Merge branch 'dev' of https://gitea.maskantech.in/gitea_admin/pdf into azeem

.gitignore

@@ -79,3 +79,4 @@ timeout-*
 
 PDF Editor Timeline.xlsx
 # Local environment config / secrets
+/third_party/pdfium-wasm/

CMakeLists.txt

@@ -53,8 +53,9 @@ endif()
 
 if(PDFENGINE_WASM)
     set(PDFENGINE_BUILD_TESTS OFF CACHE BOOL "Build engine unit/smoke tests" FORCE)
-    set(PDFENGINE_WITH_PDFIUM OFF CACHE BOOL "Link the PDFium static lib" FORCE)
     set(PDFENGINE_ENABLE_SANITIZERS OFF CACHE BOOL "ASan/UBSan" FORCE)
+    # PDFium under WASM is opt-in via the wasm-pdfium preset (links the prebuilt wasm32
+    # libpdfium.a). The plain 'wasm' preset leaves PDFENGINE_WITH_PDFIUM=OFF (mock engine).
 endif()
 
 include(CompilerWarnings)

CMakePresets.json

@@ -110,6 +110,19 @@
       }
     },
 
+    {
+      "name": "wasm-pdfium",
+      "displayName": "WASM • Emscripten + PDFium (live-preview engine)",
+      "generator": "Ninja",
+      "binaryDir": "${sourceDir}/out/build/${presetName}",
+      "toolchainFile": "${sourceDir}/cmake/toolchains/vcpkg-wasm.cmake",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Release",
+        "PDFENGINE_BUILD_TESTS": "OFF",
+        "PDFENGINE_WITH_PDFIUM": "ON"
+      }
+    },
+
     {
       "name": "fuzz-linux",
       "displayName": "Linux • libFuzzer (Clang + ASan)",
@@ -143,6 +156,7 @@
     { "name": "macos-release",   "configurePreset": "macos-release" },
     { "name": "macos-asan",      "configurePreset": "macos-asan" },
     { "name": "wasm",            "configurePreset": "wasm" },
+    { "name": "wasm-pdfium",     "configurePreset": "wasm-pdfium" },
     { "name": "fuzz-linux",       "configurePreset": "fuzz-linux" },
     { "name": "fuzz-linux-nosan", "configurePreset": "fuzz-linux-nosan" }
   ],

cmake/pdfium.cmake

@@ -7,18 +7,39 @@ if(NOT PDFENGINE_WITH_PDFIUM)
     return()
 endif()
 
-set(PDFIUM_INSTALL_DIR "${CMAKE_SOURCE_DIR}/third_party/pdfium/install"
-    CACHE PATH "Root of the PDFium install tree produced by build_pdfium.*")
+# Desktop uses the GN-built tree; WASM uses the prebuilt Emscripten static lib
+# (third_party/pdfium-wasm, fetched by get_pdfium_wasm.ps1 — a relinkable libpdfium.a).
+if(EMSCRIPTEN)
+    set(PDFIUM_INSTALL_DIR "${CMAKE_SOURCE_DIR}/third_party/pdfium-wasm"
+        CACHE PATH "Root of the wasm32 PDFium tree (prebuilt libpdfium.a + include)")
+else()
+    set(PDFIUM_INSTALL_DIR "${CMAKE_SOURCE_DIR}/third_party/pdfium/install"
+        CACHE PATH "Root of the PDFium install tree produced by build_pdfium.*")
+endif()
 
-find_path(PDFIUM_INCLUDE_DIR
-    NAMES fpdfview.h
-    PATHS "${PDFIUM_INSTALL_DIR}/include"
-    NO_DEFAULT_PATH)
+if(EMSCRIPTEN)
+    # Emscripten's toolchain sets CMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY, which makes
+    # find_library ignore custom PATHS (it only searches the emsdk sysroot). Set the
+    # prebuilt wasm32 lib + headers directly instead.
+    set(PDFIUM_INCLUDE_DIR "${PDFIUM_INSTALL_DIR}/include")
+    set(PDFIUM_LIBRARY "${PDFIUM_INSTALL_DIR}/lib/libpdfium.a")
+    if(NOT EXISTS "${PDFIUM_INCLUDE_DIR}/fpdfview.h")
+        set(PDFIUM_INCLUDE_DIR "")
+    endif()
+    if(NOT EXISTS "${PDFIUM_LIBRARY}")
+        set(PDFIUM_LIBRARY "")
+    endif()
+else()
+    find_path(PDFIUM_INCLUDE_DIR
+        NAMES fpdfview.h
+        PATHS "${PDFIUM_INSTALL_DIR}/include"
+        NO_DEFAULT_PATH)
 
-find_library(PDFIUM_LIBRARY
-    NAMES pdfium libpdfium
-    PATHS "${PDFIUM_INSTALL_DIR}/lib"
-    NO_DEFAULT_PATH)
+    find_library(PDFIUM_LIBRARY
+        NAMES pdfium libpdfium
+        PATHS "${PDFIUM_INSTALL_DIR}/lib"
+        NO_DEFAULT_PATH)
+endif()
 
 if(NOT PDFIUM_INCLUDE_DIR OR NOT PDFIUM_LIBRARY)
     message(FATAL_ERROR

engine/include/pdfengine/pdf_document.hpp

@@ -258,6 +258,10 @@ public:
 
     virtual std::expected<void, EngineError> applyEdits(const std::string& editsJson) = 0;
 
+    // Per-character caret layout (JSON, PDF coords) of the most recent reflow_paragraph op, or
+    // empty if none. Lets the WASM live preview align a caret exactly to the rendered glyphs.
+    [[nodiscard]] virtual std::string lastReflowLayout() const { return {}; }
+
     [[nodiscard]] virtual std::expected<std::vector<uint8_t>, EngineError> 
     saveIncremental() const = 0;
 

engine/src/fonts/pdf_fonts/font_subset.cpp

@@ -81,6 +81,40 @@ std::vector<uint8_t> FontSubset::buildSubset(const std::vector<uint8_t>& origina
     return result;
 }
 
+// Overwrite a few bytes of the sfnt 'name' table strings (family/full/PostScript/typographic
+// names) with a deterministic seed so the subset has a UNIQUE name. Two reflow subsets of the
+// same base font (e.g. Calibri) otherwise share the name "Calibri", and PDFium merges them —
+// making the second subset's text map to the first's glyphs (scrambled output). Identical inputs
+// hash to the same seed → same name → safe to share. The name is only used for de-dup; rendering
+// uses the embedded glyphs + cmap, so a "garbled" name is harmless.
+static void makeFontNameUnique(std::vector<uint8_t>& font, uint64_t seed) {
+    auto rd16 = [&](size_t o) -> uint32_t { return (o + 1 < font.size()) ? (uint32_t(font[o]) << 8) | font[o + 1] : 0; };
+    auto rd32 = [&](size_t o) -> uint32_t { return (o + 3 < font.size()) ? (uint32_t(font[o]) << 24) | (uint32_t(font[o + 1]) << 16) | (uint32_t(font[o + 2]) << 8) | font[o + 3] : 0; };
+    if (font.size() < 12) return;
+    uint32_t numTables = rd16(4);
+    size_t nameOff = 0;
+    for (uint32_t i = 0; i < numTables; ++i) {
+        size_t rec = 12 + size_t(i) * 16;
+        if (rec + 16 > font.size()) break;
+        if (font[rec] == 'n' && font[rec + 1] == 'a' && font[rec + 2] == 'm' && font[rec + 3] == 'e') { nameOff = rd32(rec + 8); break; }
+    }
+    if (nameOff == 0 || nameOff + 6 > font.size()) return;
+    uint32_t count = rd16(nameOff + 2);
+    size_t storageOff = nameOff + rd16(nameOff + 4);
+    for (uint32_t i = 0; i < count; ++i) {
+        size_t rec = nameOff + 6 + size_t(i) * 12;
+        if (rec + 12 > font.size()) break;
+        uint32_t nameID = rd16(rec + 6);
+        if (nameID == 1 || nameID == 4 || nameID == 6 || nameID == 16) {
+            uint32_t len = rd16(rec + 8);
+            size_t s = storageOff + rd16(rec + 10);
+            for (uint32_t b = 0; b < len && b < 8; ++b) {
+                if (s + b < font.size()) font[s + b] = static_cast<uint8_t>('A' + ((seed >> ((b % 8) * 4)) & 0x0F)); // 'A'..'P'
+            }
+        }
+    }
+}
+
 std::vector<uint8_t> FontSubset::buildSubsetByUnicode(const std::vector<uint8_t>& originalStream,
                                                       const std::vector<uint32_t>& codepoints) {
     // Subset a font down to only the glyphs needed for `codepoints`, for embedding

engine/src/parser/pdfium_document.cpp

@@ -17,6 +17,7 @@
 #include "fonts/pdf_fonts/font_fallback.hpp"
 #include "fonts/pdf_fonts/font_subset.hpp"
 #include "fonts/shaping/hb_shaper.hpp"
+#include "fonts/face/font_face.hpp"
 #include "decoration_builder.hpp"
 
 #include <nlohmann/json.hpp>
@@ -117,6 +118,93 @@ std::string code_point_to_utf8(unsigned int cp) {
     }
     return utf8;
 }
+
+// Reverse of the Windows-1252 (cp1252) byte→codepoint map: a unicode codepoint back to its
+// single cp1252 byte, or -1 if it isn't representable in cp1252. Used to undo "double-encoded"
+// ToUnicode maps (see repairMojibake).
+static int cp1252ToByte(unsigned int cp) {
+    if (cp <= 0x7F) return static_cast<int>(cp);
+    if (cp >= 0xA0 && cp <= 0xFF) return static_cast<int>(cp); // Latin-1 high range is identity
+    switch (cp) {
+        case 0x20AC: return 0x80; case 0x201A: return 0x82; case 0x0192: return 0x83;
+        case 0x201E: return 0x84; case 0x2026: return 0x85; case 0x2020: return 0x86;
+        case 0x2021: return 0x87; case 0x02C6: return 0x88; case 0x2030: return 0x89;
+        case 0x0160: return 0x8A; case 0x2039: return 0x8B; case 0x0152: return 0x8C;
+        case 0x017D: return 0x8E; case 0x2018: return 0x91; case 0x2019: return 0x92;
+        case 0x201C: return 0x93; case 0x201D: return 0x94; case 0x2022: return 0x95;
+        case 0x2013: return 0x96; case 0x2014: return 0x97; case 0x02DC: return 0x98;
+        case 0x2122: return 0x99; case 0x0161: return 0x9A; case 0x203A: return 0x9B;
+        case 0x0153: return 0x9C; case 0x017E: return 0x9E; case 0x0178: return 0x9F;
+        default: return -1;
+    }
+}
+
+// Decode a UTF-8 string into codepoints.
+static std::vector<unsigned int> utf8_to_codepoints(const std::string& s) {
+    std::vector<unsigned int> cps;
+    for (size_t i = 0; i < s.size();) {
+        unsigned char c = static_cast<unsigned char>(s[i]);
+        unsigned int cp = c; int extra = 0;
+        if (c < 0x80) { cp = c; extra = 0; }
+        else if ((c & 0xE0) == 0xC0) { cp = c & 0x1F; extra = 1; }
+        else if ((c & 0xF0) == 0xE0) { cp = c & 0x0F; extra = 2; }
+        else if ((c & 0xF8) == 0xF0) { cp = c & 0x07; extra = 3; }
+        else { cp = c; extra = 0; }
+        if (i + extra >= s.size()) { cps.push_back(c); i++; continue; }
+        bool ok = true;
+        for (int j = 1; j <= extra; ++j) { unsigned char n = static_cast<unsigned char>(s[i + j]); if ((n & 0xC0) != 0x80) { ok = false; break; } cp = (cp << 6) | (n & 0x3F); }
+        if (!ok) { cps.push_back(c); i++; continue; }
+        cps.push_back(cp); i += extra + 1;
+    }
+    return cps;
+}
+
+// Repair "double-encoded" text. Some PDFs ship a ToUnicode CMap whose multi-byte UTF-8 sequences
+// were themselves decoded through cp1252, so a single char like the en-dash (U+2013, UTF-8 bytes
+// E2 80 93) is extracted as three garbage chars (â € " = U+00E2 U+20AC U+201C). We map each char
+// back to its cp1252 byte and, where those bytes form a valid UTF-8 lead+continuation sequence,
+// re-decode them to the intended codepoint. Plain text is left untouched.
+static std::string repairMojibake(const std::string& s) {
+    auto cps = utf8_to_codepoints(s);
+    std::string out;
+    auto emit = [&](unsigned int cp) { out += code_point_to_utf8(cp); };
+    for (size_t i = 0; i < cps.size();) {
+        int b0 = cp1252ToByte(cps[i]);
+        int need = 0;
+        if (b0 >= 0xC0 && b0 <= 0xDF) need = 1;
+        else if (b0 >= 0xE0 && b0 <= 0xEF) need = 2;
+        else if (b0 >= 0xF0 && b0 <= 0xF7) need = 3;
+        if (need > 0 && i + need < cps.size()) {
+            std::string bytes; bytes.push_back(static_cast<char>(b0));
+            bool ok = true;
+            for (int j = 1; j <= need; ++j) { int b = cp1252ToByte(cps[i + j]); if (b < 0x80 || b > 0xBF) { ok = false; break; } bytes.push_back(static_cast<char>(b)); }
+            if (ok) { auto rd = utf8_to_codepoints(bytes); if (rd.size() == 1 && rd[0] > 0x7F) { emit(rd[0]); i += need + 1; continue; } }
+        }
+        emit(cps[i]); i++;
+    }
+    return out;
+}
+
+// Normalize "smart" punctuation to ASCII for reflow re-emission. Office subset fonts often map
+// en/em dashes, curly quotes, and ellipses at custom codes (not their Unicode), so re-emitting
+// them via FPDFText_SetText(unicode) drops them to a notdef glyph (the en-dash came out as "B").
+// Substituting the ASCII equivalent renders reliably in every font (embedded, base-14, or system)
+// and keeps preview == saved. Only reflowed text is affected; untouched text keeps its originals.
+std::string asciiizePunctuation(const std::string& s) {
+    auto cps = utf8_to_codepoints(s);
+    std::string out;
+    for (unsigned int cp : cps) {
+        switch (cp) {
+            case 0x2010: case 0x2011: case 0x2012: case 0x2013: case 0x2014: case 0x2015: out += '-'; break;
+            case 0x2018: case 0x2019: case 0x201B: out += '\''; break;
+            case 0x201C: case 0x201D: case 0x201F: out += '"'; break;
+            case 0x2026: out += "..."; break;
+            case 0x00A0: case 0x2002: case 0x2003: case 0x2009: case 0x202F: out += ' '; break;
+            default: out += code_point_to_utf8(cp); break;
+        }
+    }
+    return out;
+}
 } // namespace pdfengine::parser
 
 namespace {
@@ -1024,10 +1112,15 @@ std::expected<PageModel, EngineError> PdfiumPage::extractDocumentModel() const {
                 gaps.push_back(gap);
             }
         }
-        double medianGap = 0.0;
+        // Robust intra-word gap estimate. Use a LOW percentile (25th), NOT the median: on lines
+        // with many short words (or justified text), positive gaps are dominated by word-spaces,
+        // which inflates the median and makes the space threshold below swallow real spaces —
+        // gluing whole words together ("backend systems for" -> "backendsystemsfor"). The 25th
+        // percentile stays inside the small intra-word gap cluster even when spaces dominate.
+        double p25Gap = 0.0;
         if (!gaps.empty()) {
             std::sort(gaps.begin(), gaps.end());
-            medianGap = gaps[gaps.size() / 2];
+            p25Gap = gaps[gaps.size() / 4];
         }
 
