feat(v0.6.0): Edge AI — AI-IDS + ONNX Dispatcher

Implements the first two capabilities from the BESSAI v2.0 roadmap (Q3 2026):

AI-IDS (src/interfaces/ai_ids.py):
- ModbusAnomalyDetector: IsolationForest + z-score timing ensemble
- Score 0-1 (0=normal, 1=anomalous), threshold=0.65 default
- Fail-safe: returns 0.0 before fit() (no false positives on startup)
- Alerts via structlog + bess_ids_alerts_total Prometheus counter

ONNX Dispatcher (src/interfaces/onnx_dispatcher.py):
- Loads dispatch_policy.onnx at edge (no internet required)
- Input: [soc_pct, power_kw, temp_c, hour_of_day] -> target_kw
- Graceful fallback: returns None if model missing -> SafetyGuard takes over
- Async context manager with Prometheus bess_onnx_inference_ms gauge

Metrics (src/interfaces/metrics.py):
- +4 new Prometheus metrics: IDS_ALERTS_TOTAL, IDS_ANOMALY_SCORE,
  ONNX_INFERENCE_MS, ONNX_DISPATCH_COMMANDS_TOTAL

ONNX model (models/dispatch_policy.onnx):
- Dummy linear model: target_kw = soc_pct * 0.8 (for testing)
- Generated by scripts/generate_dummy_onnx.py
- Replace with trained Ray RLlib export in production

Dependencies: numpy>=1.26.0, scikit-learn>=1.4.0, onnxruntime>=1.18.0

Tests: 73/73 passed in 11.89s (+19 new: test_ai_ids.py + test_onnx_dispatcher.py)

Author: BESS Solutions

CHANGELOG.md

@@ -7,34 +7,32 @@
 
 ---
 
-## 🤖 AGENT HANDOFF — Estado actual del proyecto (2026-02-19T15:00 -03:00)
+## 🤖 AGENT HANDOFF — Estado actual del proyecto (2026-02-19T15:09 -03:00)
 
 ### Contexto del sistema
 **BESSAI Edge Gateway** (`open-bess-edge`) es el componente de borde de un sistema de gestión de baterías industriales (BESS). Adquiere telemetría via **Modbus TCP** desde inversores Huawei SUN2000, valida seguridad, y publica a **GCP Pub/Sub** con observabilidad via **OpenTelemetry** y **Prometheus**.
 
-### Estado del código — ✅ v0.5.0, COMPLETO Y VALIDADO
+### Estado del código — ✅ v0.6.0, COMPLETO Y VALIDADO
 
