diff --git a/TASKS_DAW_A.md b/TASKS_DAW_A.md
index e467278..def019e 100644
--- a/TASKS_DAW_A.md
+++ b/TASKS_DAW_A.md
@@ -12,15 +12,15 @@ Trạng thái: `[x]` hoàn thành, `[ ]` chưa. Cập nhật sau mỗi phase.
 - [x] 1.7 Pytest: chain hợp nhất → job.json đúng schema; SET_PARAM/REPORT_LATENCY round-trip.
 
 ## Phase 2 — Bridge C++: BuiltinFxChain (8 DSP)
-- [ ] 2.1 File `native_bridge/src/BuiltinFxChain.{h,cpp}` — khung chain serial + SEH guard.
-- [ ] 2.2 Port EQ 4-band (eq) từ Python `_apply_eq4`.
-- [ ] 2.3 Port EQ Pro (eqpro) RBJ 8-band từ Python `_apply_eqpro`.
-- [ ] 2.4 Port compressor/limiter từ Python.
-- [ ] 2.5 Port exciter/rebalance từ Python.
-- [ ] 2.6 Port imager/maximizer từ JS WebAudio (MASTER_MODULE_IO) — Python không có.
-- [ ] 2.7 `RealtimeFxChain::buildChain` + RenderFxJob: nhận builtin slot → chạy BuiltinFxChain xen kẽ VST3 đúng thứ tự.
-- [ ] 2.8 SET_PARAM/REPORT_LATENCY trong bridge (SHM control ring).
-- [ ] 2.9 Golden test: C++ vs Python `_apply_builtin_fx_chain` (SNR/diff ngưỡng) + order test xen kẽ.
+- [x] 2.1 File `native_bridge/src/BuiltinFxChain.{h,cpp}` — khung chain serial + SEH guard.
+- [x] 2.2 Port EQ 4-band (eq) từ Python `_apply_eq4`.
+- [x] 2.3 Port EQ Pro (eqpro) RBJ 8-band từ Python `_apply_eqpro`.
+- [x] 2.4 Port compressor/limiter từ Python.
+- [x] 2.5 Port exciter/rebalance từ Python.
+- [x] 2.6 Port imager/maximizer từ JS WebAudio (MASTER_MODULE_IO) — Python không có.
+- [x] 2.7 `RealtimeFxChain::buildChain` + RenderFxJob: nhận builtin slot → chạy BuiltinFxChain xen kẽ VST3 đúng thứ tự.
+- [x] 2.8 SET_PARAM/REPORT_LATENCY trong bridge (SHM control ring).
+- [x] 2.9 Golden test: C++ vs Python `_apply_builtin_fx_chain` (SNR/diff ngưỡng) + order test xen kẽ.
 
 ## Phase 3 — Engine Python: render + realtime chain hợp nhất
 - [ ] 3.1 `render_engine.py`: track/master đọc chain hợp nhất; Python DSP giữ làm fallback (RENDER_ENGINE≠bridge).
diff --git a/app/core/fx_realtime.py b/app/core/fx_realtime.py
index 30578bf..a21e2e8 100644
--- a/app/core/fx_realtime.py
+++ b/app/core/fx_realtime.py
@@ -4,9 +4,12 @@ spawn fx_vst_bridge --realtime-fx, đẩy/kéo block audio qua SHM ring.
 
 Layout mirror chính xác native_bridge/include/FxRealtimeIPC.h (mọi field
 u32/float 4-byte, không padding):
-  header 12 u32 (48B): magic state sampleRate blockSize running heartbeat
+  header 18 u32 (72B): magic state sampleRate blockSize running heartbeat
                        inWrite inRead inSlots outWrite outRead outSlots
+                       ctrlWrite ctrlRead ctrlSlots latWrite latRead latSlots
   inL[4][256] inR[4][256] outL[8][256] outR[8][256]
+  ctrl[8] (FxCtrlCmd 24B) — SET_PARAM ring (engine -> bridge)
+  lat[8] (FxLatReport 8B) — REPORT_LATENCY ring (bridge -> engine)
 
 Luồng dữ liệu: browser → WS → write_input() → [bridge xử lý] → read_output()
 → WS → browser. Block = FXRT_BLOCK=256 stereo float32 = 2048 bytes.
@@ -29,6 +32,8 @@ FXRT_MAGIC = 0x46585254
 FXRT_BLOCK = 256
 FXRT_IN_SLOTS = 4
 FXRT_OUT_SLOTS = 8
+FXRT_CTRL_SLOTS = 8
+FXRT_LAT_SLOTS = 8
 FXRT_STATE_STARTING = 0
 FXRT_STATE_READY = 1
 FXRT_STATE_ERROR = 2
@@ -48,15 +53,38 @@ class FxRTHeader(ctypes.Structure):
         ("out_write", ctypes.c_uint32),
         ("out_read", ctypes.c_uint32),
         ("out_slots", ctypes.c_uint32),
+        ("ctrl_write", ctypes.c_uint32),
+        ("ctrl_read", ctypes.c_uint32),
+        ("ctrl_slots", ctypes.c_uint32),
+        ("lat_write", ctypes.c_uint32),
+        ("lat_read", ctypes.c_uint32),
+        ("lat_slots", ctypes.c_uint32),
     ]
 
 
-HEADER_SIZE = ctypes.sizeof(FxRTHeader)          # 48
-IN_L_OFF = HEADER_SIZE                            # 48
+class FxCtrlCmd(ctypes.Structure):
+    _fields_ = [
+        ("slot", ctypes.c_uint32),
+        ("key", ctypes.c_char * 16),
+        ("value", ctypes.c_float),
+    ]
+
+
+class FxLatReport(ctypes.Structure):
+    _fields_ = [
+        ("slot", ctypes.c_uint32),
+        ("samples", ctypes.c_uint32),
+    ]
+
+
+HEADER_SIZE = ctypes.sizeof(FxRTHeader)          # 72
+IN_L_OFF = HEADER_SIZE                            # 72
 IN_R_OFF = IN_L_OFF + FXRT_IN_SLOTS * FXRT_BLOCK * 4
 OUT_L_OFF = IN_R_OFF + FXRT_IN_SLOTS * FXRT_BLOCK * 4
 OUT_R_OFF = OUT_L_OFF + FXRT_OUT_SLOTS * FXRT_BLOCK * 4
-TOTAL_SIZE = OUT_R_OFF + FXRT_OUT_SLOTS * FXRT_BLOCK * 4
+CTRL_OFF = OUT_R_OFF + FXRT_OUT_SLOTS * FXRT_BLOCK * 4
+LAT_OFF = CTRL_OFF + FXRT_CTRL_SLOTS * ctypes.sizeof(FxCtrlCmd)
+TOTAL_SIZE = LAT_OFF + FXRT_LAT_SLOTS * ctypes.sizeof(FxLatReport)
 
 
 class FxRealtimeSession:
@@ -81,6 +109,8 @@ class FxRealtimeSession:
         self.h.running = 1
         self.h.in_slots = FXRT_IN_SLOTS
         self.h.out_slots = FXRT_OUT_SLOTS
+        self.h.ctrl_slots = FXRT_CTRL_SLOTS
+        self.h.lat_slots = FXRT_LAT_SLOTS
         self._np = np.frombuffer(self.shm.buf, dtype=np.float32)
         self._lock = threading.Lock()
         self.proc = None
@@ -225,6 +255,28 @@ class FxRealtimeSession:
     def set_param(self, slot, key, value):
         with self._lock:
             self.pending_params.setdefault(slot, {})[str(key)] = value
+            # Phase 2.8: đẩy vào SHM control ring (bridge drain mỗi
+            # iteration). pending_params giữ làm fallback — fake test session
+            # (object.__new__) không có shm; drain_params vẫn hoạt động.
+            if getattr(self, "shm", None) is not None:
+                try:
+                    self._ctrl_enqueue_locked(slot, str(key), value)
+                except Exception:
+                    pass
+
+    def _ctrl_enqueue_locked(self, slot, key, value):
+        h = self.h
+        kb = key.encode("utf-8")[:15]
+        cmds = (FxCtrlCmd * FXRT_CTRL_SLOTS).from_buffer(self.shm.buf, CTRL_OFF)
+        while True:
+            if h.ctrl_write - h.ctrl_read < h.ctrl_slots:
+                cs = h.ctrl_write & (h.ctrl_slots - 1)
+                cmds[cs].slot = int(slot) & 0xFFFFFFFF
+                cmds[cs].key = kb + b"\00" * (16 - len(kb))
+                cmds[cs].value = float(value)
+                h.ctrl_write += 1
+                return
+            h.ctrl_read += 1  # ring đầy → drop lệnh cũ nhất
 
     def report_latency(self, slot, samples):
         with self._lock:
@@ -239,8 +291,23 @@ class FxRealtimeSession:
 
     def get_latencies(self):
         with self._lock:
+            self._lat_drain()
             return dict(self.latencies)
 