         // Phase 5F: Run Builder
@@ -1050,8 +1143,15 @@ std::expected<PageModel, EngineError> PdfiumPage::extractDocumentModel() const {
                 const auto& currG = line.glyphs[i];
                 
                 double gap = currG.bboxX - (prevG.bboxX + prevG.bboxW);
-                double spaceThreshold = (std::max)(currG.fontSize * 0.25, medianGap * 2.0);
-                
+                // A word space is a clear outlier above intra-word tracking. Anchor on a
+                // fontSize-relative floor (reliable for normal AND justified body text), and only
+                // raise it for unusually loose tracking — capped so it can never grow large
+                // enough to drop real word-spaces (the previous medianGap*2.0 bug).
+                double spaceThreshold = currG.fontSize * 0.2;
+                if (p25Gap > spaceThreshold) {
+                    spaceThreshold = (std::min)(p25Gap * 1.5, currG.fontSize * 0.38);
+                }
+
                 bool addSpace = gap > spaceThreshold && prevG.text != " " && currG.text != " ";
                 bool breakRun = currG.fontName != currentRun.fontName || 
                                 std::abs(currG.fontSize - currentRun.fontSize) > 0.1 ||
@@ -1123,6 +1223,15 @@ std::expected<PageModel, EngineError> PdfiumPage::extractDocumentModel() const {
         }
     }
 
+    // Phase 5F.5: Repair double-encoded text (e.g. an en-dash extracted as "â€"" from a malformed
+    // ToUnicode CMap) so both display and reflow re-emission use the correct codepoints.
+    for (auto& line : lines) {
+        for (auto& run : line.runs) {
+            std::string repaired = pdfengine::parser::repairMojibake(run.text);
+            if (repaired != run.text) run.text = std::move(repaired);
+        }
+    }
+
     // Phase 5G: Paragraph Builder
     std::vector<Paragraph> paragraphs;
     if (!lines.empty()) {
@@ -1765,12 +1874,155 @@ std::expected<std::shared_ptr<PdfPage>, EngineError> PdfiumDocument::getPage(int
 #endif
 }
 
+#ifdef PDFENGINE_WITH_PDFIUM
+PdfiumDocument::EmissionFont PdfiumDocument::loadEmissionFont(
+    int pageIndex, const std::string& internalFontId, double fontSize,
+    const std::vector<uint32_t>& codepoints) {
+    EmissionFont out;
+    (void)fontSize;
+
+    // Derive a base-14 name + bold/italic from the internalFontId (mirrors replace_text).
+    std::string fontName = "Helvetica";
+    bool bold = false, italic = false;
+    {
+        std::string lowerId = internalFontId;
+        std::transform(lowerId.begin(), lowerId.end(), lowerId.begin(),
+                       [](unsigned char ch) { return static_cast<char>(std::tolower(ch)); });
+        if (lowerId.find("bolditalic") != std::string::npos) { bold = true; italic = true; }
+        else if (lowerId.find("bold") != std::string::npos) { bold = true; }
+        else if (lowerId.find("italic") != std::string::npos) { italic = true; }
+        else if (lowerId.find("oblique") != std::string::npos) { italic = true; }
+        if (lowerId.find("times") != std::string::npos) {
+            fontName = bold && italic ? "Times-BoldItalic" : bold ? "Times-Bold" : italic ? "Times-Italic" : "Times-Roman";
+        } else if (lowerId.find("courier") != std::string::npos) {
+            fontName = bold && italic ? "Courier-BoldOblique" : bold ? "Courier-Bold" : italic ? "Courier-Oblique" : "Courier";
+        } else {
+            fontName = bold && italic ? "Helvetica-BoldOblique" : bold ? "Helvetica-Bold" : italic ? "Helvetica-Oblique" : "Helvetica";
+        }
+    }
+
+    // Match the FontInfo by internalFontId + resolve a FontFace (for shaping/width).
+    std::optional<FontInfo> matchedFontInfo;
+    auto fontsRes = getFonts(pageIndex, pageIndex);
+    if (fontsRes.has_value()) {
+        for (const auto& fi : *fontsRes) {
+            if (!internalFontId.empty() && fi.internalFontId == internalFontId) { matchedFontInfo = fi; break; }
+        }
+    }
+    if (matchedFontInfo.has_value()) {
+        auto r = getResolvedFont(*matchedFontInfo);
+        if (r.has_value()) out.resolved = *r;
+    }
+
+    // Glyph coverage → pick the embedding tier.
+    bool fontSupportsAll = true, subsetLacksGlyphs = false;
+    bool isSubsetFont = matchedFontInfo && matchedFontInfo->isSubset;
+    if (out.resolved) {
+        for (uint32_t cp : codepoints) {
+            if (!out.resolved->hasGlyph(cp)) {
+                if (isSubsetFont) subsetLacksGlyphs = true; else fontSupportsAll = false;
+            }
+        }
+    } else {
+        fontSupportsAll = false;
+    }
+
+    std::string cacheKey;
+    bool useEmbedded = false, useSystem = false;
+    if (matchedFontInfo && matchedFontInfo->isEmbedded) {
+        // Re-embed from the PDF's OWN font bytes (getFontData below). This path does NOT need the
+        // resolver's FontFace, so it works under WASM's no-filesystem build where getResolvedFont
+        // can return null — the reflowed text then keeps the document's real font (e.g. Calibri)
+        // instead of falling back to a generic Helvetica/Liberation. We only divert to a system
+        // substitute when we have a resolved face that PROVES the subset lacks a needed glyph.
+        bool subsetProvenLacking = out.resolved && isSubsetFont && subsetLacksGlyphs;
+        bool fullProvenLacking = out.resolved && !isSubsetFont && !fontSupportsAll;
+        if (!subsetProvenLacking && !fullProvenLacking) { cacheKey = matchedFontInfo->internalFontId; useEmbedded = true; }
+        else { cacheKey = "system_embed_" + matchedFontInfo->fontName + std::string("_") + (bold ? "B" : "") + (italic ? "I" : ""); useSystem = true; }
+    } else if (matchedFontInfo && !matchedFontInfo->isEmbedded) {
+        cacheKey = "standard_" + fontName;
+    } else {
+        cacheKey = "standard_" + fontName;
+    }
+
+    // Make the cache key CODEPOINT-AWARE. A subset is built for an exact glyph set, so a font
+    // reused with different characters (e.g. the Summary's text cached, then a bullet reflowed)
+    // must NOT reuse the prior subset — its cmap would map the new chars to the wrong glyphs,
+    // producing scrambled text ("Core Java" -> "Cuwi Le,e"). Hashing the codepoints fixes it.
+    {
+        uint64_t h = 1469598103934665603ull;
+        for (uint32_t cp : codepoints) { h ^= cp; h *= 1099511628211ull; }
+        cacheKey += "#" + std::to_string(h);
+    }
+
+    {
+        std::lock_guard<std::mutex> lock(loadedFontsMutex_);
+        if (loadedFontsCache_.count(cacheKey)) {
+            out.font = loadedFontsCache_[cacheKey];
+            auto it = loadedMeasureFaces_.find(cacheKey);
+            if (it != loadedMeasureFaces_.end()) out.measureFace = it->second;
+        }
+    }
+    if (out.font) return out;
+
+    if (useEmbedded) {
+        auto fontDataRes = getFontData(matchedFontInfo->internalFontId);
+        if (fontDataRes.has_value() && !fontDataRes.value().empty()) {
+            // Re-subset the embedded font over the glyphs we'll emit. HarfBuzz rebuilds a
+            // clean unicode cmap (no RETAIN_GIDS), so FPDFText_SetText(unicode) maps every
+            // char — loading the raw embedded bytes cid=false drops glyphs whose original
+            // cmap subtable PDFium can't follow (e.g. Office's embedded Calibri). Smart
+            // punctuation is normalized to ASCII upstream (asciiizePunctuation), so the rare
+            // "glyph at a custom code" case (the en-dash) never reaches here.
+            auto subset = fonts::pdf_fonts::FontSubset::buildSubsetByUnicode(fontDataRes.value(), codepoints);
+            std::lock_guard<std::mutex> lock(loadedFontsMutex_);
+            loadedFontDataBuffers_[cacheKey] = !subset.empty() ? std::move(subset) : fontDataRes.value();
+            const auto& bytes = loadedFontDataBuffers_[cacheKey];
+            out.font = FPDFText_LoadFont(doc_, bytes.data(), static_cast<uint32_t>(bytes.size()), FPDF_FONT_TRUETYPE, true);
+        }
+    } else if (useSystem) {
+        std::string fontPath = fonts::pdf_fonts::FontFallback::getInstance().getFallbackFontPath(
+            matchedFontInfo->normalizedFamily.empty() ? matchedFontInfo->fontName : matchedFontInfo->normalizedFamily, bold, italic);
+        std::ifstream fs(fontPath, std::ios::binary);
+        if (fs) {
+            std::vector<uint8_t> fileBytes((std::istreambuf_iterator<char>(fs)), std::istreambuf_iterator<char>());
+            if (!fileBytes.empty()) {
+                std::vector<uint8_t> subsetBytes = fonts::pdf_fonts::FontSubset::buildSubsetByUnicode(fileBytes, codepoints);
+                std::lock_guard<std::mutex> lock(loadedFontsMutex_);
+                loadedFontDataBuffers_[cacheKey] = !subsetBytes.empty() ? std::move(subsetBytes) : std::move(fileBytes);
+                const auto& bytes = loadedFontDataBuffers_[cacheKey];
+                out.font = FPDFText_LoadFont(doc_, bytes.data(), static_cast<uint32_t>(bytes.size()), FPDF_FONT_TRUETYPE, true);
+            }
+        }
+    }
+    if (!out.font) {
+        out.font = FPDFText_LoadStandardFont(doc_, fontName.c_str());
+        if (!out.font) out.font = FPDFText_LoadStandardFont(doc_, "Helvetica");
+    }
+    if (out.font) {
+        std::lock_guard<std::mutex> lock(loadedFontsMutex_);
+        loadedFontsCache_[cacheKey] = out.font;
+        // Build a measuring face from the exact bytes we loaded (embedded/system paths) so
+        // width measurement matches what PDFium renders. Standard base-14 fonts have no bytes
+        // here → measurement falls back to the resolver face.
+        auto bit = loadedFontDataBuffers_.find(cacheKey);
+        if (bit != loadedFontDataBuffers_.end() && !bit->second.empty()) {
+            auto mf = std::make_shared<fonts::FontFace>();
+            if (mf->loadFromMemory(bit->second)) { loadedMeasureFaces_[cacheKey] = mf; out.measureFace = mf; }
+        }
+    }
+    return out;
+}
+#endif
+
 std::expected<void, EngineError> PdfiumDocument::applyEdits(const std::string& editsJson) {
 #ifdef PDFENGINE_WITH_PDFIUM
     if (!doc_) {
         return std::unexpected(EngineError::Unknown);
     }
 
+    lastReflowLayout_.clear(); // repopulated only if a reflow_paragraph op runs
+
     try {
         auto root = nlohmann::json::parse(editsJson);
         if (!root.contains("version") || root["version"] != "1.0") {
@@ -2227,6 +2479,320 @@ std::expected<void, EngineError> PdfiumDocument::applyEdits(const std::string& e
                     spdlog::error("Failed to generate page content after replace_text");
                 }
 
