fix(sync): capture block context across inline splits (drop-caps)

getTextFromXPointer returns a single text node, so a position on
<p><span>C</span>onvergence of Heaven and Earth</p> sent context 'C'.
The server text search matched the first 'C' in the chapter and stored
doc-start, reopening the web reader at the chapter beginning.

When the cleaned node text is tiny (<15 runes), walk up to 3 levels to
enclosing element pointers and return the first block text >=15 runes
(keeping the longest seen as fallback). Kindle-light: at most 3 extra
local reads, only on the tiny-text path, zero extra network. Char-offset
slicing applies to the original node text only, not parent blocks.

main.lua

@@ -764,23 +764,81 @@ function Bookhoard:getContextText()
     local xp = self:getLastProgress()
     if not xp then return "" end
 
-    local text = self.ui.document:getTextFromXPointer(xp)
-    if not text or text == "" then
-        text = self.ui.document:getTextFromXPointer(self.ui.document:getNormalizedXPointer(xp))
+    local function clean(s)
+        if not s or s == "" then return "" end
+        if #s > 100 then
+            s = s:sub(1, 100)
+        end
+        return s:gsub("%s+", " "):match("^%s*(.-)%s*$") or ""
+    end
+
+    local function tryPointer(p)
+        if not p or p == "" then return "" end
+        local ok, text = pcall(function()
+            return self.ui.document:getTextFromXPointer(p)
+        end)
+        if not ok then return "" end
+        return text or ""
+    end
+
+    local text = tryPointer(xp)
+    if text == "" then
+        local ok, nxp = pcall(function()
+            return self.ui.document:getNormalizedXPointer(xp)
+        end)
+        if ok and nxp and nxp ~= "" and nxp ~= xp then
+            text = tryPointer(nxp)
+        end
+    end
+    if text == "" then return "" end
+
+    local cleaned = clean(text)
+    -- Single text nodes split by inline markup (e.g. drop-caps like
+    -- <p><span>C</span>onvergence…</p>) return just "C" here. That one
+    -- character then false-positives the server text search to the first
+    -- "C" in the chapter (doc start). Walk up to the enclosing block
+    -- element(s) so we capture words, not the node. Kindle-light: at most
+    -- 3 extra local reads, only on this tiny-text path, zero extra network.
+    local function runeLen(s)
+        -- Cheap UTF-8 aware length without dependencies.
+        local _, n = s:gsub("[^\128-\191]", "")
+        return n
+    end
+    if cleaned ~= "" and runeLen(cleaned) < 15 then
+        local parent = xp
+        for _ = 1, 3 do
+            -- Strip one trailing path segment: "/text().N" or "/span[1]" etc.
+            local stripped = parent:gsub("/[^/]+$", "")
+            if not stripped or stripped == "" or stripped == parent then
+                break
+            end
+            parent = stripped
+            -- Don't walk past the document body; element pointers above the
+            -- block would return whole-chapter text.
+            if parent:match("/body%s*$") or parent:match("DocFragment%[%d+%]$") then
+                break
+            end
+            local ptext = tryPointer(parent)
+            if ptext and ptext ~= "" then
+                local pcleaned = clean(ptext)
+                if pcleaned ~= "" and runeLen(pcleaned) >= 15 then
+                    return pcleaned
+                end
+                -- Keep the longest thing seen in case no level reaches 15.
+                if runeLen(pcleaned) > runeLen(cleaned) then
+                    cleaned = pcleaned
+                end
+            end
+        end
+        return cleaned
     end
-    if not text or text == "" then return "" end
 
     local char_offset = tonumber(xp:match("text%(%)%.?(%d+)")) or 0
     if char_offset > 0 and char_offset < #text then
         text = text:sub(char_offset + 1)
     end
 
-    if #text > 100 then
-        text = text:sub(1, 100)
-    end
-
-    text = text:gsub("%s+", " "):match("^%s*(.-)%s*$") or ""
-    return text
+    return clean(text)
 end
 
 function Bookhoard:collectBookData()