+    def _lat_drain(self):
+        shm = getattr(self, "shm", None)
+        if shm is None:
+            return
+        try:
+            h = self.h
+            lats = (FxLatReport * FXRT_LAT_SLOTS).from_buffer(shm.buf, LAT_OFF)
+            while h.lat_read < h.lat_write:
+                ls = h.lat_read & (h.lat_slots - 1)
+                self.latencies[int(lats[ls].slot)] = int(lats[ls].samples)
+                h.lat_read += 1
+        except Exception:
+            pass
+
 
 # Registry toàn tiến trình (desktop app 1 user).
 _SESSIONS = {}
diff --git a/native_bridge/CMakeLists.txt b/native_bridge/CMakeLists.txt
index b377e9d..3694b0a 100644
--- a/native_bridge/CMakeLists.txt
+++ b/native_bridge/CMakeLists.txt
@@ -106,6 +106,7 @@ add_executable(fx_vst_bridge
   src/main_fx.cpp
   src/RenderFxJob.cpp
   src/RealtimeFxLoop.cpp
+  src/BuiltinFxChain.cpp
   src/FxGuiServer.cpp
 )
 if(VST3_SDK_TARGET)
diff --git a/native_bridge/include/BuiltinFxChain.h b/native_bridge/include/BuiltinFxChain.h
new file mode 100644
index 0000000..4fd9754
--- /dev/null
+++ b/native_bridge/include/BuiltinFxChain.h
@@ -0,0 +1,47 @@
+// native_bridge/include/BuiltinFxChain.h
+// Builtin DSP chain (PLAN_DAW_A.md Phase 2): 8 custom FX được port từ Python
+// (render_engine.py `_apply_builtin_fx_chain`) + JS WebAudio (imager/maximizer,
+// MASTER_MODULE_IO). Serial, in-place, stereo — engine xử lý TOÀN chain; chạy
+// xen kẽ VST3 đúng thứ tự UI trong RenderFxJob (offline) + RealtimeFxChain.
+#pragma once
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+// Forward: sheredom JSON object (params). Khai báo để factory dùng.
+struct json_object_s;
+
+// Một FX builtin — stateful giữa các block (biquad z1/z2, env compressor...).
+class BuiltinFx {
+public:
+    virtual ~BuiltinFx() = default;
+    // Xử lý 1 block stereo in-place (L/R song song, interleaved-wise).
+    virtual void process(float* L, float* R, uint32_t n) = 0;
+    // SET_PARAM (SHM control ring, Phase 2.8): cập nhật tham số live.
+    // Trả false nếu slot không có param này (gọi từ realtime loop thread,
+    // cùng thread process → không cần lock).
+    virtual bool setParam(const std::string& key, double value) { (void)key; (void)value; return false; }
+};
+
+// Chain serial các BuiltinFx — giữ NGUYÊN thứ tự thêm vào.
+class BuiltinFxChain {
+public:
+    explicit BuiltinFxChain(double sampleRate) : sampleRate_(sampleRate) {}
+    void add(std::unique_ptr<BuiltinFx> fx, bool bypass);
+    // Chạy toàn chain serial, in-place.
+    void process(float* L, float* R, uint32_t n);
+    bool setParam(int slot, const std::string& key, double value);
+    size_t size() const { return entries_.size(); }
+private:
+    struct Entry { std::unique_ptr<BuiltinFx> fx; bool bypass = false; };
+    std::vector<Entry> entries_;
+    double sampleRate_;
+};
+
+// Factory: id ∈ eq|eqpro|imager|maximizer|compressor|limiter|exciter|rebalance
+// |gain|normalize (gain/normalize legacy giữ ở RenderFxJob — factory trả
+// nullptr cho 2 loại này). params = JSON object (sheredom), có thể null.
+std::unique_ptr<BuiltinFx> createBuiltinFx(const std::string& id,
+                                           const json_object_s* params,
+                                           double sampleRate);
diff --git a/native_bridge/include/FxRealtimeIPC.h b/native_bridge/include/FxRealtimeIPC.h
index 3e6b17d..97f9919 100644
--- a/native_bridge/include/FxRealtimeIPC.h
+++ b/native_bridge/include/FxRealtimeIPC.h
@@ -10,11 +10,13 @@
 #define FXRT_BLOCK 256
 #define FXRT_IN_SLOTS 4             // engine -> bridge
 #define FXRT_OUT_SLOTS 8            // bridge -> engine
+#define FXRT_CTRL_SLOTS 8           // SET_PARAM commands (engine -> bridge)
+#define FXRT_LAT_SLOTS 8            // REPORT_LATENCY reports (bridge -> engine)
 #define FXRT_STATE_STARTING 0
 #define FXRT_STATE_READY 1
 #define FXRT_STATE_ERROR 2
 
-// Header 48 bytes (12 u32) — giữ thứ tự khớp fx_realtime.py.
+// Header 72 bytes (18 u32) — giữ thứ tự khớp fx_realtime.py.
 struct FxRealtimeHeader {
     volatile uint32_t magic;
     volatile uint32_t state;
@@ -28,12 +30,35 @@ struct FxRealtimeHeader {
     volatile uint32_t outWrite;     // producer: bridge
     volatile uint32_t outRead;      // consumer: engine
     uint32_t outSlots;
+    volatile uint32_t ctrlWrite;    // producer: engine (SET_PARAM ring)
+    volatile uint32_t ctrlRead;     // consumer: bridge
+    uint32_t ctrlSlots;
+    volatile uint32_t latWrite;     // producer: bridge (REPORT_LATENCY ring)
+    volatile uint32_t latRead;      // consumer: engine
+    uint32_t latSlots;
 };
 
+// SET_PARAM: {slot_idx, key[16], value} — 24 bytes.
+struct FxCtrlCmd {
+    volatile uint32_t slot;
+    char key[16];
+    volatile float value;
+};
+static_assert(sizeof(FxCtrlCmd) == 24, "FxCtrlCmd must be 24 bytes");
+
+// REPORT_LATENCY: {slot_idx, samples} — 8 bytes.
+struct FxLatReport {
+    volatile uint32_t slot;
+    volatile uint32_t samples;
+};
+static_assert(sizeof(FxLatReport) == 8, "FxLatReport must be 8 bytes");
+
 struct FxRealtimeIPC {
     FxRealtimeHeader h;
     float inL[FXRT_IN_SLOTS][FXRT_BLOCK];
     float inR[FXRT_IN_SLOTS][FXRT_BLOCK];
     float outL[FXRT_OUT_SLOTS][FXRT_BLOCK];
     float outR[FXRT_OUT_SLOTS][FXRT_BLOCK];
+    FxCtrlCmd ctrl[FXRT_CTRL_SLOTS];
+    FxLatReport lat[FXRT_LAT_SLOTS];
 };
diff --git a/native_bridge/include/RenderFxJob.h b/native_bridge/include/RenderFxJob.h
index e06fff2..db3da4d 100644
--- a/native_bridge/include/RenderFxJob.h
+++ b/native_bridge/include/RenderFxJob.h
@@ -5,6 +5,7 @@
 #include <cstdint>
 #include <memory>
 #include <string>
+#include <vector>
 