+                FPDF_ClosePage(page);
+            } else if (type == "reflow_paragraph") {
+                // Word/Adobe-style paragraph reflow: re-wrap the paragraph's styled runs
+                // within its column width, re-justify, emit the new lines, and push the
+                // following in-column content down/up by the line-count delta.
+                if (!op.contains("data") || !op["data"].is_object()) {
+                    spdlog::error("reflow_paragraph missing 'data'");
+                    return std::unexpected(EngineError::InvalidFormat);
+                }
+                auto data = op["data"];
+
+                struct RunStyle { std::string text; std::string internalFontId; double fontSize; unsigned int r, g, b; };
+                std::vector<RunStyle> runs;
+                auto parseHex = [](const std::string& hex, unsigned int& r, unsigned int& g, unsigned int& b) {
+                    r = 0; g = 0; b = 0;
+                    if (hex.size() >= 7 && hex[0] == '#') {
+                        auto hv = [](char ch) -> int {
+                            if (ch >= '0' && ch <= '9') return ch - '0';
+                            if (ch >= 'a' && ch <= 'f') return ch - 'a' + 10;
+                            if (ch >= 'A' && ch <= 'F') return ch - 'A' + 10;
+                            return 0;
+                        };
+                        r = static_cast<unsigned int>(hv(hex[1]) * 16 + hv(hex[2]));
+                        g = static_cast<unsigned int>(hv(hex[3]) * 16 + hv(hex[4]));
+                        b = static_cast<unsigned int>(hv(hex[5]) * 16 + hv(hex[6]));
+                    }
+                };
+                auto parseRun = [&](const nlohmann::json& rj) {
+                    RunStyle rs;
+                    rs.text = pdfengine::parser::asciiizePunctuation(rj.value("text", ""));
+                    rs.internalFontId = rj.value("internalFontId", "");
+                    rs.fontSize = rj.value("fontSize", 12.0);
+                    parseHex(rj.value("color", std::string("#000000")), rs.r, rs.g, rs.b);
+                    runs.push_back(std::move(rs));
+                };
+                // Optional WYSIWYG mode: the frontend provides the exact visual line breaks
+                // (one inner array of styled fragments per line). When present we emit those
+                // breaks verbatim and skip our greedy line-breaker, so the saved page matches
+                // the live preview. Otherwise we break the flat `runs` ourselves.
+                std::vector<std::vector<int>> providedLines; // run indices per provided line
+                bool hasProvidedLines = data.contains("lines") && data["lines"].is_array() && !data["lines"].empty();
+                if (hasProvidedLines) {
+                    for (const auto& lineJson : data["lines"]) {
+                        std::vector<int> lineRunIdxs;
+                        if (lineJson.is_array()) {
+                            for (const auto& rj : lineJson) { lineRunIdxs.push_back(static_cast<int>(runs.size())); parseRun(rj); }
+                        }
+                        providedLines.push_back(std::move(lineRunIdxs));
+                    }
+                } else if (data.contains("runs") && data["runs"].is_array()) {
+                    for (const auto& rj : data["runs"]) parseRun(rj);
+                }
+                std::vector<int> objectIndices;
+                if (data.contains("objectIndices") && data["objectIndices"].is_array()) {
+                    for (auto& idx : data["objectIndices"]) objectIndices.push_back(idx.get<int>());
+                }
+                double columnLeft = data.value("columnLeft", 0.0);
+                double columnRight = data.value("columnRight", 0.0);
+                double firstBaselineY = data.value("firstBaselineY", 0.0);
+                double leading = data.value("leading", 0.0);
+                int oldLineCount = data.value("oldLineCount", 1);
+                std::string align = data.value("align", std::string("left"));
+                double columnWidth = columnRight - columnLeft;
+                // For a bullet/list item the TEXT reflows from a hanging indent (columnLeft), but
+                // the push-down of content below must span the FULL block width so the bullet
+                // markers (which sit left of the text indent) and other items move together.
+                // Defaults to columnLeft for ordinary paragraphs.
+                double pushColumnLeft = data.value("pushColumnLeft", columnLeft);
+
+                if (runs.empty() || objectIndices.empty() || columnWidth <= 1.0 || leading <= 0.0) {
+                    spdlog::warn("reflow_paragraph: insufficient layout data, skipping");
+                    continue;
+                }
+
+                FPDF_PAGE page = FPDF_LoadPage(doc_, pageIndex);
+                if (!page) {
+                    spdlog::error("Failed to load page {} for reflow_paragraph", pageIndex);
+                    return std::unexpected(EngineError::Unknown);
+                }
+
+                // Resolve + load each run's emission font (subset over that run's codepoints).
+                std::vector<EmissionFont> runFonts(runs.size());
+                auto toCodepoints = [](const std::string& s) {
+                    auto u16 = utf8_to_utf16le(s);
+                    std::vector<uint32_t> cps;
+                    for (size_t i = 0; i < u16.size();) {
+                        uint32_t cp = u16[i];
+                        if (cp >= 0xD800 && cp <= 0xDBFF && i + 1 < u16.size()) {
+                            uint32_t low = u16[i + 1];
+                            if (low >= 0xDC00 && low <= 0xDFFF) { cp = 0x10000 + ((cp - 0xD800) << 10) + (low - 0xDC00); i += 2; }
+                            else i += 1;
+                        } else i += 1;
+                        cps.push_back(cp);
+                    }
+                    return cps;
+                };
+                // Load each unique font ONCE, subset over the UNION of all its runs' glyphs
+                // (a font shared by several runs must cover every char it will emit).
+                {
+                    std::unordered_map<std::string, std::vector<uint32_t>> fontCps;
+                    for (const auto& rs : runs) {
+                        auto cps = toCodepoints(rs.text);
+                        auto& dst = fontCps[rs.internalFontId];
+                        dst.insert(dst.end(), cps.begin(), cps.end());
+                    }
+                    std::unordered_map<std::string, EmissionFont> fontByFid;
+                    for (const auto& rs : runs) {
+                        if (!fontByFid.count(rs.internalFontId))
+                            fontByFid[rs.internalFontId] = loadEmissionFont(pageIndex, rs.internalFontId, rs.fontSize, fontCps[rs.internalFontId]);
+                    }
+                    for (size_t ri = 0; ri < runs.size(); ++ri) runFonts[ri] = fontByFid[runs[ri].internalFontId];
+                }
+
+                // Measure a substring's width in a run's resolved font. HbShaper takes an
+                // INTEGER pixel size, so shape at a large reference size and scale to the real
+                // (fractional) font size — otherwise 9.96px rounds to 9px, every word measures
+                // ~10% narrow, and PDFium then renders them wider, collapsing the spaces.
+                fonts::HbShaper shaper;
+                constexpr unsigned int kRefSize = 1000;
+                auto measure = [&](size_t runIdx, const std::string& text) -> double {
+                    if (text.empty()) return 0.0;
+                    auto& rf = runFonts[runIdx];
+                    double size = runs[runIdx].fontSize > 0 ? runs[runIdx].fontSize : 12.0;
+                    // Prefer the face built from the EMITTED bytes (exact width parity); fall
+                    // back to the resolver face (base-14 standard fonts have no emitted bytes).
+                    fonts::FontFace* face = rf.measureFace ? rf.measureFace.get()
+                                          : (rf.resolved ? &rf.resolved->getFontFace() : nullptr);
+                    if (face) {
+                        auto glyphs = shaper.shapeRun(text, *face, kRefSize);
+                        if (!glyphs.empty()) {
+                            double w = 0.0; for (auto& gph : glyphs) w += gph.advanceX;
+                            return w * (size / static_cast<double>(kRefSize));
+                        }
+                    }
+                    return size * 0.5 * static_cast<double>(text.size()); // rough fallback
+                };
+
+                // Per-character advances (PDF units) for a run substring — used to build the
+                // caret layout so the live preview can place a caret exactly on each glyph. For
+                // ASCII (one glyph per byte) this is exact; otherwise the width is spread evenly.
+                auto perCharAdvances = [&](size_t runIdx, const std::string& text) -> std::vector<double> {
+                    std::vector<double> out(text.size(), 0.0);
+                    if (text.empty()) return out;
+                    auto& rf = runFonts[runIdx];
+                    double size = runs[runIdx].fontSize > 0 ? runs[runIdx].fontSize : 12.0;
+                    double scale = size / static_cast<double>(kRefSize);
+                    fonts::FontFace* face = rf.measureFace ? rf.measureFace.get()
+                                          : (rf.resolved ? &rf.resolved->getFontFace() : nullptr);
+                    if (face) {
+                        auto glyphs = shaper.shapeRun(text, *face, kRefSize);
+                        if (glyphs.size() == text.size()) {
+                            for (size_t i = 0; i < text.size(); ++i) out[i] = glyphs[i].advanceX * scale;
+                            return out;
+                        }
+                        double w = 0.0; for (auto& gph : glyphs) w += gph.advanceX;
+                        double per = w * scale / static_cast<double>(text.size());
+                        for (auto& v : out) v = per;
+                        return out;
+                    }
+                    for (auto& v : out) v = size * 0.5;
+                    return out;
+                };
+
+                auto isSpace = [](char ch) { return ch == ' ' || ch == '\t' || ch == '\n' || ch == '\r'; };
+                struct Seg { int runIdx; std::string text; double width; };
+                struct Word { std::vector<Seg> segs; double width; double spaceAfter; };
+                std::vector<Word> words;
+
+                // Tokenize a sequence of runs into words (non-space runs split into same-style
+                // segments), append to `words`, return the new word indices. Space width is
+                // measured in the word's own font (the model emits zero-size empty-fid space runs).
+                auto tokenize = [&](const std::vector<int>& runIdxs) -> std::vector<size_t> {
+                    std::string ft; std::vector<int> sof;
+                    for (int ri : runIdxs) for (char ch : runs[ri].text) { ft.push_back(ch); sof.push_back(ri); }
+                    std::vector<size_t> out;
+                    size_t i = 0;
+                    while (i < ft.size()) {
+                        if (isSpace(ft[i])) { i++; continue; }
+                        Word w; w.width = 0.0; w.spaceAfter = 0.0;
+                        while (i < ft.size() && !isSpace(ft[i])) {
+                            int st = sof[i]; std::string frag;
+                            while (i < ft.size() && !isSpace(ft[i]) && sof[i] == st) { frag.push_back(ft[i]); i++; }
+                            double fw = measure(static_cast<size_t>(st), frag);
+                            w.segs.push_back({st, frag, fw}); w.width += fw;
+                        }
+                        if (i < ft.size() && isSpace(ft[i])) {
+                            int styleForSpace = w.segs.empty() ? sof[i] : w.segs.back().runIdx;
+                            w.spaceAfter = measure(static_cast<size_t>(styleForSpace), " ");
+                        }
+                        out.push_back(words.size());
+                        words.push_back(std::move(w));
+                    }
+                    return out;
+                };
+
+                std::vector<std::vector<size_t>> lines;
+                if (hasProvidedLines) {
+                    // WYSIWYG: emit the frontend's exact visual line breaks.
+                    for (const auto& lineRunIdxs : providedLines) {
+                        auto wi = tokenize(lineRunIdxs);
+                        if (!wi.empty()) lines.push_back(std::move(wi));
+                    }
+                } else {
+                    // Greedy-break the whole paragraph within the column.
+                    std::vector<int> allRuns(runs.size());
+                    for (size_t ri = 0; ri < runs.size(); ++ri) allRuns[ri] = static_cast<int>(ri);
+                    auto allWords = tokenize(allRuns);
+                    std::vector<size_t> cur; double curW = 0.0;
+                    for (size_t k = 0; k < allWords.size(); ++k) {
+                        size_t wi = allWords[k];
+                        double gap = cur.empty() ? 0.0 : words[allWords[k - 1]].spaceAfter;
+                        if (!cur.empty() && curW + gap + words[wi].width > columnWidth) {
+                            lines.push_back(cur); cur.clear();
+                            cur.push_back(wi); curW = words[wi].width;
+                        } else {
+                            cur.push_back(wi); curW += gap + words[wi].width;
+                        }
+                    }
+                    if (!cur.empty()) lines.push_back(cur);
+                }
+                if (words.empty() || lines.empty()) { FPDF_ClosePage(page); continue; }
+                int newLineCount = static_cast<int>(lines.size());
+
+                // Push following in-column content down/up by the height delta (do this BEFORE
+                // deleting the old objects so indices stay valid and new objects aren't moved).
+                double deltaH = (newLineCount - oldLineCount) * leading;
+                double paragraphBottomBaseline = firstBaselineY - (oldLineCount - 1) * leading;
+                if (std::abs(deltaH) > 0.01) {
+                    double threshold = paragraphBottomBaseline - 0.5 * leading;
+                    int nObjs = FPDFPage_CountObjects(page);
+                    int pushed = 0;
+                    for (int k = 0; k < nObjs; ++k) {
+                        if (std::find(objectIndices.begin(), objectIndices.end(), k) != objectIndices.end()) continue;
+                        FPDF_PAGEOBJECT o = FPDFPage_GetObject(page, k);
+                        if (!o) continue;
+                        float l = 0, bo = 0, rr = 0, tt = 0;
+                        if (!FPDFPageObj_GetBounds(o, &l, &bo, &rr, &tt)) continue;
+                        double cY = (bo + tt) / 2.0;
+                        if (cY < threshold && rr > pushColumnLeft && l < columnRight) {
+                            FPDFPageObj_Transform(o, 1.0, 0.0, 0.0, 1.0, 0.0, -deltaH);
+                            pushed++;
+                        }
+                    }
+                    spdlog::info("reflow_paragraph: lines {}->{}, deltaH={}, pushed {} objects", oldLineCount, newLineCount, deltaH, pushed);
+                }
+
+                // Delete the old paragraph objects (descending so indices stay valid).
+                std::sort(objectIndices.begin(), objectIndices.end(), std::greater<int>());
+                int minIndex = objectIndices.back();
+                for (int idx : objectIndices) {
+                    FPDF_PAGEOBJECT o = FPDFPage_GetObject(page, idx);
+                    if (o) { FPDFPage_RemoveObject(page, o); FPDFPageObj_Destroy(o); }
+                }
+
+                // Emit the new lines (one text object per styled segment, positioned explicitly).
+                // While positioning, also capture a per-character caret layout (PDF coords) so the
+                // live preview can place a caret exactly on the rendered glyphs.
+                nlohmann::json layoutLines = nlohmann::json::array();
+                for (size_t li = 0; li < lines.size(); ++li) {
+                    double baselineY = firstBaselineY - static_cast<double>(li) * leading;
+                    auto& lw = lines[li];
+                    double naturalW = 0.0;
+                    for (size_t k = 0; k < lw.size(); ++k) {
+                        naturalW += words[lw[k]].width;
+                        if (k > 0) naturalW += words[lw[k - 1]].spaceAfter;
+                    }
+                    bool justifyThis = (align == "justify") && (li + 1 < lines.size()) && lw.size() > 1;
+                    double extraPerGap = 0.0;
+                    if (justifyThis) { double slack = columnWidth - naturalW; if (slack > 0) extraPerGap = slack / static_cast<double>(lw.size() - 1); }
+
+                    std::string lineText;
+                    std::vector<double> adv;       // advance (PDF units) of each char in lineText
+                    double lineFontSize = 0.0;
+                    double x = columnLeft;
+                    for (size_t k = 0; k < lw.size(); ++k) {
+                        size_t wi = lw[k];
+                        if (k > 0) {
+                            double gap = words[lw[k - 1]].spaceAfter + (justifyThis ? extraPerGap : 0.0);
+                            x += gap;
+                            lineText.push_back(' ');
+                            adv.push_back(gap);
+                        }
+                        double segX = x;
+                        for (auto& seg : words[wi].segs) {
+                            if (lineFontSize <= 0.0) lineFontSize = runs[seg.runIdx].fontSize;
+                            FPDF_FONT font = runFonts[seg.runIdx].font;
+                            if (font) {
+                                FPDF_PAGEOBJECT obj = FPDFPageObj_CreateTextObj(doc_, font, static_cast<float>(runs[seg.runIdx].fontSize));
+                                if (obj) {
+                                    FPDFPageObj_SetFillColor(obj, runs[seg.runIdx].r, runs[seg.runIdx].g, runs[seg.runIdx].b, 255);
+                                    auto u16 = utf8_to_utf16le(seg.text);
+                                    u16.push_back(0);
+                                    FPDFText_SetText(obj, reinterpret_cast<FPDF_WIDESTRING>(u16.data()));
+                                    FPDFPageObj_Transform(obj, 1.0, 0.0, 0.0, 1.0, segX, baselineY);
+                                    FPDFPage_InsertObjectAtIndex(page, obj, minIndex);
+                                }
+                            }
+                            auto ca = perCharAdvances(seg.runIdx, seg.text);
+                            for (size_t c = 0; c < seg.text.size(); ++c) { lineText.push_back(seg.text[c]); adv.push_back(ca[c]); }
+                            segX += seg.width;
+                        }
+                        x += words[wi].width;
+                    }
+                    layoutLines.push_back({
+                        {"baselineY", baselineY}, {"x0", columnLeft},
+                        {"fontSize", lineFontSize > 0 ? lineFontSize : leading / 1.2},
+                        {"text", lineText}, {"adv", adv},
+                    });
+                }
+                lastReflowLayout_ = nlohmann::json{{"columnLeft", columnLeft}, {"lines", layoutLines}}.dump();
+
+                if (!FPDFPage_GenerateContent(page)) {
+                    spdlog::error("Failed to generate page content after reflow_paragraph");
+                }
                 FPDF_ClosePage(page);
             } else if (type == "text_overlay" || type == "add_text") {
                 if (!op.contains("data") || !op["data"].is_object()) {

engine/src/parser/pdfium_document.hpp

@@ -13,6 +13,7 @@
 #include <mutex>
 #include <unordered_map>
 namespace pdfengine::fonts::loader { class FontResolver; }
+namespace pdfengine::fonts { class FontFace; }
 
 namespace pdfengine::parser {
 
@@ -99,7 +100,13 @@ public:
     std::expected<std::vector<uint8_t>, EngineError> saveIncremental() const override;
     std::expected<std::vector<uint8_t>, EngineError> saveFull() const override;
 
