAP = 32 * 1024;
+  // Fragmentation floor for the same gate: a tick passed the free-heap floor at
+  // 34.7 KB free but the largest block was ~11 KB, and a parse allocation inside the
+  // tick aborted anyway. Free heap says how much memory exists; maxAlloc says whether
+  // any single allocation can actually have it. 16 KB also keeps the advance-table
+  // batch path (16 KB scratch) viable during builds.
+  static constexpr size_t BACKGROUND_BUILD_MIN_MAX_ALLOC = 16 * 1024;
+  // Gate for a background build tick: true when the heap can take parse allocations.
+  // When BLE is what's squeezing the heap, sheds it (build-pending deferral in the
+  // lifecycle then holds restarts off until the window is caught up) instead of
+  // stalling the build forever below the floors.
+  bool buildTickHeapGate();
   // How many pages to keep laid out ahead of the reader for a still-building section. A page
   // turn is ~1s on e-ink and a page builds in ~30ms, so the reader can't out-click the builder
   // -- a tiny buffer is enough. The background build stops once the watermark is this far