 | Archivo | Estado | Notas |
 |---|---|---|
 | `src/core/config.py` | ✅ Producción | `INVERTER_IP` acepta IPs y hostnames. Nuevo: `HEALTH_PORT=8000` |
 | `src/core/safety.py` | ✅ Producción | check_safety + watchdog_loop async |
 | `src/core/main.py` | ✅ Producción | Integrado con HealthServer + Prometheus metrics |
 | `src/drivers/modbus_driver.py` | ✅ Producción | pymodbus 3.12, struct-based encode/decode |
-| `src/interfaces/health.py` | ✅ **NUEVO** | Servidor HTTP /health (JSON) + /metrics (Prometheus) vía aiohttp |
-| `src/interfaces/metrics.py` | ✅ **NUEVO** | Contadores/Gauges: cycles, safety_blocks, SOC, power, cycle_duration |
+| `src/interfaces/health.py` | ✅ Producción | Servidor HTTP /health (JSON) + /metrics (Prometheus) vía aiohttp |
+| `src/interfaces/metrics.py` | ✅ **AMPLIADO** | +4 métricas AI: IDS_ALERTS, IDS_SCORE, ONNX_MS, ONNX_CMDS |
+| `src/interfaces/ai_ids.py` | ✅ **NUEVO** | AI-IDS: IsolationForest + z-score ensemble, score 0-1, alertas Prometheus |
+| `src/interfaces/onnx_dispatcher.py` | ✅ **NUEVO** | ONNX Runtime offline dispatcher, fallback gracioso si no hay modelo |
 | `src/interfaces/pubsub_publisher.py` | ✅ Producción | Async context manager, GCP Pub/Sub, JSON envelope |
 | `src/interfaces/otel_setup.py` | ✅ Producción | TracerProvider + MeterProvider |
-| `infrastructure/docker/docker-compose.yml` | ✅ **MEJORADO** | +Perfil `monitoring` (Prometheus+Grafana), port 8000, healthcheck HTTP |
-| `infrastructure/prometheus/prometheus.yml` | ✅ **NUEVO** | Scrape config: gateway:8000 + otel-collector:8888 |
-| `infrastructure/grafana/provisioning/` | ✅ **NUEVO** | Auto-provisioning datasource Prometheus |
-| `infrastructure/terraform/backend.tf` | ✅ **NUEVO** | GCS remote state config (listo para habilitar) |
-| `infrastructure/terraform/terraform.tfvars.example` | ✅ **NUEVO** | Template de variables TF |
-| `pyproject.toml` | ✅ **NUEVO** | Centraliza ruff/mypy/pytest/coverage config |
-| `docs/local_development.md` | ✅ **NUEVO** | Guía completa de desarrollo local |
-| `.github/workflows/ci.yml` | ✅ **MEJORADO** | +Job `terraform-validate` (sin credenciales GCP) |
-| `tests/test_health.py` | ✅ **NUEVO** | 9 tests para /health y /metrics endpoints |
-
-**Suite de tests: 54/54 ✅ en 6.96s — Python 3.14 · pytest-asyncio 1.3.0**
+| `models/dispatch_policy.onnx` | ✅ **NUEVO** | Modelo dummy (SOC×0.8). Reemplazar con export de Ray RLlib. |
+| `scripts/generate_dummy_onnx.py` | ✅ **NUEVO** | Genera el modelo dummy + smoke test integrado |
+| `infrastructure/docker/docker-compose.yml` | ✅ Producción | Perfil `monitoring` (Prometheus+Grafana), port 8000 |
+| `infrastructure/prometheus/prometheus.yml` | ✅ Producción | Scrape config: gateway:8000 + otel-collector:8888 |
+| `infrastructure/terraform/` | ✅ Producción | apply ejecutado — 18 recursos en GCP |
+
+**Suite de tests: 73/73 ✅ en 11.89s — Python 3.14 · pytest-asyncio 1.3.0**
 
 ### 🐳 Stack Docker — OPERATIVO
 
@@ -66,13 +64,12 @@ docker compose -f infrastructure/docker/docker-compose.yml --profile simulator -
 
 ### 🟢 Próximo agente — Continuar aquí
 
-**Todos los bloqueadores resueltos.** El pipeline completo está operativo:
-- lint (ruff) → test (54/54) → tf-validate → docker-build → docker-push → Artifact Registry
+**Todos los bloqueadores resueltos.** El pipeline completo está operativo.
 
-**Próxima prioridad — BESSAI v2.0 (Q3 2026):**
-- Edge AI: ONNX Runtime (inferencia offline)
-- AI-IDS: detección de intrusiones Modbus
-- Ver roadmap: `docs/bessai_v2_roadmap.md`
+**Próxima prioridad — BESSAI v0.7.0 (Edge AI Fase 2):**
+- DRL Training: Ray RLlib (PPO/SAC) + Gymnasium + pandapower simulator
+- Federated Learning: Flower (flwr) — solo gradientes salen del edge
+- Ver roadmap: `docs/bessai_v2_roadmap.md` — Fase 2 aún en progreso
 