+    // Per-character layout (PDF coords) of the LAST reflow_paragraph op, as JSON. The WASM
+    // preview uses it to align a custom caret exactly to the rendered glyphs. Empty if the
+    // most recent applyEdits contained no reflow_paragraph operation.
+    std::string lastReflowLayout() const { return lastReflowLayout_; }
+
 private:
+    mutable std::string lastReflowLayout_;
     NativeDocHandle doc_ = nullptr;
     std::vector<uint8_t> memoryBuffer_;
 
@@ -126,7 +133,22 @@ private:
 #ifdef PDFENGINE_WITH_PDFIUM
     mutable std::unordered_map<std::string, FPDF_FONT> loadedFontsCache_;
     mutable std::unordered_map<std::string, std::vector<uint8_t>> loadedFontDataBuffers_;
+    mutable std::unordered_map<std::string, std::shared_ptr<fonts::FontFace>> loadedMeasureFaces_;
     mutable std::mutex loadedFontsMutex_;
+
+    // Resolve + load a PDFium font for emitting NEW text (mirrors the replace_text tiers:
+    // embedded full/subset reuse, base-14 standard, or system-font subset embed). Returns
+    // the FPDF_FONT (for FPDFPageObj_CreateTextObj) and the resolved font (FontFace for
+    // HarfBuzz width measurement during line-breaking). Either field may be null on failure.
+    struct EmissionFont {
+        FPDF_FONT font = nullptr;
+        std::shared_ptr<fonts::pdf_fonts::Font> resolved;
+        // A FontFace built from the EXACT bytes loaded into PDFium, so width measurement
+        // matches what PDFium renders (the resolver's face can have different advances).
+        std::shared_ptr<fonts::FontFace> measureFace;
+    };
+    EmissionFont loadEmissionFont(int pageIndex, const std::string& internalFontId,
+                                  double fontSize, const std::vector<uint32_t>& codepoints);
 #endif
 };
 

frontend/src/App.tsx

@@ -12,13 +12,12 @@ import type { CustomConfirmationOptions } from './components/custom/CustomConfir
 import { PDFViewer } from './viewer/PDFViewer';
 import type { PDFViewerRef } from './viewer/PDFViewer';
 import type { Annotation } from './viewer/AnnotationLayer';
-import type { EditableRun } from './viewer/TextEditLayer';
+import type { EditableRun, ReflowParagraphPayload } from './viewer/TextEditLayer';
 import { gatewayService, PasswordError } from './lib/gatewayService';
 import { PasswordModal } from './components/PasswordModal';
 import type { DocumentInfo, SearchResult, EditOperation, DocumentMetadata, FontInfo, OutlineItem, PDFPermissions } from './lib/gatewayService';
 import { viewportRectToPdf } from './lib/coordinateMapping';
 import type { Rect } from './lib/coordinateMapping';
-import { wasmLoader } from './lib/wasmLoader';
 import { toast } from './lib/toast';
 import { DEFAULT_TOOL_SETTINGS, TOOL_SHORTCUTS } from './lib/tools';
 import type { ToolId, ToolSettings } from './lib/tools';
@@ -126,10 +125,6 @@ function App() {
     })();
   }, [openDocument]);
 
-  useEffect(() => {
-    wasmLoader.loadEngine().then((i) => console.log(`[WASM] ${i.engineBuildInfo()}`)).catch(() => {});
-  }, []);
-
   // Load metadata, annotations, fonts when the active document version changes.
   useEffect(() => {
     if (!selectedDocId) return;
@@ -334,6 +329,13 @@ function App() {
     setActiveTool('select');
   };
 
+  // Whole-paragraph reflow (Word-style wrap + push-down) for multi-line paragraph edits.
+  const handleReflowParagraph = (pageIndex: number, payload: ReflowParagraphPayload) => {
+    if (!can('canModify')) { denyToast('Editing text'); setActiveTool('select'); return; }
+    applyOps([{ id: rid('reflow'), type: 'reflow_paragraph', pageIndex, data: payload }], 'Text reflowed');
+    setActiveTool('select');
+  };
+
   const handlePlaceStamp = (pageIndex: number, point: { x: number; y: number }) => {
     if (!activeStamp) return;
     if (!can('canAnnotate')) { denyToast('Stamping'); return; }
@@ -614,6 +616,7 @@ function App() {
                 onRedactArea={handleRedactArea}
                 onPlaceText={handlePlaceText}
                 onEditText={handleEditText}
+                onReflowParagraph={handleReflowParagraph}
                 onPlaceStamp={handlePlaceStamp}
                 onPlaceSignature={handlePlaceSignature}
                 onDecorateText={handleDecorateText}

frontend/src/lib/gatewayService.ts

@@ -220,6 +220,7 @@ export type EditOperationDataMap = {
   delete_annotation: DeleteAnnotationData;
   update_annotation: UpdateAnnotationData;
   replace_text: ReplaceTextData;
+  reflow_paragraph: ReflowParagraphData;
   underline: DecorationData;
   strikeout: DecorationData;
   squiggly: DecorationData;
@@ -244,6 +245,32 @@ export interface ReplaceTextData {
   fontSize: number;
 }
 
+// Whole-paragraph re-layout (Word-style wrap + push-down). The engine line-breaks the
+// styled runs within [columnLeft, columnRight], re-justifies, and shifts following
+// in-column content by the line-count delta.
+export interface ReflowFragment {
+  text: string;
+  internalFontId: string;
+  fontSize: number;
+  color: string;
+}
+
+export interface ReflowParagraphData {
+  objectIndices: number[];
+  runs: ReflowFragment[];
+  columnLeft: number;
+  columnRight: number;
+  firstBaselineY: number;
+  leading: number;
+  oldLineCount: number;
+  align: 'left' | 'justify';
+  // Push-down column left (defaults to columnLeft). Lets a bullet item reflow its text from a
+  // hanging indent while the push-down still moves markers + items below by the full width.
+  pushColumnLeft?: number;
+  // WYSIWYG mode: exact visual line breaks from the live editor (one inner array per line).
+  lines?: ReflowFragment[][];
+}
+
 export interface DeleteAnnotationData {
   annotationId: string;
 }
@@ -423,6 +450,17 @@ class GatewayService {
     return response.json();
   }
 
+  // Raw PDF bytes of the current document version (for the in-browser WASM preview engine).
+  async getDocumentRaw(documentId: string): Promise<ArrayBuffer | null> {
+    try {
+      const r = await fetch(`${this.baseUrl}/documents/${documentId}/raw`);
+      if (!r.ok) return null;
+      return await r.arrayBuffer();
+    } catch {
+      return null;
+    }
+  }
+
   // Raw embedded/substitute font bytes for an in-place-editing preview. Returns null
   // when the font isn't browser-loadable (Type1/non-sfnt → 404) so callers fall back
   // to a CSS font.

frontend/src/lib/pdfiumEngine.ts

@@ -0,0 +1,122 @@
+// Lazy-loaded in-browser PDFium engine (WASM). Renders pages + edit previews PIXEL-IDENTICAL
+// to the gateway (same C++ engine compiled to WASM). Used only for the live-edit preview;
+// the gateway stays authoritative for commit/save. Loaded on demand (the .wasm is ~6.7MB).
+
+interface PdfiumModule {
+  _malloc(n: number): number;
+  _free(p: number): void;
+  HEAPU8: Uint8Array;
+  ccall(name: string, ret: string | null, argTypes: string[], args: unknown[]): number | string;
+}
+
+// Per-character caret layout the engine emits for the reflowed paragraph (PDF coords). Lets the
+// live preview place a caret EXACTLY on the rendered glyphs (same wrap + metrics as the image).
+export interface ReflowLayoutLine { baselineY: number; x0: number; fontSize: number; text: string; adv: number[]; }
+export interface ReflowLayout { columnLeft: number; lines: ReflowLayoutLine[]; }
+export interface PreviewResult { blob: Blob | null; layout: ReflowLayout | null; }
+
+let modulePromise: Promise<PdfiumModule | null> | null = null;
+const docHandles = new Map<string, number>(); // documentId -> wasm doc handle
+
+// Load the Emscripten module once (mirrors the existing wasmLoader mechanism).
+function getModule(): Promise<PdfiumModule | null> {
+  if (!modulePromise) {
+    modulePromise = (async () => {
+      try {
+        // Vite refuses to import /public files as modules ("can only be referenced via HTML
+        // tags"). So fetch the Emscripten module as TEXT and import it via a Blob URL — Vite
+        // never sees it as an import, and the browser loads it as a normal ES module.
+        // Cache-buster: bump V whenever the engine is rebuilt so the browser can never serve a
+        // stale .mjs/.wasm (a normal hard-reload sometimes keeps the multi-MB .wasm cached).
+        const V = '20260615e';
+        const resp = await fetch(`/pdfium-engine.mjs?v=${V}`, { cache: 'no-store' });
+        if (!resp.ok) throw new Error(`pdfium-engine.mjs ${resp.status}`);
+        const blobUrl = URL.createObjectURL(new Blob([await resp.text()], { type: 'text/javascript' }));
+        let createModule: (opts: object) => Promise<PdfiumModule>;
+        try {
+          createModule = (await import(/* @vite-ignore */ blobUrl)).default;
+        } finally {
+          URL.revokeObjectURL(blobUrl);
+        }
+        const M = await createModule({
+          locateFile: (path: string) => (path.endsWith('.wasm') ? `/pdfium-engine.wasm?v=${V}` : path),
+        });
+        console.log(`[pdfiumEngine] ✅ WASM engine v${V} loaded & instantiated`);
+        return M as PdfiumModule;
+      } catch (e) {
+        console.error('[pdfiumEngine] failed to load WASM engine', e);
+        return null;
+      }
+    })();
+  }
+  return modulePromise;
+}
+
+export async function isReady(): Promise<boolean> {
+  return (await getModule()) !== null;
+}
+
+// Load a document into the WASM engine (keyed by documentId). Idempotent per id.
+export async function wasmLoadDocument(documentId: string, bytes: ArrayBuffer): Promise<boolean> {
+  const M = await getModule();
+  if (!M) return false;
+  if (docHandles.has(documentId)) return true;
+  const u8 = new Uint8Array(bytes);
+  const ptr = M._malloc(u8.length);
+  M.HEAPU8.set(u8, ptr);
+  const h = M.ccall('loadDocument', 'number', ['number', 'number'], [ptr, u8.length]) as number;
+  M._free(ptr);
+  console.log(`[pdfiumEngine] loadDocument(${u8.length} bytes) -> handle ${h}`);
+  if (h < 0) return false;
+  docHandles.set(documentId, h);
+  return true;
+}
+
+export function wasmHasDocument(documentId: string): boolean {
+  return docHandles.has(documentId);
+}
+
+function lastRenderBlob(M: PdfiumModule, len: number): Blob | null {
+  if (len <= 0) return null;
+  const ptr = M.ccall('lastRenderPtr', 'number', [], []) as number;
+  // Copy out of the WASM heap before it can move (ALLOW_MEMORY_GROWTH).
+  const bytes = M.HEAPU8.slice(ptr, ptr + len);
+  return new Blob([bytes], { type: 'image/png' });
+}
+
+function lastLayout(M: PdfiumModule): ReflowLayout | null {
+  const json = M.ccall('lastLayoutJson', 'string', [], []) as string;
+  if (!json) return null;
+  try { return JSON.parse(json) as ReflowLayout; } catch { return null; }
+}
+
+// Render the page as-is (PNG blob).
+export async function wasmRenderPage(documentId: string, pageIndex: number, dpi: number): Promise<Blob | null> {
+  const M = await getModule();
+  if (!M) return null;
+  const h = docHandles.get(documentId);
+  if (h === undefined) return null;
+  const len = M.ccall('renderPagePng', 'number', ['number', 'number', 'number'], [h, pageIndex, dpi]) as number;
+  return lastRenderBlob(M, len);
+}
+
+// Render a PREVIEW of an edit (e.g. a reflow_paragraph op) without mutating the loaded doc.
+// Returns the PNG blob AND the per-character caret layout the engine computed for the reflow.
+export async function wasmPreviewRender(
+  documentId: string, pageIndex: number, dpi: number, editsJson: string,
+): Promise<PreviewResult> {
+  const M = await getModule();
+  if (!M) return { blob: null, layout: null };
+  const h = docHandles.get(documentId);
+  if (h === undefined) return { blob: null, layout: null };
+  const len = M.ccall('previewRender', 'number', ['number', 'number', 'number', 'string'], [h, pageIndex, dpi, editsJson]) as number;
+  return { blob: lastRenderBlob(M, len), layout: lastLayout(M) };
+}
+
+export async function wasmFreeDocument(documentId: string): Promise<void> {
+  const h = docHandles.get(documentId);
+  if (h === undefined) return;
+  docHandles.delete(documentId);
+  const M = await getModule();
+  if (M) M.ccall('freeDocument', null, ['number'], [h]);
+}

frontend/src/lib/wasmLoader.ts

@@ -25,10 +25,17 @@ class WasmLoader {
     }
 
     try {
-      // Dynamic import from the public folder / static route
-      // @ts-ignore
-      const moduleUrl = (await import('/pdfengine.mjs?url')).default;
-      const createModule = (await import(/* @vite-ignore */ moduleUrl)).default;
+      // Vite won't import /public files as modules, so fetch the module text and import it
+      // via a Blob URL (the browser loads it as an ES module; Vite never sees the import).
+      const resp = await fetch('/pdfengine.mjs');
+      if (!resp.ok) throw new Error(`pdfengine.mjs ${resp.status}`);
+      const blobUrl = URL.createObjectURL(new Blob([await resp.text()], { type: 'text/javascript' }));
+      let createModule: (opts: object) => Promise<any>;
+      try {
+        createModule = (await import(/* @vite-ignore */ blobUrl)).default;
+      } finally {
+        URL.revokeObjectURL(blobUrl);
+      }
       const Module = await createModule({
         locateFile: (path: string) => {
           if (path.endsWith('.wasm')) {