 // Offline FX render mode (Phase 0 of PLAN_MASTERBUS_FX_RACK_VST.md):
 // `daw_vst_bridge --render-fx <job.json> --in <input.wav> --out <output.wav>`
@@ -72,6 +73,15 @@ public:
     // JSON: [{"path":"...","bypass":true|false},...] — chain entries replace
     // the previous chain atomically. May be called from any thread.
     void setChain(const std::string& json, double sampleRate, int32_t blockSize);
+    // SET_PARAM (Phase 2.8): áp dụng live vào builtin slot (slot index theo
+    // chain JSON). Gọi từ loop thread.
+    void setParam(int slot, const std::string& key, double value);
+    // Phase 2.8 (SHM control ring): chain generation — tăng mỗi lần worker
+    // swap chain; RealtimeFxLoop so gen → báo lại latency qua lat ring.
+    uint64_t chainGen();
+    // Per-slot latency samples (thứ tự khớp chain JSON): builtin = 0 (IIR
+    // zero-latency), VST3 = getLatencySamples (đọc lúc load) — PDC Phase 3.
+    std::vector<uint32_t> entryLatencies();
     // Audio thread only: run the chain in-place over one stereo block.
     void process(float* inL, float* inR, uint32_t n);
     void shutdown();
diff --git a/native_bridge/src/BuiltinFxChain.cpp b/native_bridge/src/BuiltinFxChain.cpp
new file mode 100644
index 0000000..151993c
--- /dev/null
+++ b/native_bridge/src/BuiltinFxChain.cpp
@@ -0,0 +1,530 @@
+// native_bridge/src/BuiltinFxChain.cpp
+// 8 builtin DSP (PLAN_DAW_A.md Phase 2) — port 1:1 từ:
+//   - Python app/core/render_engine.py `_apply_builtin_fx_chain` (6 loại:
+//     eq, eqpro, compressor, limiter, exciter, rebalance)
+//   - JS WebAudio MASTER_MODULE_IO (2 loại: imager, maximizer)
+// Biquad: RBJ cookbook, Direct Form II transposed (cùng công thức Python
+// `_rbj_*`). Block-wise, stateful giữa block — khớp lfilter full-file.
+#define _USE_MATH_DEFINES
+#include "BuiltinFxChain.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <string>
+#include <vector>
+
+#include "sheredom_json.h"
+
+namespace {
+
+// ── JSON helpers (mirror RenderFxJob.cpp anonymous-namespace versions) ──────
+const json_value_s* memberValue(const json_object_s* o, const char* name) {
+    for (const json_object_element_s* e = o->start; e; e = e->next) {
+        if (e->name && e->name->string && std::strcmp(e->name->string, name) == 0)
+            return e->value;
+    }
+    return nullptr;
+}
+double memberNumber(const json_object_s* o, const char* name, double def) {
+    const json_value_s* v = memberValue(o, name);
+    if (v && v->type == json_type_number) return std::atof(static_cast<const json_number_s*>(v->payload)->number);
+    return def;
+}
+bool memberBool(const json_object_s* o, const char* name, bool def) {
+    const json_value_s* v = memberValue(o, name);
+    if (v && v->type == json_type_true) return true;
+    if (v && v->type == json_type_false) return false;
+    return def;
+}
+std::string memberString(const json_object_s* o, const char* name, const std::string& def) {
+    const json_value_s* v = memberValue(o, name);
+    if (v && v->type == json_type_string && static_cast<const json_string_s*>(v->payload)->string)
+        return std::string(static_cast<const json_string_s*>(v->payload)->string, static_cast<const json_string_s*>(v->payload)->string_size);
+    return def;
+}
+
+// ── Biquad (RBJ, Direct Form II transposed) ─────────────────────────────────
+class Biquad {
+public:
+    void setCoeffs(double B0, double B1, double B2, double A1, double A2) {
+        b0_ = B0; b1_ = B1; b2_ = B2; a1_ = A1; a2_ = A2;
+    }
+    // Reset state (đổi params → filter mới: state cũ vô nghĩa, tránh pop).
+    void reset() { z1_[0] = z1_[1] = 0; z2_[0] = z2_[1] = 0; }
+    float step(int ch, float x) {
+        const double y = b0_ * x + z1_[ch];
+        z1_[ch] = b1_ * x - a1_ * y + z2_[ch];
+        z2_[ch] = b2_ * x - a2_ * y;
+        return (float)y;
+    }
+    void process(float* L, float* R, uint32_t n) {
+        for (uint32_t i = 0; i < n; ++i) { L[i] = step(0, L[i]); R[i] = step(1, R[i]); }
+    }
+private:
+    double b0_ = 1, b1_ = 0, b2_ = 0, a1_ = 0, a2_ = 0;
+    double z1_[2] = {0, 0}, z2_[2] = {0, 0};
+};
+
+// RBJ coefficient generators — công thức y hệt Python `_rbj_*`.
+struct RBJ { double b0, b1, b2, a1, a2; };
+RBJ rbjPeaking(double f0, double gdb, double q, double sr) {
+    const double A = std::pow(10.0, gdb / 40.0);
+    const double w0 = 2.0 * M_PI * f0 / sr;
+    const double alpha = std::sin(w0) / (2.0 * q);
+    const double cw = std::cos(w0);
+    const double a0 = 1.0 + alpha / A;
+    RBJ r;
+    r.b0 = (1.0 + alpha * A) / a0;
+    r.b1 = (-2.0 * cw) / a0;
+    r.b2 = (1.0 - alpha * A) / a0;
+    r.a1 = (-2.0 * cw) / a0;
+    r.a2 = (1.0 - alpha / A) / a0;
+    return r;
+}
+RBJ rbjShelf(double f0, double gdb, double q, double sr, bool low) {
+    const double A = std::pow(10.0, gdb / 40.0);
+    const double w0 = 2.0 * M_PI * f0 / sr;
+    const double alpha = std::sin(w0) / (2.0 * q);
+    const double cw = std::cos(w0);
+    const double sA = 2.0 * std::sqrt(A) * alpha;
+    RBJ r;
+    if (low) {
+        const double a0 = (A + 1) + (A - 1) * cw + sA;
+        r.b0 = A * ((A + 1) - (A - 1) * cw + sA) / a0;
+        r.b1 = 2.0 * A * ((A - 1) - (A + 1) * cw) / a0;
+        r.b2 = A * ((A + 1) - (A - 1) * cw - sA) / a0;
+        r.a1 = -2.0 * ((A - 1) + (A + 1) * cw) / a0;
+        r.a2 = ((A + 1) + (A - 1) * cw - sA) / a0;
+    } else {
+        const double a0 = (A + 1) - (A - 1) * cw + sA;
+        r.b0 = A * ((A + 1) + (A - 1) * cw + sA) / a0;
+        r.b1 = -2.0 * A * ((A - 1) + (A + 1) * cw) / a0;
+        r.b2 = A * ((A + 1) + (A - 1) * cw - sA) / a0;
+        r.a1 = 2.0 * ((A - 1) - (A + 1) * cw) / a0;
+        r.a2 = ((A + 1) - (A - 1) * cw - sA) / a0;
+    }
+    return r;
+}
+RBJ rbjHighpass(double f0, double q, double sr) {
+    const double w0 = 2.0 * M_PI * f0 / sr;
+    const double alpha = std::sin(w0) / (2.0 * q);
+    const double cw = std::cos(w0);
+    const double a0 = 1.0 + alpha;
+    RBJ r;
+    r.b0 = ((1.0 + cw) / 2.0) / a0;
+    r.b1 = (-(1.0 + cw)) / a0;
+    r.b2 = ((1.0 + cw) / 2.0) / a0;
+    r.a1 = (-2.0 * cw) / a0;
+    r.a2 = (1.0 - alpha) / a0;
+    return r;
+}
+RBJ rbjLowpass(double f0, double q, double sr) {
+    const double w0 = 2.0 * M_PI * f0 / sr;
+    const double alpha = std::sin(w0) / (2.0 * q);
+    const double cw = std::cos(w0);
+    const double a0 = 1.0 + alpha;
+    RBJ r;
+    r.b0 = ((1.0 - cw) / 2.0) / a0;
+    r.b1 = (1.0 - cw) / a0;
+    r.b2 = ((1.0 - cw) / 2.0) / a0;
+    r.a1 = (-2.0 * cw) / a0;
+    r.a2 = (1.0 - alpha) / a0;
+    return r;
+}
+
+// ── EQ 4-band (track 'eq'): lowshelf 100Hz, peaking 800Hz Q0.7, peaking
+//    3200Hz Q1.2, highshelf 10kHz — cùng thứ tự/đáp ứng Python `_apply_eq4`.
+class Eq4Fx : public BuiltinFx {
+public:
+    Eq4Fx(const json_object_s* params, double sr) : sr_(sr) {
+        static const double kF0[4] = {100, 800, 3200, 10000};
+        static const double kQ[4] = {0.707, 0.7, 1.2, 0.707};
+        static const bool kLow[4] = {true, false, false, false};
+        for (int i = 0; i < 4; ++i) { f0_[i] = kF0[i]; q_[i] = kQ[i]; low_[i] = kLow[i]; }
+        if (params) for (int i = 0; i < 4; ++i)
+            gains_[i] = memberNumber(params, ("g" + std::to_string(i + 1)).c_str(), 0.0);
+        rebuild();
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        for (int i = 0; i < 4; ++i) if (gains_[i] != 0.0) bands_[i].process(L, R, n);
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key.size() == 2 && key[0] == 'g' && key[1] >= '1' && key[1] <= '4') {
+            gains_[key[1] - '1'] = value;
+            rebuild();
+            return true;
+        }
+        return false;
+    }
+private:
+    void rebuild() {
+        for (int i = 0; i < 4; ++i) {
+            const RBJ r = low_[i] ? rbjShelf(f0_[i], gains_[i], q_[i], sr_, true)
+                                  : (i == 3 ? rbjShelf(f0_[i], gains_[i], q_[i], sr_, false)
+                                            : rbjPeaking(f0_[i], gains_[i], q_[i], sr_));
+            bands_[i].setCoeffs(r.b0, r.b1, r.b2, r.a1, r.a2);
+            bands_[i].reset();
+        }
+    }
+    double sr_, gains_[4] = {0, 0, 0, 0}, f0_[4], q_[4];
+    bool low_[4];
+    Biquad bands_[4];
+};
+
+// ── EQ Pro: RBJ per-band từ params.bands[] + amount — Python `_apply_eqpro`.
+class EqProFx : public BuiltinFx {
+public:
+    EqProFx(const json_object_s* params, double sr) : sr_(sr) {
+        if (params) {
+            amount_ = memberNumber(params, "amount", 100.0) / 100.0;
+            const json_value_s* bv = memberValue(params, "bands");
+            if (bv && bv->type == json_type_array) {
+                const json_array_s* arr = static_cast<const json_array_s*>(bv->payload);
+                for (const json_array_element_s* el = arr->start; el; el = el->next) {