 ### 📂 Estructura de archivos clave
 ```

models/dispatch_policy.onnx

@@ -54,3 +54,13 @@ prometheus_client>=0.20.0
 # HTTP server — Health check & metrics endpoint
 # ---------------------------------------------------------------------------
 aiohttp>=3.9.5
+
+# ---------------------------------------------------------------------------
+# Edge AI — AI-IDS (Modbus anomaly detection) + ONNX inference engine
+# ---------------------------------------------------------------------------
+# Numerical computing (required by scikit-learn and onnxruntime)
+numpy>=1.26.0
+# IsolationForest for Modbus anomaly detection (AI-IDS)
+scikit-learn>=1.4.0
+# ONNX Runtime for offline dispatch policy inference
+onnxruntime>=1.18.0

requirements.txt

@@ -0,0 +1,106 @@
+"""
+scripts/generate_dummy_onnx.py
+================================
+Generates a minimal ONNX model for local testing of the ONNXDispatcher.
+
+Model: Linear dispatch policy
+  Input:  [soc_pct, power_kw, temp_c, hour_of_day]  (shape: 1x4, float32)
+  Output: [dispatch_target_kw]                        (shape: 1x1, float32)
+  Formula: target_kw = soc_pct * 0.8
+
+This is a placeholder model. In production, the ONNX model is produced by
+training a Ray RLlib policy and exporting it via onnxmltools or torch.onnx.
+
+Usage:
+    python scripts/generate_dummy_onnx.py
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+
+try:
+    import onnx
+    from onnx import TensorProto, helper
+    _ONNX_AVAILABLE = True
+except ImportError:
+    _ONNX_AVAILABLE = False
+
+try:
+    import onnxruntime as ort
+    _ORT_AVAILABLE = True
+except ImportError:
+    _ORT_AVAILABLE = False
+
+
+def create_dummy_onnx_model() -> "onnx.ModelProto":
+    """Create a minimal linear dispatch ONNX model.
+
+    Graph: input (1x4) → MatMul (4x1 weights) → output (1x1)
+    Weights: [0.8, 0.0, 0.0, 0.0] → target_kw ≈ soc_pct * 0.8
+    """
+    # Weight matrix [4, 1]: only soc_pct (index 0) has weight 0.8
+    weights = np.array([[0.8], [0.0], [0.0], [0.0]], dtype=np.float32)
+    weight_tensor = helper.make_tensor(
+        name="W",
+        data_type=TensorProto.FLOAT,
+        dims=[4, 1],
+        vals=weights.flatten().tolist(),
+    )
+
+    # Graph definition
+    input_ = helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 4])
+    output = helper.make_tensor_value_info("output", TensorProto.FLOAT, [1, 1])
+
+    matmul_node = helper.make_node(
+        op_type="MatMul",
+        inputs=["input", "W"],
+        outputs=["output"],
+    )
+
+    graph = helper.make_graph(
+        nodes=[matmul_node],
+        name="DummyDispatchPolicy",
+        inputs=[input_],
+        outputs=[output],
+        initializer=[weight_tensor],
+    )
+
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 17)])
+    model.doc_string = (
+        "Dummy dispatch policy for BESSAI testing. "
+        "target_kw = soc_pct * 0.8. "
+        "Replace with a trained Ray RLlib export in production."
+    )
+    onnx.checker.check_model(model)
+    return model
+
+
+def main() -> None:
+    if not _ONNX_AVAILABLE:
+        print("ERROR: 'onnx' package not installed. Run: pip install onnx")
+        raise SystemExit(1)
+
+    model = create_dummy_onnx_model()
+    output_path = Path("models/dispatch_policy.onnx")
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    onnx.save(model, str(output_path))
+    print(f"[OK] Dummy ONNX model saved: {output_path}")
+
+    # Quick smoke test with onnxruntime
+    if _ORT_AVAILABLE:
+        sess = ort.InferenceSession(str(output_path))
+        test_input = np.array([[90.0, 50.0, 25.0, 14.0]], dtype=np.float32)
+        outputs = sess.run(None, {"input": test_input})
+        target_kw = float(outputs[0][0][0])
+        print(f"  Smoke test -> input SOC=90% -> dispatch_target_kw={target_kw:.2f} kW")
+        assert abs(target_kw - 72.0) < 1.0, f"Expected ~72.0, got {target_kw}"
+        print("  [OK] Smoke test passed.")
+    else:
+        print("  [WARN] onnxruntime not installed - skipping smoke test.")
+
+
+if __name__ == "__main__":
+    main()