frontend/src/viewer/PDFViewer.tsx

@@ -5,7 +5,7 @@ import { AnnotationLayer } from './AnnotationLayer';
 import type { Annotation } from './AnnotationLayer';
 import { OverlayLayer } from './OverlayLayer';
 import { TextEditLayer } from './TextEditLayer';
-import type { EditableRun } from './TextEditLayer';
+import type { EditableRun, ReflowParagraphPayload } from './TextEditLayer';
 import { SearchOverlayLayer } from './SearchOverlayLayer';
 import type { Rect } from '../lib/coordinateMapping';
 import { gatewayService } from '../lib/gatewayService';
@@ -36,6 +36,7 @@ interface PDFViewerProps {
   onRedactArea?: (pageIndex: number, bounds: Rect) => void;
   onPlaceText?: (pageIndex: number, rectPts: Rect, text: string) => void;
   onEditText?: (pageIndex: number, run: EditableRun, newText: string) => void;
+  onReflowParagraph?: (pageIndex: number, payload: ReflowParagraphPayload) => void;
   onPlaceStamp?: (pageIndex: number, pointPts: { x: number; y: number }) => void;
   onPlaceSignature?: (pageIndex: number, pointPts: { x: number; y: number }) => void;
   onDecorateText?: (pageIndex: number, lines: Rect[], type: 'underline' | 'strikeout' | 'squiggly', color: string) => void;
@@ -78,6 +79,7 @@ export const PDFViewer = React.forwardRef<PDFViewerRef, PDFViewerProps>(({
   onRedactArea,
   onPlaceText,
   onEditText,
+  onReflowParagraph,
   onPlaceStamp,
   onPlaceSignature,
   onDecorateText,
@@ -482,7 +484,9 @@ export const PDFViewer = React.forwardRef<PDFViewerRef, PDFViewerProps>(({
                       width={page.width}
                       height={page.height}
                       zoom={zoom}
+                      pageImageUrl={imageUrl}
                       onEditText={onEditText}
+                      onReflowParagraph={onReflowParagraph}
                     />
                   )}
 

frontend/src/viewer/ParagraphEditor.tsx

@@ -0,0 +1,445 @@
+import React, { useEffect, useMemo, useRef, useState } from 'react';
+import { gatewayService } from '../lib/gatewayService';
+import { wasmLoadDocument, wasmHasDocument, wasmPreviewRender } from '../lib/pdfiumEngine';
+import type { ReflowLayout } from '../lib/pdfiumEngine';
+import type { ReflowParagraphPayload } from './TextEditLayer';
+import type { ReflowFragment } from '../lib/gatewayService';
+
+// Live paragraph editor with a PIXEL-PERFECT preview. A transparent contentEditable handles
+// input + caret (native), while the in-browser WASM PDFium engine renders the edited paragraph
+// band IDENTICALLY to the page (same C++ engine) and paints it underneath. So clicking in
+// doesn't change the look, and typing reflows in the document's true rendering. Commit goes to
+// the gateway (authoritative). Falls back to plain browser text if the WASM engine is unavailable.
+
+interface SeedRun { text: string; fid: string; size: number; color: string; fontName: string; }
+interface ParagraphLayout {
+  columnLeft: number; columnRight: number; firstBaselineY: number; leading: number;
+  oldLineCount: number; align: 'left' | 'justify'; objectIndices: number[]; seedRuns: SeedRun[];
+}
+interface ParagraphEditorProps {
+  documentId: string;
+  pageIndex: number;
+  para: any;
+  // Bullet-item reflow: full-width push-down left, and overrides for spacing/alignment that the
+  // sub-paragraph can't infer on its own (a single-line item has no leading; bullets aren't justified).
+  pushColumnLeft?: number;
+  leadingOverride?: number;
+  alignOverride?: 'left' | 'justify';
+  caretClick?: { x: number; y: number } | null;
+  heightPts: number;
+  zoom: number;
+  pageWidthPx: number;
+  pageHeightPx: number;
+  onCommit: (payload: ReflowParagraphPayload) => void;
+  onCancel: () => void;
+}
+
+function median(xs: number[]): number {
+  if (xs.length === 0) return 0;
+  const s = [...xs].sort((a, b) => a - b);
+  return s[Math.floor(s.length / 2)];
+}
+function computeLayout(para: any): ParagraphLayout {
+  const lines = para?.lines ?? [];
+  let columnLeft = Infinity, columnRight = -Infinity, firstBaselineY = -Infinity;
+  const baselines: number[] = [], rightEdges: number[] = [], objectIndices: number[] = [];
+  const seedRuns: SeedRun[] = [];
+  for (let li = 0; li < lines.length; li++) {
+    const line = lines[li];
+    if (typeof line.baseline_y === 'number') { baselines.push(line.baseline_y); firstBaselineY = Math.max(firstBaselineY, line.baseline_y); }
+    columnLeft = Math.min(columnLeft, line.x);
+    columnRight = Math.max(columnRight, line.x + line.w);
+    rightEdges.push(line.x + line.w);
+    const lineRuns = line.runs ?? [];
+    for (let ri = 0; ri < lineRuns.length; ri++) {
+      const r = lineRuns[ri];
+      (Array.isArray(r.object_indices) ? r.object_indices : []).forEach((o: number) => objectIndices.push(o));
+      let text = r.text ?? '';
+      if (li > 0 && ri === 0 && seedRuns.length > 0) {
+        const prev = seedRuns[seedRuns.length - 1].text;
+        if (prev && !/\s$/.test(prev) && !/^\s/.test(text)) text = ' ' + text;
+      }
+      seedRuns.push({ text, fid: r.internal_font_id ?? '', size: r.font_size ?? 12, color: typeof r.color === 'string' ? r.color : '#000000', fontName: r.font_name ?? '' });
+    }
+  }
+  const deltas: number[] = [];
+  for (let i = 0; i < baselines.length - 1; i++) deltas.push(baselines[i] - baselines[i + 1]);
+  const domSize = seedRuns.find((r) => r.text.trim())?.size ?? 12;
+  const leading = deltas.length ? Math.abs(median(deltas)) : domSize * 1.2;
+  const colW = columnRight - columnLeft;
+  let align: 'left' | 'justify' = 'left';
+  if (lines.length >= 2) {
+    let reaching = 0;
+    for (let i = 0; i < rightEdges.length - 1; i++) if (rightEdges[i] >= columnRight - colW * 0.04) reaching++;
+    if (reaching >= (lines.length - 1) * 0.7) align = 'justify';
+  }
+  return { columnLeft, columnRight, firstBaselineY, leading, oldLineCount: lines.length, align, objectIndices, seedRuns };
+}
+
+// Flat styled runs from the editable's current text (each text node → a fragment). The engine
+// wraps these into lines itself (greedy break within the column), so the wrap + justification of
+// the rendered image and the exported caret layout always agree — no browser layout involved.
+function extractFlatRuns(editable: HTMLElement, dominantFid: string): ReflowFragment[] {
+  const out: ReflowFragment[] = [];
+  const walker = document.createTreeWalker(editable, NodeFilter.SHOW_TEXT);
+  let node = walker.nextNode() as Text | null;
+  while (node) {
+    const el = node.parentElement;
+    const fid = el?.getAttribute('data-fid') || dominantFid;
+    const size = parseFloat(el?.getAttribute('data-size') ?? '12') || 12;
+    const color = el?.getAttribute('data-color') ?? '#000000';
+    const text = node.textContent ?? '';
+    if (text) out.push({ text, internalFontId: fid, fontSize: size, color });
+    node = walker.nextNode() as Text | null;
+  }
+  return out;
+}
+
+// The paragraph's ORIGINAL line breaks as fragments (one inner array per line). Used for the
+// initial render so opening the editor shows the document's existing wrapping verbatim — the
+// engine only re-wraps once the user actually edits.
+function buildOriginalLines(para: any, dominantFid: string): ReflowFragment[][] {
+  const out: ReflowFragment[][] = [];
+  for (const line of para?.lines ?? []) {
+    const frags: ReflowFragment[] = [];
+    for (const r of line.runs ?? []) {
+      const text = r.text ?? '';
+      if (text) frags.push({ text, internalFontId: r.internal_font_id || dominantFid, fontSize: r.font_size ?? 12, color: typeof r.color === 'string' ? r.color : '#000000' });
+    }
+    if (frags.length) out.push(frags);
+  }
+  return out;
+}
+
+// Global character offset of the DOM caret within the editable.
+function globalCaretOffset(el: HTMLElement): number {
+  const sel = window.getSelection();
+  if (!sel || sel.rangeCount === 0) return 0;
+  const range = sel.getRangeAt(0);
+  if (!el.contains(range.startContainer)) return 0;
+  const pre = document.createRange();
+  pre.selectNodeContents(el);
+  pre.setEnd(range.startContainer, range.startOffset);
+  return pre.toString().length;
+}
+
+// Move the DOM caret to a global character offset (so typing inserts at the right place).
+function setGlobalCaretOffset(el: HTMLElement, target: number): void {
+  const walker = document.createTreeWalker(el, NodeFilter.SHOW_TEXT);
+  let acc = 0;
+  let node = walker.nextNode() as Text | null;
+  const sel = window.getSelection();
+  while (node) {
+    const len = node.textContent?.length ?? 0;
+    if (acc + len >= target) {
+      const range = document.createRange();
+      range.setStart(node, Math.max(0, Math.min(target - acc, len)));
+      range.collapse(true);
+      sel?.removeAllRanges(); sel?.addRange(range);
+      return;
+    }
+    acc += len;
+    node = walker.nextNode() as Text | null;
+  }
+  const range = document.createRange();
+  range.selectNodeContents(el); range.collapse(false);
+  sel?.removeAllRanges(); sel?.addRange(range);
+}
+
+// Global char index where each engine line starts in the full text (continuation lines skip the
+// whitespace the engine consumed at the wrap point).
+function lineStarts(layout: ReflowLayout, fullText: string): number[] {
+  const starts: number[] = [];
+  let pos = 0;
+  for (let i = 0; i < layout.lines.length; i++) {
+    if (i > 0) { while (pos < fullText.length && /\s/.test(fullText[pos])) pos++; }
+    starts.push(pos);
+    pos += layout.lines[i].text.length;
+  }
+  return starts;
+}
+
+export const ParagraphEditor: React.FC<ParagraphEditorProps> = ({
+  documentId, pageIndex, para, pushColumnLeft, leadingOverride, alignOverride,
+  caretClick, heightPts, zoom, pageWidthPx, pageHeightPx, onCommit, onCancel,
+}) => {
+  const layout = useMemo(() => computeLayout(para), [para]);
+  // A bullet item supplies the paragraph's true leading + forces left alignment; a flowing
+  // paragraph uses what computeLayout inferred.
+  const leading = leadingOverride ?? layout.leading;
+  const align = alignOverride ?? layout.align;
+  const editRef = useRef<HTMLDivElement>(null);
+  const committedRef = useRef(false);
+  const initialTextRef = useRef('');
+  const debounceRef = useRef<number | null>(null);
+  const blobUrlRef = useRef<string | null>(null);
+  const engineLayoutRef = useRef<ReflowLayout | null>(null);
+  const initialCaretApplied = useRef(false);
+  // Until the user actually edits, the preview keeps the document's ORIGINAL line breaks (no
+  // re-wrap), so merely opening the editor doesn't shift anything. Re-wrap kicks in on first edit.
+  const editedRef = useRef(false);
+  const [previewUrl, setPreviewUrl] = useState<string | null>(null);
+  const [caretBox, setCaretBox] = useState<{ left: number; top: number; height: number } | null>(null);
+  // True once the user actually edits. Until then we show the UNTOUCHED original page (pixel-
+  // perfect) rather than the reflow render, so merely opening the editor changes nothing.
+  const [edited, setEdited] = useState(false);
+  // Falls back to the plain (visible) browser editor only if the WASM engine fails to produce
+  // a render in a few seconds — so the user is never stuck editing invisible text.
+  const [wasmFailed, setWasmFailed] = useState(false);
+
+  // The font most of the paragraph uses — empty-fid runs (spaces, freshly typed text) adopt it
+  // so the engine emits them in the real document font instead of a generic fallback.
+  const dominantFid = layout.seedRuns.find((r) => r.text.trim() && r.fid)?.fid
+    ?? layout.seedRuns.find((r) => r.fid)?.fid ?? '';
+  const domSize = layout.seedRuns.find((r) => r.text.trim())?.size ?? layout.seedRuns[0]?.size ?? 12;
+  const fontPx = domSize * zoom;
+  const leadingPx = leading * zoom;
+  const colLeftPx = layout.columnLeft * zoom;
+  const colWidthPx = (layout.columnRight - layout.columnLeft) * zoom;
+  const firstBaselineScreen = (heightPts - layout.firstBaselineY) * zoom;
+  const editorTop = firstBaselineScreen - (leadingPx + fontPx * 0.7) / 2;
+  // The WASM preview repaints the page from the paragraph top down (reflow pushes content below).
+  const bandTop = Math.max(0, editorTop - leadingPx * 0.5);
+
+  // Reflow op JSON for the engine. With `lines` (initial render) the engine emits those exact
+  // breaks — opening the editor changes nothing. Without `lines` (after an edit) it greedy-wraps
+  // within the column, and its render + exported caret layout agree exactly (no browser layout).
+  const buildOpJson = (runs: ReflowFragment[], originalLines?: ReflowFragment[][]): string => JSON.stringify({
+    version: '1.0',
+    operations: [{ id: 'preview', type: 'reflow_paragraph', pageIndex, data: {
+      objectIndices: layout.objectIndices,
+      runs: runs.length ? runs : [{ text: ' ', internalFontId: dominantFid, fontSize: domSize, color: '#000000' }],
+      ...(originalLines && originalLines.length ? { lines: originalLines } : {}),
+      columnLeft: layout.columnLeft, columnRight: layout.columnRight,
+      pushColumnLeft: pushColumnLeft ?? layout.columnLeft,
+      firstBaselineY: layout.firstBaselineY, leading,
+      oldLineCount: layout.oldLineCount, align,
+    } }],
+  });
+
+  // Caret screen box (px in the page container) for a global char offset, from the engine layout.
+  const caretBoxFor = (global: number, lay: ReflowLayout, fullText: string) => {
+    if (!lay.lines.length) return null;
+    const starts = lineStarts(lay, fullText);
+    let li = lay.lines.length - 1;
+    for (let i = 0; i < lay.lines.length; i++) {
+      if (global <= starts[i] + lay.lines[i].text.length) { li = i; break; }
+    }
+    const line = lay.lines[li];
+    const offset = Math.max(0, Math.min(global - starts[li], line.adv.length));
+    let x = line.x0;
+    for (let k = 0; k < offset; k++) x += line.adv[k] ?? 0;
+    const fpx = line.fontSize * zoom;
+    return { left: x * zoom, top: (heightPts - line.baselineY) * zoom - fpx * 0.82, height: fpx * 1.04 };
+  };
+
+  // Map a screen click to a global char offset via the engine layout (line by baseline, char by adv).
+  const globalFromPoint = (clientX: number, clientY: number, lay: ReflowLayout, fullText: string) => {
+    const el = editRef.current;
+    const container = el?.offsetParent as HTMLElement | null;
+    if (!container || !lay.lines.length) return 0;
+    const rect = container.getBoundingClientRect();
+    const pdfX = (clientX - rect.left) / zoom;
+    const localY = clientY - rect.top;
+    let li = 0, bestD = Infinity;
+    for (let i = 0; i < lay.lines.length; i++) {
+      const fpx = lay.lines[i].fontSize * zoom;
+      const mid = (heightPts - lay.lines[i].baselineY) * zoom - fpx * 0.35;
+      const d = Math.abs(localY - mid);
+      if (d < bestD) { bestD = d; li = i; }
+    }
+    const line = lay.lines[li];
+    let x = line.x0, offset = 0;
+    for (let k = 0; k < line.adv.length; k++) {
+      const next = x + line.adv[k];
+      if (pdfX < (x + next) / 2) break;
+      x = next; offset = k + 1;
+    }
+    return lineStarts(lay, fullText)[li] + offset;
+  };
+
+  const positionCaret = () => {
+    const el = editRef.current, lay = engineLayoutRef.current;
+    if (!el || !lay) return;
+    setCaretBox(caretBoxFor(globalCaretOffset(el), lay, el.textContent ?? ''));
+  };
+
+  // Render the current edit state via the WASM engine (pixel-identical to the page) + reposition
+  // the caret from the engine's own glyph layout.
+  const renderPreview = async () => {
+    const el = editRef.current;
+    if (!el) return;
+    const dpi = Math.round(72 * zoom);
+    // Before the first edit, keep the document's original line breaks (no re-wrap on open).
+    const originalLines = editedRef.current ? undefined : buildOriginalLines(para, dominantFid);
+    const { blob, layout: lay } = await wasmPreviewRender(documentId, pageIndex, dpi, buildOpJson(extractFlatRuns(el, dominantFid), originalLines));
+    if (!blob) { console.warn('[ParagraphEditor] WASM preview returned null'); return; }
+    engineLayoutRef.current = lay;
+    const url = URL.createObjectURL(blob);
+    if (blobUrlRef.current) URL.revokeObjectURL(blobUrlRef.current);
+    blobUrlRef.current = url;
+    setPreviewUrl(url);
+    // On the very first render, place the caret where the user clicked (mapped via engine layout).
+    if (!initialCaretApplied.current) {
+      initialCaretApplied.current = true;
+      if (caretClick && lay) setGlobalCaretOffset(el, globalFromPoint(caretClick.x, caretClick.y, lay, el.textContent ?? ''));
+    }
+    positionCaret();
+  };
+
+  // Load the document into the WASM engine, then do the initial render.
+  useEffect(() => {
+    let cancelled = false;
+    (async () => {
+      if (!wasmHasDocument(documentId)) {
+        const bytes = await gatewayService.getDocumentRaw(documentId);
+        if (bytes) await wasmLoadDocument(documentId, bytes);
+      }
+      if (cancelled) return;
+      renderPreview();
+    })();
+    return () => {
+      cancelled = true;
+      if (debounceRef.current) window.clearTimeout(debounceRef.current);
+      if (blobUrlRef.current) URL.revokeObjectURL(blobUrlRef.current);
+    };
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [documentId]);
+
+  // Seed the contentEditable ONCE (one span per run; spaces/empty-fid adopt the dominant font).
+  useEffect(() => {
+    const el = editRef.current;
+    if (!el) return;
+    el.innerHTML = '';
+    for (const r of layout.seedRuns) {
+      if (!r.text) continue;
+      const effSize = !r.text.trim() || r.size <= 1 ? domSize : r.size;
+      const span = document.createElement('span');
+      span.setAttribute('data-fid', r.fid || dominantFid);
+      span.setAttribute('data-size', String(effSize));
+      span.setAttribute('data-color', r.color);
+      span.setAttribute('data-fontname', r.fontName);
+      span.style.fontSize = `${effSize * zoom}px`;
+      span.style.color = 'transparent';
+      span.textContent = r.text;
+      el.appendChild(span);
+    }
+    el.focus();
+    initialTextRef.current = (el.textContent ?? '').replace(/\s+/g, ' ').trim();
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []);
+
+  const imageReady = previewUrl !== null;
+  const fallbackVisible = wasmFailed && !imageReady;
+  // In the fallback (WASM unavailable) the plain glyphs become visible and we use the native caret.
+  useEffect(() => {
+    const el = editRef.current;
+    if (!el) return;
+    el.querySelectorAll<HTMLElement>('span[data-fid]').forEach((span) => {
+      span.style.color = fallbackVisible ? (span.getAttribute('data-color') ?? '#000000') : 'transparent';
+    });
+  }, [fallbackVisible]);
+
+  // Fall back to the visible browser editor only if no render arrives within ~4s.
+  useEffect(() => {
+    if (previewUrl) return;
+    const t = window.setTimeout(() => setWasmFailed(true), 4000);
+    return () => window.clearTimeout(t);
+  }, [previewUrl]);
+
+  // Re-render shortly after each keystroke (the caret repositions from the new layout).
+  const onInput = () => {
+    editedRef.current = true; // now the engine may re-wrap (the user is actually editing)
+    if (!edited) setEdited(true); // swap from the untouched original page to the reflow render
+    positionCaret(); // immediate (approximate, from the prior layout) for responsiveness
+    if (debounceRef.current) window.clearTimeout(debounceRef.current);
+    debounceRef.current = window.setTimeout(renderPreview, 110);
+  };
+
+  // Place the caret on click via the engine layout. Using onClick (not onMouseDown+preventDefault)
+  // means focus is already established, so a SINGLE click lands the caret correctly.
+  const onClickEditor = (e: React.MouseEvent) => {
+    const lay = engineLayoutRef.current, el = editRef.current;
+    if (!lay || !el || fallbackVisible) return;
+    setGlobalCaretOffset(el, globalFromPoint(e.clientX, e.clientY, lay, el.textContent ?? ''));
+    positionCaret();
+  };
+
+  const commit = () => {
+    if (committedRef.current) return;
+    committedRef.current = true;
+    const el = editRef.current;
+    if (!el) { onCancel(); return; }
+    const nowText = (el.textContent ?? '').replace(/\s+/g, ' ').trim();
+    if (nowText === initialTextRef.current) { onCancel(); return; }
+    const flat = extractFlatRuns(el, dominantFid);
+    if (flat.length === 0) { onCancel(); return; }
+    // Commit with FLAT runs (no lines) so the gateway wraps identically to the live preview.
+    onCommit({
+      objectIndices: layout.objectIndices, runs: flat,
+      columnLeft: layout.columnLeft, columnRight: layout.columnRight,
+      pushColumnLeft: pushColumnLeft ?? layout.columnLeft,
+      firstBaselineY: layout.firstBaselineY, leading,
+      oldLineCount: layout.oldLineCount, align,
+    });
+  };
+  const cancel = () => { committedRef.current = true; onCancel(); };
+
+  return (
+    <>
+      <style>{`@keyframes pe-caret-blink{0%,49%{opacity:1}50%,100%{opacity:0}}`}</style>
+      {/* White cover ONLY in the fallback case (WASM failed). */}
+      {fallbackVisible && (
+        <div
+          className="absolute z-[36] bg-white"
+          style={{ left: colLeftPx - 2, top: editorTop - 2, width: colWidthPx + 4, height: layout.oldLineCount * leadingPx + 8 }}
+        />
+      )}
+      {/* Reflow render — shown ONLY after the user edits. Before that, the untouched original
+          page (the canvas below this layer) shows through, so opening the editor changes nothing. */}
+      {edited && previewUrl && (
+        <div
+          className="absolute z-[37] overflow-hidden"
+          style={{ left: 0, top: bandTop, width: pageWidthPx, height: Math.max(0, pageHeightPx - bandTop), pointerEvents: 'none' }}
+        >
+          <img src={previewUrl} alt="" style={{ position: 'absolute', left: 0, top: -bandTop, width: pageWidthPx, height: pageHeightPx, maxWidth: 'none' }} />
+        </div>
+      )}
+
+      {/* Custom caret, positioned from the ENGINE's glyph layout so it lands exactly on the
+          rendered text (the browser's own layout is never used for positioning). */}
+      {!fallbackVisible && caretBox && (
+        <div
+          className="absolute z-[39]"
+          style={{ left: caretBox.left, top: caretBox.top, height: caretBox.height, width: 1.6, background: '#2563eb', animation: 'pe-caret-blink 1s step-end infinite', pointerEvents: 'none' }}
+        />
+      )}
+
+      {/* Transparent input layer: holds the text + receives keys; its OWN layout is ignored. */}
+      <div
+        ref={editRef}
+        contentEditable
+        suppressContentEditableWarning
+        spellCheck={false}
+        onInput={onInput}
+        onClick={onClickEditor}
+        onKeyUp={positionCaret}
+        onKeyDown={(e) => {
+          if (e.key === 'Enter') { e.preventDefault(); commit(); }
+          if (e.key === 'Escape') { e.preventDefault(); cancel(); }
+        }}
+        onBlur={commit}
+        onPaste={(e) => { e.preventDefault(); document.execCommand('insertText', false, e.clipboardData.getData('text/plain')); }}
+        className="absolute z-[38] outline-none"
+        style={{
+          left: colLeftPx, top: editorTop, width: colWidthPx, minHeight: layout.oldLineCount * leadingPx,
+          fontSize: `${fontPx}px`, lineHeight: `${leadingPx}px`,
+          caretColor: fallbackVisible ? '#2563eb' : 'transparent',
+          color: fallbackVisible ? undefined : 'transparent',
+          background: fallbackVisible ? '#ffffff' : 'transparent',
+          whiteSpace: 'normal', overflowWrap: 'break-word',
+        }}
+      />
+    </>
+  );
+};

