diff --git a/server/src/global/globals.h b/server/src/global/globals.h
index 31823a44..ab228475 100644
--- a/server/src/global/globals.h
+++ b/server/src/global/globals.h
@@ -12,6 +12,7 @@
 #include <vector>
 #include <windows.h>
 #include "engine/core/word_item.h"
+#include "ipc/candidate_render_sync.h"
 #include "window/candidate_view_model.h"
 
 namespace GlobalIme
@@ -256,7 +257,10 @@ inline void PublishRenderedCandidatePageGeneration(std::uint64_t generation)
 {
     {
         std::lock_guard lock(candidate_render_mutex);
-        rendered_candidate_page_generation.store(generation, std::memory_order_release);
+        rendered_candidate_page_generation.store(
+            FanyImeIpc::AdvanceRenderedCandidateGeneration(
+                rendered_candidate_page_generation.load(std::memory_order_relaxed), generation),
+            std::memory_order_release);
     }
     candidate_render_cv.notify_all();
 }
diff --git a/server/src/ipc/candidate_render_sync.h b/server/src/ipc/candidate_render_sync.h
index 9f4bac36..ffac3d35 100644
--- a/server/src/ipc/candidate_render_sync.h
+++ b/server/src/ipc/candidate_render_sync.h
@@ -36,4 +36,13 @@ constexpr bool ShouldWaitForCandidateRender(std::uint64_t rendered, std::uint64_
     }
     return rendered < current;
 }
+
+// The render echo only moves forward. WebView2 echoes from async DOM callbacks, so a content-only
+// update's callback can land after a dedup of the same content already echoed the newer generation;
+// storing the older one left rendered one behind published and every selection waited out the full
+// bound.
+constexpr std::uint64_t AdvanceRenderedCandidateGeneration(std::uint64_t rendered, std::uint64_t echoed) noexcept
+{
+    return echoed > rendered ? echoed : rendered;
+}
 } // namespace FanyImeIpc
diff --git a/server/src/window/ime_windows.cpp b/server/src/window/ime_windows.cpp
index 13fead4d..e6c53990 100644
--- a/server/src/window/ime_windows.cpp
+++ b/server/src/window/ime_windows.cpp
@@ -2492,7 +2492,16 @@ LRESULT CALLBACK WndProcCandWindow(HWND hwnd, UINT message, WPARAM wParam, LPARA
         {
             CAND_DIAG_LOGF(L"show folded into in-flight fine-tune caret=({},{})", Global::Point[0], Global::Point[1]);
             CAND_DIAG_LOGF(L"candidate-frame path=folded-inflight content_gen={}", contentGeneration);
-            InflateCandWnd(str);
+            // The in-flight FineTune echoes only the page it captured, which may predate this one. Without
+            // an echo here a digit/space selection right after a fast burst waits out the full render-sync
+            // timeout, felt as a stall on commit.
+            InflateCandWnd(str, [contentGeneration, pageGeneration = candidatePage->generation]() {
+                if (!::is_global_wnd_cand_shown || contentGeneration != g_candidate_content_generation.load())
+                {
+                    return;
+                }
+                Global::PublishRenderedCandidatePageGeneration(pageGeneration);
+            });
             KillTimer(hwnd, TIMER_ID_CANDIDATE_MOVE_SETTLE);
             SetTimer(hwnd, TIMER_ID_CANDIDATE_MOVE_SETTLE, kCandidateMoveSettleMs, nullptr);
             return 0;
diff --git a/server/tests/src/test_candidate_render_sync.cpp b/server/tests/src/test_candidate_render_sync.cpp
index b2a9edc2..0e3fdbbc 100644
--- a/server/tests/src/test_candidate_render_sync.cpp
+++ b/server/tests/src/test_candidate_render_sync.cpp
@@ -22,6 +22,23 @@ TEST_CASE(candidate_render_sync_waits_only_when_a_painted_page_lags_the_publishe
     REQUIRE(!ShouldWaitForCandidateRender(8, 7, false, true));
 }
 
+TEST_CASE(candidate_render_echo_never_moves_backwards)
+{
+    using FanyImeIpc::AdvanceRenderedCandidateGeneration;
+    using FanyImeIpc::ShouldWaitForCandidateRender;
+
+    // WebView2: a dedup echoes 332 synchronously, then the content-only DOM callback for the same
+    // content lands and echoes 331. The stale echo must not pull rendered back behind published.
+    std::uint64_t rendered = 330;
+    rendered = AdvanceRenderedCandidateGeneration(rendered, 332);
+    rendered = AdvanceRenderedCandidateGeneration(rendered, 331);
+    REQUIRE_EQ(rendered, 332u);
+    REQUIRE(!ShouldWaitForCandidateRender(rendered, 332, false, true));
+
+    // A newer echo still advances.
+    REQUIRE_EQ(AdvanceRenderedCandidateGeneration(332, 333), 333u);
+}
+
 TEST_CASE(candidate_render_sync_wait_is_bounded_to_a_visible_stall)
 {
     // The bound is a UX contract and a transport constraint: a wedged UI thread may delay a