feat: tách luồng âm instrument theo môi trường — docker dùng FluidSynthWASM, standalone dùng native pyfluidsynth (+Carla VSTi)

- runtime: detect()/capabilities() trả environment (docker|standalone)
- vst_engine: render_soundfont_midi_to_audio (pyfluidsynth native, event-stream, release tail)
- plugins: POST /soundfont-render — render MIDI notes -> WAV (auth, fallback default SF)
- frontend: SonicRuntime.environment + dataset.environment; app.jsx route preview/transport:
  standalone+SF -> soundfont-render WAV (playNativeSfNote/scheduleNativeSfItem),
  docker giữ WASM, VSTi standalone giữ Carla bridge

app/core/vst_engine.py

@@ -8,6 +8,74 @@ from ctypes import c_char_p
 def midi_note_to_freq(note_number: int) -> float:
     return 440.0 * (2.0 ** ((note_number - 69) / 12.0))
 
+
+def render_soundfont_midi_to_audio(midi_events: list, sf_path: str, bank: int = 0,
+                                   program: int = 0, sr: int = 44100, bpm: float = 120.0) -> np.ndarray:
+    """Native FluidSynth (pyfluidsynth) — render MIDI events bằng soundfont THẬT.
+
+    Dùng cho môi trường standalone: âm instrument xử lí trực tiếp trên OS
+    (backend pyfluidsynth), không phải FluidSynthWASM trong browser.
+    Xử lí đúng note CHỒNG nhau + đuôi release (event-stream noteon/noteoff).
+    Trả stereo float32 (2, N). Raise RuntimeError nếu pyfluidsynth thiếu / SF
+    không load được."""
+    import fluidsynth as _fs
+    settings = _fs.new_fluid_settings()
+    _fs.fluid_settings_setnum(settings, b'synth.sample-rate', float(sr))
+    synth = _fs.new_fluid_synth(settings)
+    try:
+        fid = _fs.fluid_synth_sfload(synth, os.fspath(sf_path).encode('utf-8'), 1)
+        if fid == -1:
+            raise RuntimeError(f"SoundFont load failed: {sf_path}")
+        _fs.fluid_synth_program_select(synth, 0, fid, int(bank), int(program))
+        beat_sec = 60.0 / max(30.0, bpm)
+        events = []
+        total_sec = 0.0
+        for ev in midi_events:
+            start_s = int(float(ev.get('start_beat', 0)) * beat_sec * sr)
+            dur_s = int(float(ev.get('duration_beats', 1)) * beat_sec * sr)
+            note = int(ev.get('note', 60))
+            vel = min(127, max(1, int(float(ev.get('velocity', 100)))))
+            events.append((start_s, 'on', note, vel))
+            events.append((start_s + dur_s, 'off', note, 0))
+            end_sec = (start_s + dur_s) / float(sr)
+            if end_sec > total_sec:
+                total_sec = end_sec
+        events.sort(key=lambda e: e[0])
+        total_sec = max(total_sec, 0.25) + 0.5  # đuôi reverb/release
+        n = int(total_sec * sr)
+        out = np.zeros((2, n), dtype=np.float32)
+        pos = 0
+
+        def _render_until(target: int):
+            """Render từ pos tới target, CỘNG vào out (giữ âm đang ngân)."""
+            nonlocal pos
+            target = min(target, n)
+            while pos < target:
+                chunk = min(target - pos, 44100)
+                block = _fs.fluid_synth_write_s16_stereo(synth, chunk)
+                b = block.astype(np.float32).reshape(-1, 2).T / 32768.0
+                take = min(chunk, b.shape[1], target - pos)
+                if take > 0:
+                    out[:, pos:pos + take] += b[:, :take]
+                pos += take
+
+        for t, kind, note, vel in events:
+            if t > pos:
+                _render_until(t)
+            if pos >= n:
+                break
+            if kind == 'on':
+                _fs.fluid_synth_noteon(synth, 0, note, vel)
+            else:
+                _fs.fluid_synth_noteoff(synth, 0, note)
+        _render_until(n)  # đuôi release của notes cuối
+        return out
+    finally:
+        try:
+            _fs.delete_fluid_synth(synth)
+        except Exception:
+            pass
+
 def render_midi_events_to_audio(midi_events: list, sr: int = 44100, bpm: float = 120.0, instrument: str = 'synth') -> np.ndarray:
     beat_duration_sec = 60.0 / max(30.0, bpm)
     max_duration_sec = 2.0