frontend/src/viewer/TextEditLayer.tsx

@@ -1,6 +1,7 @@
 import React, { useEffect, useRef, useState } from 'react';
 import { gatewayService } from '../lib/gatewayService';
 import { loadPdfFont, releaseDocumentFonts } from '../lib/fontFaceLoader';
+import { ParagraphEditor } from './ParagraphEditor';
 
 // A single editable run. Geometry (x/y/w/h, baselineY) is PDF BOTTOM-LEFT (as
 // getPageModel returns) and is used only to position the inline editor. The
@@ -17,6 +18,104 @@ export interface EditableRun {
   internalFontId: string;
   fontName: string;
   color: string;
+  // Position of this run within the raw page model, so a commit can reconstruct its paragraph.
+  paraIndex: number;
+  lineIndex: number;
+  runIndex: number;
+}
+
+// A whole-paragraph re-layout request (Word-style wrap + push-down), built on commit when the
+// edited run's paragraph spans multiple lines. The engine owns the actual line-breaking.
+export interface ReflowFragment {
+  text: string;
+  internalFontId: string;
+  fontSize: number;
+  color: string;
+}
+
+export interface ReflowParagraphPayload {
+  objectIndices: number[];
+  runs: ReflowFragment[];
+  columnLeft: number;
+  columnRight: number;
+  firstBaselineY: number;
+  leading: number;
+  oldLineCount: number;
+  align: 'left' | 'justify';
+  pushColumnLeft?: number; // full-width push-down left for bullet items (see engine)
+  // WYSIWYG: exact visual line breaks from the live editor (one inner array per line).
+  lines?: ReflowFragment[][];
+}
+
+// A bullet marker glyph at the start of a list item (•, -, ▪, etc.).
+export function isBulletMarker(text?: string): boolean {
+  if (!text) return false;
+  const t = text.trim();
+  return t.length <= 2 && /^[•▪◦‣·●○■□*‒–—\-]$/.test(t);
+}
+
+// One logical item to reflow inside a non-flowing paragraph (a single bullet + its wrapped lines,
+// or a leading non-bullet block). Returns a sub-paragraph whose lines exclude the bullet marker
+// (so it stays put) with the text column starting at the hanging indent, plus the full-width
+// pushColumnLeft and the paragraph's true leading.
+export function buildBulletItem(para: any, runLineIndex: number): { subPara: any; pushColumnLeft: number; leading: number } | null {
+  const lines = para?.lines ?? [];
+  if (!lines.length) return null;
+  const colLeft = Math.min(...lines.map((l: any) => l.x));
+  const isStart = (l: any) => isBulletMarker((l.runs ?? [])[0]?.text);
+  // Item = from the nearest marker line at/above the click (or paragraph start) to before the next.
+  let start = runLineIndex;
+  while (start > 0 && !isStart(lines[start])) start--;
+  let end = runLineIndex + 1;
+  while (end < lines.length && !isStart(lines[end])) end++;
+  const itemLines = lines.slice(start, end);
+  if (!itemLines.length) return null;
+
+  // Paragraph leading (line-spacing) from all baselines — used so a newly-wrapped line gets the
+  // document's true spacing, not a guess.
+  const baselines = lines.map((l: any) => l.baseline_y).filter((b: any) => typeof b === 'number');
+  const deltas: number[] = [];
+  for (let i = 0; i < baselines.length - 1; i++) deltas.push(Math.abs(baselines[i] - baselines[i + 1]));
+  const leading = deltas.length ? median(deltas) : (itemLines[0].runs?.[0]?.font_size ?? 12) * 1.2;
+
+  // Strip a leading bullet marker (+ any leading space) from the first line; the text indent is
+  // where the real text begins (which the wrapped lines already hang to).
+  const firstRuns = itemLines[0].runs ?? [];
+  let subLines = itemLines;
+  if (isBulletMarker(firstRuns[0]?.text)) {
+    let ti = 1;
+    while (ti < firstRuns.length && !(firstRuns[ti].text ?? '').trim()) ti++;
+    const textRuns = firstRuns.slice(ti);
+    if (!textRuns.length) return null;
+    const textIndent = textRuns[0].x;
+    subLines = itemLines.map((l: any, idx: number) => idx === 0
+      ? { ...l, runs: textRuns, x: textIndent, w: (l.x + l.w) - textIndent }
+      : l);
+  }
+  return { subPara: { ...para, lines: subLines }, pushColumnLeft: colLeft, leading };
+}
+
+// True only for a genuine FLOWING paragraph — multiple lines that fill the column from a common
+// left edge (e.g. the Professional Summary). Bullet lists / structured blocks get grouped into a
+// single "paragraph" by the model too, but they must NOT be reflowed (it would merge the bullets
+// into one justified blob and mangle markers like • and –). Those fall back to per-line editing.
+function isFlowingParagraph(para: any): boolean {
+  const lines = para?.lines ?? [];
+  if (lines.length < 2) return false;
+  const lefts = lines.map((l: any) => l.x);
+  const colLeft = Math.min(...lefts);
+  const colRight = Math.max(...lines.map((l: any) => l.x + l.w));
+  const colW = colRight - colLeft;
+  if (colW <= 0) return false;
+  // Most non-last lines must reach near the right edge (a filled column, not ragged list items).
+  let reaching = 0;
+  for (let i = 0; i < lines.length - 1; i++) {
+    if (lines[i].x + lines[i].w >= colRight - colW * 0.06) reaching++;
+  }
+  const filled = reaching >= (lines.length - 1) * 0.7;
+  // And every line must start at (roughly) the same left edge — bullets have hanging indents.
+  const consistentLeft = Math.max(...lefts.map((x: number) => Math.abs(x - colLeft))) < colW * 0.12;
+  return filled && consistentLeft;
 }
 
 interface TextEditLayerProps {
@@ -25,7 +124,9 @@ interface TextEditLayerProps {
   width: number; // page width in ZOOMED px (= widthPts * zoom)
   height: number; // page height in ZOOMED px (= heightPts * zoom)
   zoom: number;
+  pageImageUrl?: string; // rendered page image, for the live push-down preview
   onEditText?: (pageIndex: number, run: EditableRun, newText: string) => void;
+  onReflowParagraph?: (pageIndex: number, payload: ReflowParagraphPayload) => void;
 }
 
