fix(fxrt): chan out ring overflow trong bridge (backpressure, het pointer collision)

native_bridge/src/RealtimeFxLoop.cpp

@@ -205,7 +205,7 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
     uint64_t processed = 0;
     uint64_t lastGen = 0;  // chain generation đã báo latency
     // ---- FXRT_PERF instrumentation (tạm, đo crackle) ----
-    uint64_t perfIter = 0, perfProc = 0, perfIdle = 0;
+    uint64_t perfIter = 0, perfProc = 0, perfIdle = 0, perfOutFull = 0;
     double perfProcSum = 0.0, perfProcMax = 0.0, perfProcMin = 1e9;
     double perfLoopSum = 0.0, perfLoopMax = 0.0;
     uint32_t perfTakeSum = 0;
@@ -264,6 +264,21 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
             std::cerr << "[RealtimeFxLoop] latency reported: " << lats.size()
                       << " slot(s)" << std::endl;
         }
+        // OUT-ring backpressure (fix crackle "wobble" — pointer collision):
+        // kiểm tra chỗ trống TRƯỚC khi consume input. Nếu out ring đầy (engine
+        // pump chậm/stall: WS send bị block khi main-thread browser busy) mà
+        // bridge vẫn ghi → ghi đè slot engine đang đọc = read/write pointer
+        // giẫm nhau → mẫu torn → méo "wobble" kéo dài tới khi reset. Khi đầy:
+        // KHÔNG xử lí batch này (nhường engine drain), input backlog dồn về
+        // engine → engine drop batch cũ nhất (drop sạch, worklet hold+fade che
+        // gap) — không bao giờ ghi đè out slot chưa đọc.
+        const uint32_t outFree = ipc->h.outSlots - (ipc->h.outWrite - ipc->h.outRead);
+        if (outFree < std::min<uint32_t>(ipc->h.inWrite - ipc->h.inRead, FXRT_IN_SLOTS)) {
+            ++perfOutFull;
+            std::this_thread::sleep_until(std::chrono::steady_clock::now() +
+                                          std::chrono::milliseconds(1));
+            continue;
+        }
         uint32_t avail = ipc->h.inWrite - ipc->h.inRead;
         if (avail == 0) {
             ++perfIdle;
@@ -315,12 +330,13 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
                       << " loopAvg=" << (perfIter ? perfLoopSum / perfIter : 0.0) << "ms"
                       << " loopMax=" << perfLoopMax << "ms"
                       << " idle=" << perfIdle
+                      << " outFull=" << perfOutFull
                       << " takeAvg=" << (perfProc ? (double)perfTakeSum / perfProc : 0.0)
                       << " procBlocks=" << processed
                       << " rate=" << (runSec > 0 ? processed / runSec : 0.0) << "blk/s"
                       << " outDepth=" << (ipc->h.outWrite - ipc->h.outRead)
                       << std::endl;
-            perfProc = 0; perfProcSum = 0.0; perfTakeSum = 0; perfIdle = 0;
+            perfProc = 0; perfProcSum = 0.0; perfTakeSum = 0; perfIdle = 0; perfOutFull = 0;
         }
     }
     ipc->h.state = FXRT_STATE_STARTING;  // đã dừng