+                    if (!el->value || el->value->type != json_type_object) continue;
+                    const json_object_s* bo = static_cast<const json_object_s*>(el->value->payload);
+                    Band b;
+                    b.active = memberBool(bo, "active", true);
+                    b.gain = memberNumber(bo, "gain", 0.0);
+                    b.freq = memberNumber(bo, "freq", 1000.0);
+                    b.q = memberNumber(bo, "q", 1.0);
+                    b.type = memberString(bo, "type", "peaking");
+                    bands_.push_back(b);
+                }
+            }
+        }
+        rebuild();
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        for (size_t i = 0; i < bqs_.size(); ++i) {
+            if (!bands_[i].active || bands_[i].gain == 0.0) continue;
+            bqs_[i].process(L, R, n);
+        }
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key == "amount") { amount_ = value / 100.0; rebuild(); return true; }
+        return false;
+    }
+private:
+    struct Band { bool active = true; double gain = 0, freq = 1000, q = 1; std::string type; };
+    void rebuild() {
+        bqs_.clear();
+        for (const auto& b : bands_) {
+            const double g = b.gain * amount_;
+            RBJ r;
+            if (b.type == "lowshelf") r = rbjShelf(b.freq, g, b.q, sr_, true);
+            else if (b.type == "highshelf") r = rbjShelf(b.freq, g, b.q, sr_, false);
+            else if (b.type == "highpass") r = rbjHighpass(b.freq, b.q, sr_);
+            else r = rbjPeaking(b.freq, g, b.q, sr_);  // peaking/lowpass/notch/bandpass
+            Biquad bq;
+            bq.setCoeffs(r.b0, r.b1, r.b2, r.a1, r.a2);
+            bqs_.push_back(bq);
+        }
+    }
+    double sr_, amount_ = 1.0;
+    std::vector<Band> bands_;
+    std::vector<Biquad> bqs_;
+};
+
+// ── Compressor: block-256 peak env, release 250ms — mirror Python 1:1.
+class CompressorFx : public BuiltinFx {
+public:
+    explicit CompressorFx(const json_object_s* params, double sr) : sr_(sr) {
+        if (params) {
+            threshold_ = memberNumber(params, "threshold", -16.0);
+            ratio_ = std::max(1.0, memberNumber(params, "ratio", 3.0));
+            makeupDb_ = memberNumber(params, "makeup", 0.0);
+        }
+        rel_ = std::exp(-1.0 / (sr_ * 0.25));
+        makeup_ = std::pow(10.0, makeupDb_ / 20.0);
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        const uint32_t block = 256;
+        float* chans[2] = {L, R};
+        for (int c = 0; c < 2; ++c) {
+            float* x = chans[c];
+            double env = env_[c];
+            for (uint32_t pos = 0; pos < n; pos += block) {
+                const uint32_t nb = std::min<uint32_t>(block, n - pos);
+                float peak = 0.f;
+                for (uint32_t i = 0; i < nb; ++i) peak = std::max(peak, std::fabs(x[pos + i]));
+                env = std::max((double)peak, env * rel_);
+                float g = (float)makeup_;
+                if (env > 1e-9) {
+                    const double db = 20.0 * std::log10(env);
+                    const double over = db - threshold_;
+                    if (over > 0.0) {
+                        const double gdb = -over * (1.0 - 1.0 / ratio_);
+                        g = (float)(std::pow(10.0, gdb / 20.0) * makeup_);
+                    }
+                }
+                for (uint32_t i = 0; i < nb; ++i) x[pos + i] *= g;
+            }
+            env_[c] = env;
+        }
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key == "threshold") { threshold_ = value; return true; }
+        if (key == "ratio") { ratio_ = std::max(1.0, value); return true; }
+        if (key == "makeup") { makeupDb_ = value; makeup_ = std::pow(10.0, makeupDb_ / 20.0); return true; }
+        return false;
+    }
+private:
+    double sr_, threshold_ = -16.0, ratio_ = 3.0, makeupDb_ = 0.0, rel_, makeup_ = 1.0;
+    double env_[2] = {0, 0};
+};
+
+// ── Limiter: tanh brickwall — mirror Python `_apply_limiter` / JS limNode.
+class LimiterFx : public BuiltinFx {
+public:
+    explicit LimiterFx(const json_object_s* params) {
+        if (params) ceilingDb_ = std::min(0.0, memberNumber(params, "ceiling", -1.0));
+        rebuild();
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        for (uint32_t i = 0; i < n; ++i) {
+            const float x = std::max(-1.f, std::min(1.f, L[i]));
+            L[i] = (float)(std::tanh((double)x * k_) / tanhK_);
+            const float y = std::max(-1.f, std::min(1.f, R[i]));
+            R[i] = (float)(std::tanh((double)y * k_) / tanhK_);
+        }
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key == "ceiling") { ceilingDb_ = std::min(0.0, value); rebuild(); return true; }
+        return false;
+    }
+private:
+    void rebuild() {
+        const double th = std::pow(10.0, ceilingDb_ / 20.0);
+        k_ = 1.0 / std::max(0.02, th);
+        tanhK_ = std::tanh(k_);
+    }
+    double ceilingDb_ = -1.0, k_, tanhK_;
+};
+
+// ── Exciter: highpass 2kHz Q0.7 + tanh — mirror Python `_apply_exciter`.
+class ExciterFx : public BuiltinFx {
+public:
+    explicit ExciterFx(const json_object_s* params, double sr) : sr_(sr) {
+        if (params) drive_ = memberNumber(params, "drive", 40.0);
+        wet_ = (drive_ / 100.0) * 0.6;
+        const RBJ r = rbjHighpass(2000.0, 0.7, sr_);
+        hp_.setCoeffs(r.b0, r.b1, r.b2, r.a1, r.a2);
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        for (uint32_t i = 0; i < n; ++i) {
+            const float yl = hp_.step(0, L[i]);
+            L[i] = L[i] + (float)(std::tanh(yl * 3.0) * wet_);
+            const float yr = hp_.step(1, R[i]);
+            R[i] = R[i] + (float)(std::tanh(yr * 3.0) * wet_);
+        }
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key == "drive") { drive_ = value; wet_ = (drive_ / 100.0) * 0.6; return true; }
+        return false;
+    }
+private:
+    double sr_, drive_ = 40.0, wet_;
+    Biquad hp_;
+};
+
+// ── Rebalance: M/S crossfeed L'=a·L+b·R, R'=b·L+a·R — Python `_apply_rebalance`.
+class RebalanceFx : public BuiltinFx {
+public:
+    explicit RebalanceFx(const json_object_s* params) {
+        if (params) {
+            mid_ = memberNumber(params, "mid", 0.0);
+            side_ = memberNumber(params, "side", 0.0);
+        }
+        rebuild();
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        for (uint32_t i = 0; i < n; ++i) {
+            const float l = L[i], r = R[i];
+            L[i] = (float)(a_ * l + b_ * r);
+            R[i] = (float)(b_ * l + a_ * r);
+        }
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key == "mid") { mid_ = value; rebuild(); return true; }
+        if (key == "side") { side_ = value; rebuild(); return true; }
+        return false;
+    }
+private:
+    void rebuild() {
+        const double mid = std::pow(10.0, mid_ / 20.0);
+        const double side = std::pow(10.0, side_ / 20.0);
+        a_ = (mid + side) / 2.0;
+        b_ = (mid - side) / 2.0;
+    }
+    double mid_ = 0, side_ = 0, a_ = 1, b_ = 0;
+};
+
+// ── Imager: 4-band crossover (100/1000/6000Hz, WebAudio biquad Q=1) + M/S
+//    width per band — JS MASTER_MODULE_IO imager (updateImagerBand).
+//    Band1=LP100, Band2=HP100+LP1000, Band3=HP1000+LP6000, Band4=HP6000.
+class ImagerFx : public BuiltinFx {
+public:
+    explicit ImagerFx(const json_object_s* params, double sr) : sr_(sr) {
+        for (int i = 0; i < 4; ++i) {
+            w_[i] = 100.0;
+            if (params) w_[i] = memberNumber(params, ("w" + std::to_string(i + 1)).c_str(), 100.0);
+        }
+        // Crossover filter graph (mỗi band 1-2 biquad nối tiếp).
+        // Mặc định: band 0/3 chỉ 1 stage; band 1/2 có 2 stage. Stage thừa bị
+        // tắt (active=false) — `{}` init sẽ zero mọi field nên phải set lại.
+        for (int b = 0; b < 4; ++b) for (int f = 0; f < 2; ++f)
+            band_[b][f].active = true;
+        const RBJ lp100 = rbjLowpass(100, 1.0, sr_), hp100 = rbjHighpass(100, 1.0, sr_);
+        const RBJ lp1k = rbjLowpass(1000, 1.0, sr_), hp1k = rbjHighpass(1000, 1.0, sr_);
+        const RBJ lp6k = rbjLowpass(6000, 1.0, sr_), hp6k = rbjHighpass(6000, 1.0, sr_);
+        apply(lp100, band_[0][0]);
+        apply(hp100, band_[1][0]); apply(lp1k, band_[1][1]);
+        apply(hp1k, band_[2][0]); apply(lp6k, band_[2][1]);
+        apply(hp6k, band_[3][0]);
+        band_[0][1].active = false;
+        band_[3][1].active = false;
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        // Biquad xử lý in-place trên bản sao band (không hủy input).
+        std::vector<float> bl(n), br(n);
+        std::vector<float> accL(n), accR(n);
+        for (int b = 0; b < 4; ++b) {
+            std::memcpy(bl.data(), L, n * sizeof(float));
+            std::memcpy(br.data(), R, n * sizeof(float));