 // Nearest character index to an x-offset (px from the run's left edge), measured in the
@@ -58,14 +159,19 @@ function fallbackFamily(fontName: string): string {
   return 'Arial, "Helvetica Neue", Helvetica, sans-serif';
 }
 
-// Flatten the dynamic page-model JSON into a flat list of editable runs.
+// Flatten the dynamic page-model JSON into a flat list of editable runs, tagging each with its
+// (paragraph, line, run) position in the raw model so a commit can reconstruct the paragraph.
 function flattenRuns(model: any): EditableRun[] {
   const runs: EditableRun[] = [];
   const paragraphs = model?.paragraphs ?? [];
-  for (const p of paragraphs) {
-    for (const line of p.lines ?? []) {
+  for (let pi = 0; pi < paragraphs.length; pi++) {
+    const lines = paragraphs[pi].lines ?? [];
+    for (let li = 0; li < lines.length; li++) {
+      const line = lines[li];
       const baselineY = typeof line.baseline_y === 'number' ? line.baseline_y : 0;
-      for (const r of line.runs ?? []) {
+      const lineRuns = line.runs ?? [];
+      for (let ri = 0; ri < lineRuns.length; ri++) {
+        const r = lineRuns[ri];
         const objectIndices: number[] = Array.isArray(r.object_indices) ? r.object_indices : [];
         // Only runs backed by real page objects are editable (replace_text targets them).
         if (typeof r.text === 'string' && r.text.trim() && r.w > 0 && r.h > 0 && objectIndices.length > 0) {
@@ -78,6 +184,7 @@ function flattenRuns(model: any): EditableRun[] {
             internalFontId: r.internal_font_id ?? '',
             fontName: r.font_name ?? '',
             color: typeof r.color === 'string' ? r.color : '#000000',
+            paraIndex: pi, lineIndex: li, runIndex: ri,
           });
         }
       }
@@ -86,6 +193,76 @@ function flattenRuns(model: any): EditableRun[] {
   return runs;
 }
 
+// Median of an array (used for line leading).
+function median(xs: number[]): number {
+  if (xs.length === 0) return 0;
+  const s = [...xs].sort((a, b) => a - b);
+  return s[Math.floor(s.length / 2)];
+}
+
+// Reconstruct the edited run's paragraph as a reflow request: collect every run in reading
+// order (substituting the edited run's new text), infer column bounds / leading / alignment.
+function buildReflowPayload(model: any, run: EditableRun, newText: string): ReflowParagraphPayload | null {
+  const para = model?.paragraphs?.[run.paraIndex];
+  if (!para || !Array.isArray(para.lines)) return null;
+
+  const runs: ReflowParagraphPayload['runs'] = [];
+  const objectIndices: number[] = [];
+  let columnLeft = Infinity, columnRight = -Infinity, firstBaselineY = -Infinity;
+  const baselines: number[] = [];
+  const rightEdges: number[] = [];
+
+  for (let li = 0; li < para.lines.length; li++) {
+    const line = para.lines[li];
+    if (typeof line.baseline_y === 'number') baselines.push(line.baseline_y);
+    columnLeft = Math.min(columnLeft, line.x);
+    columnRight = Math.max(columnRight, line.x + line.w);
+    rightEdges.push(line.x + line.w);
+    if (typeof line.baseline_y === 'number') firstBaselineY = Math.max(firstBaselineY, line.baseline_y);
+    for (let ri = 0; ri < (line.runs ?? []).length; ri++) {
+      const r = line.runs[ri];
+      const oi: number[] = Array.isArray(r.object_indices) ? r.object_indices : [];
+      objectIndices.push(...oi);
+      const isEdited = li === run.lineIndex && ri === run.runIndex;
+      let text = isEdited ? newText : (r.text ?? '');
+      // A model line break carries no space character (the gap was positional), so insert
+      // one at line boundaries — otherwise "high-performance" + "backend" merge on reflow.
+      if (li > 0 && ri === 0 && runs.length > 0) {
+        const prev = runs[runs.length - 1].text;
+        if (prev && !/\s$/.test(prev) && !/^\s/.test(text)) text = ' ' + text;
+      }
+      runs.push({
+        text,
+        internalFontId: r.internal_font_id ?? '',
+        fontSize: r.font_size ?? run.fontSize,
+        color: typeof r.color === 'string' ? r.color : '#000000',
+      });
+    }
+  }
+  if (runs.length === 0 || objectIndices.length === 0) return null;
+
+  // Leading = median baseline-to-baseline gap (bottom-left → positive going down).
+  const deltas: number[] = [];
+  for (let i = 0; i < baselines.length - 1; i++) deltas.push(baselines[i] - baselines[i + 1]);
+  const leading = deltas.length ? Math.abs(median(deltas)) : run.fontSize * 1.2;
+
+  // Alignment: if most non-last lines reach the right margin, it's justified.
+  const colW = columnRight - columnLeft;
+  let align: 'left' | 'justify' = 'left';
+  if (para.lines.length >= 2) {
+    let reaching = 0;
+    for (let i = 0; i < rightEdges.length - 1; i++) {
+      if (rightEdges[i] >= columnRight - colW * 0.04) reaching++;
+    }
+    if (reaching >= (para.lines.length - 1) * 0.7) align = 'justify';
+  }
+
+  return {
+    objectIndices, runs, columnLeft, columnRight, firstBaselineY,
+    leading, oldLineCount: para.lines.length, align,
+  };
+}
+
 export const TextEditLayer: React.FC<TextEditLayerProps> = ({
   documentId,
   pageIndex,
@@ -93,6 +270,7 @@ export const TextEditLayer: React.FC<TextEditLayerProps> = ({
   height,
   zoom,
   onEditText,
+  onReflowParagraph,
 }) => {
   const [runs, setRuns] = useState<EditableRun[]>([]);
   const [editing, setEditing] = useState<number | null>(null);
@@ -102,18 +280,28 @@ export const TextEditLayer: React.FC<TextEditLayerProps> = ({
   const committedRef = useRef(false);
   const inputRef = useRef<HTMLInputElement>(null);
   const caretIdxRef = useRef<number | null>(null);
+  const modelRef = useRef<any>(null);
+  // When set, the live reflow editor is open. `para` is the (sub-)paragraph to reflow — the whole
+  // paragraph for a flowing block, or a single bullet item (marker stripped) for a list.
+  const [paraEdit, setParaEdit] = useState<{
+    para: any; pushColumnLeft?: number; leading?: number; align?: 'left' | 'justify';
+  } | null>(null);
+  // Screen coords of the click that opened the paragraph editor, so the caret lands there
+  // (instead of jumping to the paragraph start).
+  const [caretClick, setCaretClick] = useState<{ x: number; y: number } | null>(null);
 
   // Fetch the page model once per document/page.
   useEffect(() => {
     let cancelled = false;
     setEditing(null);
+    setParaEdit(null);
     gatewayService
       .getPageModel(documentId, pageIndex)
       .then((model) => {
-        if (!cancelled) setRuns(flattenRuns(model));
+        if (!cancelled) { modelRef.current = model; setRuns(flattenRuns(model)); }
       })
       .catch(() => {
-        if (!cancelled) setRuns([]);
+        if (!cancelled) { modelRef.current = null; setRuns([]); }
       });
     return () => {
       cancelled = true;
@@ -135,9 +323,27 @@ export const TextEditLayer: React.FC<TextEditLayerProps> = ({
     height: r.h * zoom,
   });
 
-  const openEditor = (i: number, clickX: number) => {
-    committedRef.current = false;
+  const openEditor = (i: number, clickX: number, clientX?: number, clientY?: number) => {
     const run = runs[i];
+    // Multi-line content → live reflow editor. A FLOWING paragraph reflows whole; a bullet/list
+    // reflows just the clicked ITEM (marker kept in place, hanging indent preserved).
+    const para = modelRef.current?.paragraphs?.[run.paraIndex];
+    if (Array.isArray(para?.lines) && para.lines.length > 1 && onReflowParagraph) {
+      const click = clientX != null && clientY != null ? { x: clientX, y: clientY } : null;
+      if (isFlowingParagraph(para)) {
+        setCaretClick(click);
+        setParaEdit({ para });
+        return;
+      }
+      const item = buildBulletItem(para, run.lineIndex);
+      if (item) {
+        setCaretClick(click);
+        setParaEdit({ para: item.subPara, pushColumnLeft: item.pushColumnLeft, leading: item.leading, align: 'left' });
+        return;
+      }
+      // else: couldn't scope an item → fall through to per-line in-place editing
+    }
+    committedRef.current = false;
     const fb = fallbackFamily(run.fontName);
     // Caret lands at the clicked character (measured in the fallback font — close enough
     // to pick the index; it stays correct after the real font swaps in).
@@ -158,7 +364,17 @@ export const TextEditLayer: React.FC<TextEditLayerProps> = ({
     const run = runs[editing];
     const next = value;
     setEditing(null);
-    if (next !== run.text) onEditText?.(pageIndex, run, next);
+    if (next === run.text) return;
+
+    // FLOWING multi-line paragraph → full reflow (re-wrap + push-down). Single lines, headings,
+    // labels, and bullet/list items → the precise, low-risk replace_text path (no re-wrapping,
+    // so list structure and markers stay intact).
+    const para = modelRef.current?.paragraphs?.[run.paraIndex];
+    if (onReflowParagraph && isFlowingParagraph(para)) {
+      const payload = buildReflowPayload(modelRef.current, run, next);
+      if (payload) { onReflowParagraph(pageIndex, payload); return; }
+    }
+    onEditText?.(pageIndex, run, next);
   };
 
   const cancel = () => {
@@ -170,14 +386,33 @@ export const TextEditLayer: React.FC<TextEditLayerProps> = ({
 
   return (
     <div className="absolute top-0 left-0 z-[35]" style={{ width: `${width}px`, height: `${height}px` }}>
+      {/* Live reflow editor (wraps + pushes down as you type) — whole paragraph or one bullet item. */}
+      {paraEdit !== null && (
+        <ParagraphEditor
+          documentId={documentId}
+          pageIndex={pageIndex}
+          para={paraEdit.para}
+          pushColumnLeft={paraEdit.pushColumnLeft}
+          leadingOverride={paraEdit.leading}
+          alignOverride={paraEdit.align}
+          caretClick={caretClick}
+          heightPts={heightPts}
+          zoom={zoom}
+          pageWidthPx={width}
+          pageHeightPx={height}
+          onCommit={(payload) => { setParaEdit(null); onReflowParagraph?.(pageIndex, payload); }}
+          onCancel={() => setParaEdit(null)}
+        />
+      )}
+
       {/* Per-run hit targets (visible hint on hover). */}
-      {editing === null &&
+      {editing === null && paraEdit === null &&
         runs.map((r, i) => {
           const box = rectOf(r);
           return (
             <div
               key={i}
-              onClick={(e) => openEditor(i, e.nativeEvent.offsetX)}
+              onClick={(e) => openEditor(i, e.nativeEvent.offsetX, e.clientX, e.clientY)}
               className="absolute cursor-text hover:bg-[rgba(37,99,235,0.06)]"
               style={{ left: box.left, top: box.top, width: box.width, height: box.height }}
             />

gateway/app/routers/documents.py

@@ -298,6 +298,22 @@ def get_font_bytes(document_id: str, internal_font_id: str) -> Response:
     )
 
 
+@router.get("/{document_id}/raw")
+def get_document_raw(document_id: str) -> Response:
+    """Raw PDF bytes of the current version — loaded into the in-browser WASM engine for the
+    pixel-identical live-edit preview. Gated on copy permission (same as export)."""
+    doc_info = document_store.get_document(document_id)
+    if not doc_info:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Document not found")
+    perms = doc_info.get("permissions") or {}
+    if perms.get("canCopy", True) is False:
+        raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="Not permitted (canCopy).")
+    data = doc_info.get("bytes_data")
+    if not data:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Document bytes unavailable")
+    return Response(content=bytes(data), media_type="application/pdf")
+
+
 class SearchRect(BaseModel):
     x: float
     y: float

gateway/app/routers/edits.py

@@ -223,6 +223,41 @@ class ReplaceTextOperation(BaseModel):
     data: ReplaceTextData
 
 
+class ReflowRun(BaseModel):
+    text: str
+    internalFontId: str
+    fontSize: float
+    color: str = "#000000"
+
+
+class ReflowParagraphData(BaseModel):
+    objectIndices: list[int]
+    runs: list[ReflowRun]
+    columnLeft: float
+    columnRight: float
+    firstBaselineY: float
+    leading: float
+    oldLineCount: int = 1
+    align: Literal["left", "justify"] = "left"
+    # Push-down column left edge (defaults to columnLeft). For a bullet item the text reflows from
+    # a hanging indent (columnLeft) but the push-down spans the full width from pushColumnLeft so
+    # bullet markers and items below move together.
+    pushColumnLeft: float | None = None
+    # Optional WYSIWYG mode: exact visual line breaks from the live editor (one inner list of
+    # styled fragments per line). When present the engine emits these breaks verbatim.
+    lines: list[list[ReflowRun]] | None = None
+    # Optional WYSIWYG mode: exact visual line breaks from the live editor (one inner list of
+    # styled fragments per line). When present the engine emits these breaks verbatim.
+    lines: list[list[ReflowRun]] | None = None
+
+
+class ReflowParagraphOperation(BaseModel):
+    id: str
+    type: Literal["reflow_paragraph"]
+    pageIndex: int = Field(..., ge=0)
+    data: ReflowParagraphData
+
+
 class DecorationData(BaseModel):
     # Text-markup annotation (underline/strikeout/squiggly) over one or more text
     # lines — quadpoints in PDF top-down space, mirroring HighlightData.
@@ -264,6 +299,7 @@ EditOperation = Annotated[
     | DeleteAnnotationOperation
     | UpdateAnnotationOperation
     | ReplaceTextOperation
+    | ReflowParagraphOperation
     | UnderlineOperation
     | StrikeoutOperation
     | SquigglyOperation,
@@ -283,7 +319,7 @@ _OP_PERMISSION = {
     "squiggly": "canAnnotate", "comment": "canAnnotate", "freehand": "canAnnotate",
     "free_text": "canAnnotate", "text_overlay": "canAnnotate", "image_overlay": "canAnnotate",
     "delete_annotation": "canAnnotate", "update_annotation": "canAnnotate",
-    "replace_text": "canModify", "redaction": "canModify",
+    "replace_text": "canModify", "reflow_paragraph": "canModify", "redaction": "canModify",
     "update_field": "canFillForms",
     "page_rotation": "canAssemble", "page_deletion": "canAssemble", "page_reorder": "canAssemble",
 }
@@ -367,7 +403,7 @@ def apply_edits_impl(document_id: str, request: EditsRequest):
 
