M2: GUI thread + host window trong juce_fx_bridge (Reaper-style — editor cua CHINH instance DSP)

JuceFxGuiHost: GUI thread rieng chay juce MessageManager runDispatchLoopUntil(10) + tick
30ms inline cadence (JUCE Timer KHONG fire tren GUI thread rieng — da xac minh selftest),
PluginWindow (DocumentWindow native 'FX slot N') attach editor, reconcile dong editor+
window, teardown thu tu. runOnGui(): plugin instance phai tao tren GUI thread — JUCE plugin
gan MM vao thread tao instance, editor tao thread khac -> treo (M1 chi chay vi main thread
vua tao instance vua la message thread); prepare/shutdown bao qua runOnGui.
JuceFxLoop: guiHost.start() TRUOC engine.prepare; re-prepare sau waitEditorsClosed(1500);
teardown runOnGui(shutdown) truoc stop(). Idle ~0 CPU (khong feeder, khong instance 2).
juce_fx_gui_selftest (khong ship) PASS tren purecomp + fast-compressor: open/hide/show/
resize/close, teardown sach; window native chup duoc (PrintWindow). Regression M1 selftest OK.

native_bridge/juce_fx/JuceFxLoop.cpp

@@ -6,6 +6,7 @@
 // Validate header mỗi iteration: sampleRate/blockSize đổi giữa chừng →
 // teardown + prepareToPlay lại + report latency mới qua FxLatReport.
 #include "JuceFxEngine.h"
+#include "JuceFxGuiHost.h"
 #include "FxRealtimeIPC.h"
 #include "FxShm.h"
 
@@ -132,7 +133,16 @@ int run_juce_fx_loop(const std::string& jobPath, const std::string& shmName,
     engine.setChain(chainJson);
     if (ipc->h.sampleRate) curSr = ipc->h.sampleRate;
     if (ipc->h.blockSize) curBlock = ipc->h.blockSize;
-    engine.prepare(curSr, curBlock);
+    // M2: GUI thread + host window (Reaper-style — editor của CHÍNH instance
+    // DSP trong graph, không instance 2 / feeder). GUI thread cũng là thread
+    // tạo plugin instance: JUCE plugin gắn MM vào thread tạo instance — editor
+    // tạo trên thread khác sẽ treo (selftest: createEditorIfNeeded hang khi
+    // instance tạo trên main). Vì vậy start() TRƯỚC prepare, prepare chạy qua
+    // runOnGui; main thread sau này chỉ engine.process (JUCE cho phép
+    // cross-thread process như host chuẩn).
+    JuceFxGuiHost guiHost(engine);
+    guiHost.start();
+    guiHost.runOnGui([&] { engine.prepare(curSr, curBlock); });
     std::cerr << "[JuceFxLoop] prepared sr=" << curSr << " block=" << curBlock
               << " latency=" << engine.latencySamples() << std::endl;
 
@@ -160,6 +170,8 @@ int run_juce_fx_loop(const std::string& jobPath, const std::string& shmName,
     ipc->h.state = FXRT_STATE_READY;
     std::cerr << "[JuceFxLoop] ready" << std::endl;
 
+    // (guiHost đã start TRƯỚC engine.prepare ở trên — instance phải tạo trên
+    // GUI thread.) Editor chỉ mở khi có lệnh (M3 ctrl ring); idle ≈ 0 CPU.
     std::thread hb([&]() {
         while (ipc->h.running && ipc->h.state == FXRT_STATE_READY) {
             ipc->h.heartbeat++;
@@ -192,7 +204,11 @@ int run_juce_fx_loop(const std::string& jobPath, const std::string& shmName,
                                   ipc->h.blockSize != curBlock)) {
             curSr = ipc->h.sampleRate;
             curBlock = ipc->h.blockSize;
-            engine.prepare(curSr, curBlock);
+            // M2: editor là JUCE Component — đóng trên GUI thread TRƯỚC khi
+            // re-prepare; prepare() chạy trên GUI thread (thread tạo instance —
+            // xem startup, M2b).
+            guiHost.waitEditorsClosed(1500);
+            guiHost.runOnGui([&] { engine.prepare(curSr, curBlock); });
             reportLatencies();
             std::cerr << "[JuceFxLoop] re-prepared sr=" << curSr
                       << " block=" << curBlock
@@ -270,7 +286,8 @@ int run_juce_fx_loop(const std::string& jobPath, const std::string& shmName,
     }
     ipc->h.state = FXRT_STATE_STARTING;
     hb.join();
-    engine.shutdown();
+    guiHost.runOnGui([&] { engine.shutdown(); });  // graph clear trên GUI thread
+    guiHost.stop();      // M2: đóng editor (GUI thread) + dừng GUI thread + join
     fxshm::closeShm(v);
 #ifdef _WIN32
     timeEndPeriod(1);