+            for (int f = 0; f < 2 && band_[b][f].active; ++f) {
+                band_[b][f].bq.process(bl.data(), br.data(), n);
+            }
+            const double width = std::max(0.0, std::min(200.0, w_[b]));
+            const double g1 = (width + 100.0) / 200.0;
+            const double g2 = (100.0 - width) / 200.0;
+            for (uint32_t i = 0; i < n; ++i) {
+                accL[i] += (float)(g1 * bl[i] + g2 * br[i]);
+                accR[i] += (float)(g1 * br[i] + g2 * bl[i]);
+            }
+        }
+        std::memcpy(L, accL.data(), n * sizeof(float));
+        std::memcpy(R, accR.data(), n * sizeof(float));
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key.size() == 2 && key[0] == 'w' && key[1] >= '1' && key[1] <= '4') {
+            w_[key[1] - '1'] = value;
+            return true;
+        }
+        return false;
+    }
+private:
+    struct BandFilter { Biquad bq; bool active = true; };
+    static void apply(const RBJ& r, BandFilter& f) {
+        f.bq.setCoeffs(r.b0, r.b1, r.b2, r.a1, r.a2);
+    }
+    double sr_, w_[4];
+    BandFilter band_[4][2] = {};  // [band][stage]; stage 2 inactive → skip
+};
+
+// ── Maximizer: boost → soft-clip → (+upward comp) → ceiling clip — JS
+//    MASTER_MODULE_IO maximizer. Upward compressor approximates WebAudio
+//    DynamicsCompressor (soft-knee, attack/release) — `ponytail: port Chromium
+//    algorithm chính xác nếu cần bit-parity; default maxUpward=0 → no-op`.
+class MaximizerFx : public BuiltinFx {
+public:
+    explicit MaximizerFx(const json_object_s* params, double sr) : sr_(sr) {
+        if (params) {
+            boostDb_ = memberNumber(params, "boost_db", 0.0);
+            softClip_ = memberNumber(params, "soft_clip", 0.0);
+            upward_ = memberNumber(params, "upward", 0.0);
+            ceilingDb_ = memberNumber(params, "ceiling_db", -0.1);
+        }
+        rebuild();
+        att_ = std::exp(-1.0 / (sr_ * 0.01));
+        rel_ = std::exp(-1.0 / (sr_ * 0.1));
+    }
+    void process(float* L, float* R, uint32_t n) override {
+        for (uint32_t i = 0; i < n; ++i) {
+            L[i] = processSample(0, L[i]);
+            R[i] = processSample(1, R[i]);
+        }
+    }
+    bool setParam(const std::string& key, double value) override {
+        if (key == "boost_db") { boostDb_ = value; rebuild(); return true; }
+        if (key == "soft_clip") { softClip_ = value; rebuild(); return true; }
+        if (key == "upward") { upward_ = value; rebuild(); return true; }
+        if (key == "ceiling_db") { ceilingDb_ = value; rebuild(); return true; }
+        return false;
+    }
+private:
+    void rebuild() {
+        boost_ = std::pow(10.0, std::max(-60.0, std::min(30.0, boostDb_)) / 20.0);
+        const double p = std::max(0.0, std::min(100.0, softClip_));
+        t_ = 1.0 - (p / 100.0) * 0.8;
+        inv_ = 1.0 - t_;
+        upwardGain_ = upward_ > 0 ? std::pow(10.0, std::max(0.0, std::min(30.0, upward_)) / 20.0) - 1.0 : 0.0;
+        ceiling_ = std::pow(10.0, std::max(-60.0, std::min(0.0, ceilingDb_)) / 20.0);
+    }
+    float softClip(float x) const {
+        const float ax = std::fabs(x);
+        if (ax < (float)t_) return x;
+        return (float)((x < 0 ? -1.0 : 1.0) * (t_ + inv_ * std::tanh((ax - t_) / inv_)));
+    }
+    // Soft-knee feedforward compressor (DynamicsCompressor-ish).
+    float upwardComp(int ch, float x) {
+        const float ax = std::fabs(x);
+        if (ax > env_[ch]) env_[ch] = att_ * env_[ch] + (1.0 - att_) * ax;
+        else env_[ch] = rel_ * env_[ch] + (1.0 - rel_) * ax;
+        double db = 20.0 * std::log10(std::max(env_[ch], 1e-12));
+        double y = db - (-30.0);  // threshold -30dB
+        double g = 1.0;
+        if (2.0 * y > 10.0) {          // beyond knee → slope 1/ratio
+            g = std::pow(10.0, -y * (1.0 - 1.0 / 4.0) / 20.0);
+        } else if (2.0 * y > -10.0) {  // soft knee
+            const double k = 10.0;
+            g = std::pow(10.0, -(1.0 - 1.0 / 4.0) * (y + k / 2.0) * (y + k / 2.0) / (2.0 * k) / 20.0);
+        }
+        return (float)(g * x);
+    }
+    float processSample(int ch, float x) {
+        const float xb = (float)(boost_ * x);
+        const float dry = softClip(xb);
+        const float wet = upwardGain_ * upwardComp(ch, xb);
+        const float sum = dry + wet;
+        return std::max(-(float)ceiling_, std::min((float)ceiling_, sum));
+    }
+    double sr_, boostDb_ = 0, softClip_ = 0, upward_ = 0, ceilingDb_ = -0.1;
+    double boost_ = 1.0, t_ = 1.0, inv_ = 0.0, upwardGain_ = 0.0, ceiling_ = 0.9886;
+    double att_, rel_;
+    double env_[2] = {0, 0};
+};
+
+}  // namespace
+
+// ── BuiltinFxChain ──────────────────────────────────────────────────────────
+void BuiltinFxChain::add(std::unique_ptr<BuiltinFx> fx, bool bypass) {
+    entries_.push_back(Entry{std::move(fx), bypass});
+}
+void BuiltinFxChain::process(float* L, float* R, uint32_t n) {
+    for (auto& e : entries_) {
+        if (e.bypass || !e.fx) continue;
+        e.fx->process(L, R, n);
+    }
+}
+bool BuiltinFxChain::setParam(int slot, const std::string& key, double value) {
+    if (slot < 0 || (size_t)slot >= entries_.size()) return false;
+    auto& e = entries_[(size_t)slot];
+    return e.fx && e.fx->setParam(key, value);
+}
+
+std::unique_ptr<BuiltinFx> createBuiltinFx(const std::string& id,
+                                           const json_object_s* params,
+                                           double sampleRate) {
+    if (id == "eq") return std::make_unique<Eq4Fx>(params, sampleRate);
+    if (id == "eqpro") return std::make_unique<EqProFx>(params, sampleRate);
+    if (id == "compressor") return std::make_unique<CompressorFx>(params, sampleRate);
+    if (id == "limiter") return std::make_unique<LimiterFx>(params);
+    if (id == "exciter") return std::make_unique<ExciterFx>(params, sampleRate);
+    if (id == "rebalance") return std::make_unique<RebalanceFx>(params);
+    if (id == "imager") return std::make_unique<ImagerFx>(params, sampleRate);
+    if (id == "maximizer") return std::make_unique<MaximizerFx>(params, sampleRate);
+    return nullptr;  // gain/normalize = legacy, xử lý riêng ở RenderFxJob
+}
diff --git a/native_bridge/src/RealtimeFxLoop.cpp b/native_bridge/src/RealtimeFxLoop.cpp
index bb7cf8a..a3202fd 100644
--- a/native_bridge/src/RealtimeFxLoop.cpp
+++ b/native_bridge/src/RealtimeFxLoop.cpp
@@ -166,11 +166,41 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
     uint32_t inMask = ipc->h.inSlots - 1;
     uint32_t outMask = ipc->h.outSlots - 1;
     uint64_t processed = 0;
+    uint64_t lastGen = 0;  // chain generation đã báo latency
     while (ipc->h.running) {
         if (parentPid && !parentAlive(parentPid)) {
             std::cerr << "[RealtimeFxLoop] parent gone — exiting" << std::endl;
             break;
         }
+        // SET_PARAM ring (engine -> bridge, Phase 2.8): drain mỗi iteration —
+        // áp dụng vào chain TRƯỚC block kế tiếp. Ring đầy → drop lệnh cũ.
+        while (ipc->h.ctrlRead < ipc->h.ctrlWrite) {
+            const uint32_t cs = ipc->h.ctrlRead & (FXRT_CTRL_SLOTS - 1);
+            const FxCtrlCmd& cmd = ipc->ctrl[cs];
+            const char* kend = std::find(cmd.key, cmd.key + sizeof(cmd.key), '\0');
+            chain.setParam((int)cmd.slot,
+                           std::string(cmd.key, (size_t)(kend - cmd.key)),
+                           (double)cmd.value);
+            MemoryBarrier();
+            ipc->h.ctrlRead++;
+        }
+        // REPORT_LATENCY (bridge -> engine): chain swap (gen đổi) → báo lại
+        // latency từng slot (builtin=0; VST3=getLatencySamples). Ring 8 —
+        // engine drain định kỳ, đầy thì overwrite (drop cũ).
+        const uint64_t gen = chain.chainGen();
+        if (gen != lastGen) {
+            const std::vector<uint32_t> lats = chain.entryLatencies();
+            for (size_t i = 0; i < lats.size(); ++i) {
+                const uint32_t ls = ipc->h.latWrite & (FXRT_LAT_SLOTS - 1);
+                ipc->lat[ls].slot = (uint32_t)i;
+                ipc->lat[ls].samples = lats[i];
+                MemoryBarrier();
+                ipc->h.latWrite++;
+            }
+            lastGen = gen;
+            std::cerr << "[RealtimeFxLoop] latency reported: " << lats.size()
+                      << " slot(s)" << std::endl;
+        }
         const uint32_t avail = ipc->h.inWrite - ipc->h.inRead;
         if (avail == 0) { sleepMs(1); continue; }
         const uint32_t slot = ipc->h.inRead & inMask;
diff --git a/native_bridge/src/RenderFxJob.cpp b/native_bridge/src/RenderFxJob.cpp
index f8d7781..98ef77e 100644
--- a/native_bridge/src/RenderFxJob.cpp
+++ b/native_bridge/src/RenderFxJob.cpp
@@ -38,8 +38,10 @@
 #include <windows.h>
 #endif
 #include "FxRealtimeIPC.h"
+#include "BuiltinFxChain.h"
 