         # Redaction and in-place text rewrites mutate existing objects, which do
         # not round-trip cleanly through an incremental save — force a full save.
-        full_save_types = {"redaction", "replace_text"}
+        full_save_types = {"redaction", "replace_text", "reflow_paragraph"}
         needs_full = any(op.get("type") in full_save_types for op in req_dict.get("operations", []))
         new_bytes = doc_copy.save_full() if needs_full else doc_copy.save_incremental()
 

tests/edits/test_reflow_paragraph.py

@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+"""Tests for `reflow_paragraph` — Word/Adobe-style paragraph re-layout.
+
+Editing paragraph text re-wraps it within the column width, re-justifies, and pushes the
+following content down/up by the line-count delta. This is what fixes the justified-text
+"word collision" that the old line-level shift produced.
+
+Run: gateway/.venv/Scripts/python.exe tests/edits/test_reflow_paragraph.py
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT / "gateway"))
+import pdfengine  # noqa: E402
+
+
+def _build_pdf() -> bytes:
+    # A 3-line paragraph (Helvetica 11, baselines 700/686/672) + a footer line at y=600.
+    content = (
+        b"BT /F1 11 Tf 72 700 Td (The quick brown fox jumps over the lazy) Tj ET\n"
+        b"BT /F1 11 Tf 72 686 Td (dog near the river bank on a sunny) Tj ET\n"
+        b"BT /F1 11 Tf 72 672 Td (afternoon in early spring.) Tj ET\n"
+        b"BT /F1 11 Tf 72 600 Td (FOOTER LINE) Tj ET\n"
+    )
+    objs = [
+        b"<< /Type /Catalog /Pages 2 0 R >>",
+        b"<< /Type /Pages /Kids [3 0 R] /Count 1 >>",
+        b"<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] "
+        b"/Resources << /Font << /F1 5 0 R >> >> /Contents 4 0 R >>",
+        b"<< /Length %d >>\nstream\n" % len(content) + content + b"endstream",
+        b"<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>",
+    ]
+    pdf = b"%PDF-1.7\n"
+    offs = []
+    for i, o in enumerate(objs, 1):
+        offs.append(len(pdf))
+        pdf += b"%d 0 obj\n" % i + o + b"\nendobj\n"
+    xref = len(pdf)
+    pdf += b"xref\n0 %d\n0000000000 65535 f \n" % (len(objs) + 1)
+    for o in offs:
+        pdf += b"%010d 00000 n \n" % o
+    pdf += b"trailer\n<< /Size %d /Root 1 0 R >>\nstartxref\n%d\n%%%%EOF\n" % (len(objs) + 1, xref)
+    return pdf
+
+
+def _para_text(p) -> str:
+    return " ".join("".join(r.text for r in ln.runs) for ln in p.lines)
+
+
+def _footer_baseline(model):
+    for p in model.paragraphs:
+        for ln in p.lines:
+            if "FOOTER" in "".join(r.text for r in ln.runs):
+                return ln.baseline_y
+    return None
+
+
+def _reflow(longer: bool):
+    doc = pdfengine.PdfDocument.load_from_memory(_build_pdf(), "")
+    m = doc.get_page(0).extract_document_model()
+    para = m.paragraphs[0]
+    obj_idxs = [oi for ln in para.lines for r in ln.runs for oi in r.object_indices]
+    col_left = min(ln.x for ln in para.lines)
+    col_right = max(ln.x + ln.w for ln in para.lines)
+    first_baseline = max(ln.baseline_y for ln in para.lines)
+    bls = sorted((ln.baseline_y for ln in para.lines), reverse=True)
+    leading = abs(bls[0] - bls[1])
+
+    if longer:
+        new_text = ("The quick brown fox jumps over the lazy dog near the river bank on a sunny "
+                    "afternoon in early spring while birds sing softly and the gentle breeze "
+                    "carries the scent of fresh blossoms across the meadow.")
+    else:
+        new_text = "The quick brown fox."
+
+    op = {"version": "1.0", "operations": [{
+        "id": "r1", "type": "reflow_paragraph", "pageIndex": 0, "data": {
+            "objectIndices": obj_idxs,
+            "runs": [{"text": new_text, "internalFontId": para.lines[0].runs[0].internal_font_id,
+                      "fontSize": 11.0, "color": "#000000"}],
+            "columnLeft": col_left, "columnRight": col_right, "firstBaselineY": first_baseline,
+            "leading": leading, "oldLineCount": len(para.lines), "align": "justify"}}]}
+    doc.apply_edits(json.dumps(op))
+    out = doc.save_full()
+    m2 = pdfengine.PdfDocument.load_from_memory(out, "").get_page(0).extract_document_model()
+    return m2, leading, len(para.lines)
+
+
+def test_grow_wraps_and_pushes_down():
+    m2, leading, old_lines = _reflow(longer=True)
+    new_lines = len(m2.paragraphs[0].lines)
+    assert new_lines > old_lines, f"expected more lines after growing text, got {new_lines} (was {old_lines})"
+    footer = _footer_baseline(m2)
+    expected = 600 - (new_lines - old_lines) * leading
+    assert footer is not None and abs(footer - expected) < 2.0, \
+        f"footer should be pushed to ~{expected:.0f}, got {footer}"
+    print(f"  ok  grow: {old_lines}->{new_lines} lines, footer 600->{footer:.0f} (pushed {(new_lines-old_lines)*leading:.0f})")
+
+
+def test_shrink_pulls_up():
+    m2, leading, old_lines = _reflow(longer=False)
+    new_lines = len(m2.paragraphs[0].lines)
+    assert new_lines <= old_lines
+    footer = _footer_baseline(m2)
+    expected = 600 - (new_lines - old_lines) * leading  # new<old -> negative delta -> footer moves UP
+    assert footer is not None and abs(footer - expected) < 2.0, \
+        f"footer should pull up to ~{expected:.0f}, got {footer}"
+    print(f"  ok  shrink: {old_lines}->{new_lines} lines, footer 600->{footer:.0f}")
+
+
+def test_no_word_overlap():
+    # After a justified reflow, words on a line must not overlap (the old bug collided them).
+    m2, _, _ = _reflow(longer=True)
+    for ln in m2.paragraphs[0].lines:
+        runs = sorted(ln.runs, key=lambda r: r.x)
+        for a, b in zip(runs, runs[1:]):
+            assert a.x + a.w <= b.x + 1.0, f"overlap: run ends {a.x + a.w:.1f} > next start {b.x:.1f}"
+    print("  ok  no-overlap: justified lines have monotonic, non-colliding runs")
+
+
+def main() -> int:
+    try:
+        test_grow_wraps_and_pushes_down()
+        test_shrink_pulls_up()
+        test_no_word_overlap()
+    except AssertionError as exc:
+        print(f"  FAIL: {exc}")
+        return 1
+    print("\n3/3 passed.")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

wasm/CMakeLists.txt

@@ -43,7 +43,6 @@ message(STATUS "  Output ............... ${CMAKE_BINARY_DIR}/bin/hello.mjs (+ he
 # Phase 1/2 WASM Engine target
 add_executable(pdfengine_wasm
     bindings/wasm_engine.cpp
-    bindings/pdf_engine_facade.cpp
 )
 
 set_target_properties(pdfengine_wasm PROPERTIES
@@ -62,9 +61,10 @@ target_link_options(pdfengine_wasm PRIVATE
     "-sMODULARIZE=1"
     "-sEXPORT_ES6=1"
     "-sENVIRONMENT=node,web"
-    "-sFILESYSTEM=0"
     "-sALLOW_MEMORY_GROWTH=1"
-    "-sEXPORTED_FUNCTIONS=['_loadDocument','_renderPage','_freeDocument','_engineBuildInfo','_engineHasSkia','_malloc','_free','_getPageTextJson']"
+    "-sWASM_BIGINT"                 # PDFium uses i64
+    "-sSTACK_SIZE=5MB"              # PDFium render is stack-heavy
+    "-sEXPORTED_FUNCTIONS=['_loadDocument','_pageCount','_renderPagePng','_previewRender','_lastRenderPtr','_lastRenderW','_lastRenderH','_lastLayoutJson','_freeDocument','_engineBuildInfo','_malloc','_free']"
     "-sEXPORTED_RUNTIME_METHODS=['ccall','cwrap','getValue','setValue','HEAPU8']"
 )
 

wasm/bindings/wasm_engine.cpp

@@ -1,49 +1,88 @@
+// Real engine bindings for WASM: load a PDF, render a page (PNG), and render a PREVIEW of
+// an edit (e.g. reflow_paragraph) without mutating the loaded document. Rendering uses the
+// SAME C++ engine path as the gateway (PdfDocument::render), so the browser preview is
+// pixel-identical to the server-rendered page.
 #include <emscripten/emscripten.h>
-#include <cstdint>
+#include <pdfengine/pdf_document.hpp>
+#include <unordered_map>
+#include <memory>
+#include <vector>
 #include <string>
-#include "pdf_engine_facade.hpp"
+#include <cstdint>
 
-// Global facade instance
-static PdfEngineFacade g_facade;
+namespace {
+struct DocEntry {
+    std::shared_ptr<pdfengine::PdfDocument> doc;
+    std::vector<uint8_t> bytes; // kept so a preview can re-load a clean copy
+};
+std::unordered_map<int, DocEntry> g_docs;
+int g_nextHandle = 1;
+
+// Holds the most recent render output (PNG) so JS can read it via lastRender* accessors.
+std::vector<uint8_t> g_lastPng;
+int g_lastW = 0, g_lastH = 0;
+// Per-character caret layout (JSON) from the most recent preview reflow, read via lastLayoutJson().
+std::string g_lastLayout;
+
+int renderInto(pdfengine::PdfDocument& doc, int pageIndex, int dpi) {
+    auto page = doc.getPage(pageIndex);
+    if (!page) return -1;
+    auto img = (*page)->render(dpi);
+    if (!img) return -1;
+    g_lastPng = std::move(img->data);
+    g_lastW = img->width;
+    g_lastH = img->height;
+    return static_cast<int>(g_lastPng.size());
+}
+} // namespace
 
 extern "C" {
 
 EMSCRIPTEN_KEEPALIVE int loadDocument(const uint8_t* buffer, int size) {
-    return g_facade.loadDocument(buffer, size);
+    std::vector<uint8_t> bytes(buffer, buffer + size);
+    auto res = pdfengine::PdfDocument::loadFromMemory(bytes, "");
+    if (!res) return -1;
+    int h = g_nextHandle++;
+    g_docs[h] = DocEntry{*res, std::move(bytes)};
+    return h;
 }
 
-EMSCRIPTEN_KEEPALIVE int renderPage(int docHandle, int pageIndex, float scale, uint8_t* outputBuffer, int width, int height) {
-    return g_facade.renderPage(docHandle, pageIndex, scale, outputBuffer, width, height) ? 1 : 0;
+EMSCRIPTEN_KEEPALIVE int pageCount(int handle) {
+    auto it = g_docs.find(handle);
+    return it == g_docs.end() ? -1 : it->second.doc->pageCount();
 }
 
-EMSCRIPTEN_KEEPALIVE void freeDocument(int docHandle) {
-    g_facade.freeDocument(docHandle);
+// Render the loaded page as-is (PNG). Returns byte length (read via lastRenderPtr()).
+EMSCRIPTEN_KEEPALIVE int renderPagePng(int handle, int pageIndex, int dpi) {
+    auto it = g_docs.find(handle);
+    if (it == g_docs.end()) return -1;
+    return renderInto(*it->second.doc, pageIndex, dpi);
 }
 
-EMSCRIPTEN_KEEPALIVE const char* engineBuildInfo() {
-    return PdfEngineFacade::buildInfo();
+// Render a PREVIEW of an edit: load a fresh copy from the original bytes, apply the edits
+// (e.g. a reflow_paragraph op), render the page — the loaded document is untouched.
+EMSCRIPTEN_KEEPALIVE int previewRender(int handle, int pageIndex, int dpi, const char* editsJson) {
+    auto it = g_docs.find(handle);
+    if (it == g_docs.end()) return -1;
+    auto fresh = pdfengine::PdfDocument::loadFromMemory(it->second.bytes, "");
+    if (!fresh) return -1;
+    g_lastLayout.clear();
+    if (editsJson && editsJson[0]) {
+        auto r = (*fresh)->applyEdits(editsJson);
+        if (!r) return -1;
+        g_lastLayout = (*fresh)->lastReflowLayout();
+    }
+    return renderInto(**fresh, pageIndex, dpi);
 }
 
-EMSCRIPTEN_KEEPALIVE int engineHasSkia() {
-    return PdfEngineFacade::hasSkia() ? 1 : 0;
-}
+EMSCRIPTEN_KEEPALIVE const uint8_t* lastRenderPtr() { return g_lastPng.data(); }
+EMSCRIPTEN_KEEPALIVE int lastRenderW() { return g_lastW; }
+EMSCRIPTEN_KEEPALIVE int lastRenderH() { return g_lastH; }
+// Caret layout JSON for the most recent previewRender (empty if it had no reflow).
+EMSCRIPTEN_KEEPALIVE const char* lastLayoutJson() { return g_lastLayout.c_str(); }
 
-EMSCRIPTEN_KEEPALIVE const char* getDocumentFonts(int docHandle, int startPage, int endPage) {
-    static std::string s_buf;
-    s_buf = g_facade.getDocumentFonts(docHandle, startPage, endPage);
-    return s_buf.c_str();
-}
+EMSCRIPTEN_KEEPALIVE void freeDocument(int handle) { g_docs.erase(handle); }
 
-EMSCRIPTEN_KEEPALIVE const char* getPageFonts(int docHandle, int pageIndex) {
-    static std::string s_buf;
-    s_buf = g_facade.getPageFonts(docHandle, pageIndex);
-    return s_buf.c_str();
-}
-
-EMSCRIPTEN_KEEPALIVE const char* getPageTextJson(int docHandle, int pageIndex) {
-    static std::string s_buf;
-    s_buf = g_facade.getPageTextJson(docHandle, pageIndex);
-    return s_buf.c_str();
-}
+EMSCRIPTEN_KEEPALIVE const char* engineBuildInfo() { return "pdfengine-wasm+pdfium"; }
 
 }