 #include <algorithm>
+#include <atomic>
 #include <cctype>
 #include <cmath>
 #include <cstdio>
@@ -382,6 +384,7 @@ struct Vst3FxState {
     HostComponentHandlerFx componentHandler;
     int32 inputChannels = 2;
     int32 outputChannels = 2;
+    int32 latencySamples = 0;  // PDC (REPORT_LATENCY, Phase 2.8)
     bool controllerIsComponent = false;
     Steinberg::IPlugFrame* plugFrame = nullptr;  // owned; freed in closeEditor
     IPtr<IPlugView> view;                        // editor view while GUI open
@@ -517,6 +520,8 @@ bool vst3FxLoadInner(Vst3FxState* s, const std::string& path, double sampleRate,
     std::cerr << "[RenderFx] STEP setActive OK" << std::endl;
     processor->setProcessing(true);
     std::cerr << "[RenderFx] STEP setProcessing OK" << std::endl;
+    s->latencySamples = processor->getLatencySamples();
+    std::cerr << "[RenderFx] STEP getLatencySamples = " << s->latencySamples << std::endl;
     if (!s->processData.prepare(*component, maxBlockSize, kSample32)) {
         err = "processData.prepare failed"; return false;
     }
@@ -850,12 +855,22 @@ public:
 #endif
     }
     bool loaded() const { return state_ != nullptr; }
+    // Plugin-reported latency samples (PDC) — 0 nếu plugin không báo.
+    int32 latencySamples() const {
+#ifndef HAVE_VST3SDK
+        return 0;
+#else
+        return state_ ? static_cast<Vst3FxState*>(state_)->latencySamples : 0;
+#endif
+    }
 
 private:
     void* state_ = nullptr;
 };
 
-// One FX slot from the job. builtin id: "gain" | "normalize".
+// One FX slot from the job. builtin id: "gain" | "normalize" (legacy) hoặc 8
+// DSP (PLAN_DAW_A Phase 2) — id ∈ eq|eqpro|imager|maximizer|compressor|limiter
+// |exciter|rebalance → bfx (BuiltinFxChain).
 struct FxSlot {
     bool vst = false;        // true = VST3, false = builtin
     std::string path;        // vst only
@@ -865,6 +880,7 @@ struct FxSlot {
     double db = 0.0;         // gain param
     double peak = 0.95;      // normalize param
     std::unique_ptr<Vst3Fx> fx;
+    std::unique_ptr<BuiltinFx> bfx;  // 8 DSP builtin (Phase 2)
 };
 
 }  // namespace
@@ -887,6 +903,22 @@ static bool runFxSlotSafe(const FxSlot& sl, const float* inL, const float* inR,
 #endif
 }
 
+// SEH-guarded builtin DSP slot (in-place). Frame chỉ raw pointers.
+static bool runBuiltinSlotSafe(BuiltinFx* fx, float* L, float* R, uint32_t n,
+                               uint32_t* crashCode) {
+#ifdef _WIN32
+    __try {
+        fx->process(L, R, n);
+        return true;
+    } __except (*crashCode = (uint32_t)GetExceptionCode(), EXCEPTION_EXECUTE_HANDLER) {
+        return false;
+    }
+#else
+    fx->process(L, R, n);
+    return true;
+#endif
+}
+
 int run_render_fx_job(const std::string& jobPath, const std::string& inPath,
                       const std::string& outPath) {
 #ifdef _WIN32
@@ -951,11 +983,18 @@ int run_render_fx_job(const std::string& jobPath, const std::string& inPath,
             } else if (type == "builtin") {
                 sl.vst = false;
                 sl.builtinId = memberString(o, "id", "");
+                const json_object_s* po = nullptr;
                 const json_value_s* pv = memberValue(o, "params");
-                if (pv && pv->type == json_type_object) {
-                    const json_object_s* po = static_cast<const json_object_s*>(pv->payload);
-                    double d = sl.db; memberNumber(po, "db", d); sl.db = d;
-                    double pk = sl.peak; memberNumber(po, "peak", pk); sl.peak = pk;
+                if (pv && pv->type == json_type_object)
+                    po = static_cast<const json_object_s*>(pv->payload);
+                double d = sl.db; if (po) memberNumber(po, "db", d); sl.db = d;
+                double pk = sl.peak; if (po) memberNumber(po, "peak", pk); sl.peak = pk;
+                if (sl.builtinId != "gain" && sl.builtinId != "normalize") {
+                    // 8 DSP (PLAN_DAW_A Phase 2) — khớp Python _apply_builtin_fx_chain.
+                    sl.bfx = createBuiltinFx(sl.builtinId, po, srD);
+                    if (!sl.bfx) {
+                        fail("unknown builtin id: " + sl.builtinId); return rc;
+                    }
                 }
             } else {
                 fail("unknown fx_chain slot type: " + type); return rc;
@@ -1004,6 +1043,14 @@ int run_render_fx_job(const std::string& jobPath, const std::string& inPath,
                                         + std::to_string(crashCode) + ")"); return rc; }
                 std::memcpy(in.L.data() + pos, L0.data(), n * sizeof(float));
                 std::memcpy(in.R.data() + pos, R0.data(), n * sizeof(float));
+            } else if (sl.bfx) {
+                // Builtin DSP (8 loại) — in-place, xen kẽ VST3 đúng thứ tự UI.
+                bool ok = runBuiltinSlotSafe(sl.bfx.get(), in.L.data() + pos,
+                                             in.R.data() + pos, n, &crashCode);
+                if (!ok) {
+                    rc = 3; fail("builtin DSP crashed (SEH code=0x"
+                                 + std::to_string(crashCode) + ")"); return rc;
+                }
             } else {
                 for (uint32_t i = 0; i < n; ++i) {
                     in.L[pos + i] = curL[i] * (float)gainLin;
@@ -1058,32 +1105,38 @@ int run_render_fx_job(const std::string& jobPath, const std::string& inPath,
 namespace {
 
 // SEH frame holds only raw pointers / trivials (C2712 — no unwinding locals).
-static void realtimeRunChain(Vst3Fx** fxs, const bool* bypass, uint32_t count,
-                             float* inL, float* inR, uint32_t n) {
+static void realtimeRunChain(Vst3Fx** vst, BuiltinFx** bfx, const bool* bypass,
+                             uint32_t count, float* inL, float* inR, uint32_t n) {
 #ifdef _WIN32
     __try {
         for (uint32_t i = 0; i < count; ++i) {
-            if (!fxs[i]) continue;
             if (bypass && bypass[i]) continue;
-            fxs[i]->processAudio(inL, inR, inL, inR, n);  // in-place safe
+            if (vst[i]) vst[i]->processAudio(inL, inR, inL, inR, n);  // in-place safe
+            else if (bfx[i]) bfx[i]->process(inL, inR, n);           // builtin in-place
         }
     } __except (EXCEPTION_EXECUTE_HANDLER) {
-        std::cerr << "[RealtimeFx] plugin crashed in process — chain bypassed" << std::endl;
+        std::cerr << "[RealtimeFx] chain crashed in process — slot bypassed" << std::endl;
     }
 #else
-    (void)fxs; (void)bypass; (void)count; (void)inL; (void)inR; (void)n;
+    (void)vst; (void)bfx; (void)bypass; (void)count; (void)inL; (void)inR; (void)n;
 #endif
 }
 
 }  // namespace
 
 struct RealtimeFxChain::Impl {
-    struct Entry { std::shared_ptr<Vst3Fx> fx; bool bypass = false; };
+    // VST3 (fx) hoặc builtin DSP (bfx) — 1 trong 2; xen kẽ đúng thứ tự UI.
+    struct Entry {
+        std::shared_ptr<Vst3Fx> fx;
+        std::unique_ptr<BuiltinFx> bfx;
+        bool bypass = false;
+    };
     struct Chain { std::vector<Entry> entries; };
 
     std::mutex mutex_;               // guards chain_ / retired_
     std::shared_ptr<Chain> chain_;
     std::shared_ptr<Chain> retired_; // dtor deferred to the worker thread
+    std::atomic<uint64_t> gen_{0};   // chain generation (Phase 2.8)
 
     std::mutex qmutex_;              // guards q_ / quit_
     std::condition_variable qcv_;
@@ -1140,16 +1193,45 @@ void RealtimeFxChain::process(float* inL, float* inR, uint32_t n) {
     }
     if (!c || c->entries.empty()) return;
     constexpr uint32_t kMaxSlots = 16;  // v1 cap
-    Vst3Fx* fxs[kMaxSlots];
+    Vst3Fx* vst[kMaxSlots];
+    BuiltinFx* bfx[kMaxSlots];
     bool bypass[kMaxSlots];
     uint32_t count = 0;
     for (const auto& e : c->entries) {
         if (count >= kMaxSlots) break;
-        fxs[count] = e.fx.get();
+        vst[count] = e.fx.get();
+        bfx[count] = e.bfx.get();
         bypass[count] = e.bypass;
         ++count;
     }
-    realtimeRunChain(fxs, bypass, count, inL, inR, n);
+    realtimeRunChain(vst, bfx, bypass, count, inL, inR, n);
+}
+
+void RealtimeFxChain::setParam(int slot, const std::string& key, double value) {
+    // SET_PARAM (SHM control ring, Phase 2.8): áp dụng live vào builtin slot.
+    // VST3 param automation qua VST3 API — ponytail: them khi co nhu cau.
+    if (!impl_) return;
+    std::shared_ptr<Impl::Chain> c;
+    { std::lock_guard<std::mutex> lk(impl_->mutex_); c = impl_->chain_; }
+    if (!c || slot < 0 || (size_t)slot >= c->entries.size()) return;
+    auto& e = c->entries[(size_t)slot];
+    if (e.bfx) e.bfx->setParam(key, value);
+}
+
+uint64_t RealtimeFxChain::chainGen() {
+    return impl_ ? impl_->gen_.load(std::memory_order_acquire) : 0;
+}
+
+std::vector<uint32_t> RealtimeFxChain::entryLatencies() {
+    std::vector<uint32_t> out;
+    if (!impl_) return out;
+    std::shared_ptr<Impl::Chain> c;
+    { std::lock_guard<std::mutex> lk(impl_->mutex_); c = impl_->chain_; }
+    if (!c) return out;
+    out.reserve(c->entries.size());
+    for (const auto& e : c->entries)
+        out.push_back(e.fx ? (uint32_t)e.fx->latencySamples() : 0u);
+    return out;
 }
 
 std::shared_ptr<RealtimeFxChain::Impl::Chain> RealtimeFxChain::Impl::buildChain(const std::string& json) {
@@ -1166,6 +1248,25 @@ std::shared_ptr<RealtimeFxChain::Impl::Chain> RealtimeFxChain::Impl::buildChain(
     for (const json_array_element_s* e = arr->start; e; e = e->next) {
         if (!e->value || e->value->type != json_type_object) continue;
         const json_object_s* o = static_cast<const json_object_s*>(e->value->payload);
+        const std::string type = memberString(o, "type", "");
+        if (type == "builtin") {
+            // Builtin DSP (Phase 2): id ∈ 8 loại — xen kẽ VST3 đúng thứ tự.
+            const std::string id = memberString(o, "id", "");
+            if (id.empty()) continue;
+            const json_object_s* po = nullptr;
+            const json_value_s* pv = memberValue(o, "params");
+            if (pv && pv->type == json_type_object)
+                po = static_cast<const json_object_s*>(pv->payload);
+            Entry ent;
+            ent.bypass = memberBool(o, "bypass", false);
+            ent.bfx = createBuiltinFx(id, po, sampleRate_);
+            if (!ent.bfx) {
+                std::cerr << "[RealtimeFx] unknown builtin id — slot skipped: " << id << std::endl;
+                continue;
+            }
+            chain->entries.push_back(std::move(ent));
+            continue;
+        }
         const std::string path = memberString(o, "path", "");
         if (path.empty()) continue;
         Entry ent;
@@ -1213,6 +1314,7 @@ void RealtimeFxChain::Impl::workerLoop() {
             retired_ = std::move(chain_);  // old chain destroyed on THIS thread
             chain_ = std::move(next);
         }
+        gen_.fetch_add(1, std::memory_order_release);  // loop báo lại latency
     }
 #ifdef _WIN32
     CoUninitialize();
diff --git a/tests/test_builtin_fx_golden.py b/tests/test_builtin_fx_golden.py
new file mode 100644
index 0000000..f904fe4
--- /dev/null
+++ b/tests/test_builtin_fx_golden.py
@@ -0,0 +1,139 @@
+# -*- coding: utf-8 -*-
+"""Golden test PLAN_DAW_A.md 2.9: C++ BuiltinFxChain vs Python
+_apply_builtin_fx_chain (render_engine.py) — SNR per-chain, kể cả order test.
+
+Skip nếu thiếu fx_vst_bridge.exe (build/Release) hoặc scipy.
+"""
+import io, json, math, os, subprocess, sys, tempfile, wave
+
+import numpy as np
+import pytest
+
+REPO = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, REPO)
+
+from app.core.render_engine import _apply_builtin_fx_chain  # noqa: E402
+
+BRIDGE_CANDIDATES = [
+    os.environ.get("SF_FX_BRIDGE_PATH", ""),
+    os.path.join(REPO, "native_bridge", "build", "Release", "fx_vst_bridge.exe"),
+    os.path.join(REPO, "install", "fx_vst_bridge.exe"),
+]
+
+
+def _find_bridge():
+    for c in BRIDGE_CANDIDATES:
+        if c and os.path.isfile(c):
+            return c
+    return None
+
+
+@pytest.fixture(scope="module")
+def bridge():
+    b = _find_bridge()
+    if not b:
+        pytest.skip("fx_vst_bridge.exe not found (native_bridge/build/Release)")
+    return b
+
+
+def _write_wav(path, data, sr):
+    data = np.clip(data, -1.0, 1.0)
+    # data shape (2, n) → interleave (n, 2) trước khi ghi
+    pcm = (data.T * 32767).astype(np.int16)
+    with wave.open(path, "wb") as w:
+        w.setnchannels(2)
+        w.setsampwidth(2)
+        w.setframerate(sr)
+        w.writeframes(pcm.tobytes())
+
+
+def _read_wav(path):
+    import soundfile as sf
+    data, sr = sf.read(path, dtype="float32", always_2d=True)
+    return data.T, sr
+
+
+def _render_bridge(bridge, wav_in, sr, chain):
+    job = {"sample_rate": sr, "block_size": 512, "fx_chain": chain}
+    with tempfile.TemporaryDirectory() as td:
+        job_path = os.path.join(td, "job.json")
+        out_path = os.path.join(td, "out.wav")
+        with io.open(job_path, "w", encoding="utf-8") as f:
+            json.dump(job, f)
+        r = subprocess.run([bridge, "--render-fx", job_path, "--in", wav_in,
+                            "--out", out_path],
+                           capture_output=True, text=True, timeout=120)
+        assert r.returncode == 0, f"bridge rc={r.returncode}: {r.stdout} {r.stderr}"
+        assert os.path.isfile(out_path), f"no out wav: {r.stdout} {r.stderr}"
+        return _read_wav(out_path)[0]
+
+
+def _snr_db(ref, got):
+    ref = ref.astype(np.float64)
+    got = got.astype(np.float64)
+    if ref.shape != got.shape:
+        raise AssertionError(f"shape mismatch {ref.shape} vs {got.shape}")
+    n = min(ref.shape[1], got.shape[1])
+    if n == 0:
+        return math.inf
+    ref, got = ref[:, :n], got[:, :n]
+    denom = float(np.sum((ref - got) ** 2))
+    if denom < 1e-12:
+        return math.inf
+    return 10.0 * math.log10(float(np.sum(ref ** 2)) / denom)
+
+
+def _run_case(bridge, sr, seed, chain):
+    rng = np.random.default_rng(seed)
+    n = sr // 4
+    t = np.arange(n) / sr
+    sig = (0.25 * np.sin(2 * np.pi * 220 * t)
+           + 0.12 * np.sin(2 * np.pi * 1100 * t)
+           + 0.06 * np.sin(2 * np.pi * 5000 * t))
+    noise = 0.01 * rng.standard_normal(n)
+    audio = np.stack([sig + noise, 0.9 * sig + 0.02 * rng.standard_normal(n)])
+    audio = audio.astype(np.float32)
+
+    with tempfile.TemporaryDirectory() as td:
+        wav_in = os.path.join(td, "in.wav")
+        _write_wav(wav_in, audio, sr)
+        got = _render_bridge(bridge, wav_in, sr, chain)
+
+    # Python reference dùng schema cũ {type: eq} — convert từ job schema mới
+    # {type: builtin, id: eq} (bridge nhận schema mới, spec Phase 2).
+    py_chain = [{"type": s.get("id", s.get("type")), "params": s.get("params") or {}}
+                for s in chain]
+    ref = _apply_builtin_fx_chain(audio.copy(), py_chain, sr)
+    snr = _snr_db(ref, got)
+    return snr
+
+
+def _b(id, params):
+    return {"type": "builtin", "id": id, "params": params or {}}
+
+
+CHAINS = {
+    "eq": [_b("eq", {"g1": 3.0, "g2": -2.0, "g3": 1.5, "g4": 2.0})],
+    "eqpro": [_b("eqpro", {"amount": 80, "bands": [
+        {"active": True, "type": "lowshelf", "freq": 120, "gain": 2.5, "q": 0.8},
+        {"active": True, "type": "peaking", "freq": 900, "gain": -3.0, "q": 1.2},
+        {"active": True, "type": "highpass", "freq": 60, "gain": 0.0, "q": 0.7},
+        {"active": True, "type": "highshelf", "freq": 8000, "gain": 1.8, "q": 0.7},
+    ]})],
+    "compressor": [_b("compressor", {"threshold": -20.0, "ratio": 4.0, "makeup": 2.0})],
+    "limiter": [_b("limiter", {"ceiling": -3.0})],
+    "exciter": [_b("exciter", {"drive": 60.0})],
+    "rebalance": [_b("rebalance", {"mid": 2.0, "side": -1.5})],
+    "order_ec": [_b("eq", {"g1": 2.0, "g3": -1.0}),
+                 _b("compressor", {"threshold": -18.0, "ratio": 3.0})],
+    "order_ce": [_b("compressor", {"threshold": -18.0, "ratio": 3.0}),
+                 _b("eq", {"g1": 2.0, "g3": -1.0})],
+}
+
+MIN_SNR = 30.0
+
+
+@pytest.mark.parametrize("name", sorted(CHAINS))
+def test_builtin_golden(bridge, name):
+    snr = _run_case(bridge, 44100, seed=1234, chain=CHAINS[name])
+    assert snr >= MIN_SNR, f"{name}: SNR {snr:.1f} dB < {MIN_SNR}"