feat: alert engine stage 1 — rules, events, state machine, CRUD

Replaces the POC single-threshold check with a real per-rule engine over
the live monitor feed.

- AlertRule / AlertEvent tables (auto-created via create_all; no migration).
  Rule = {metric, comparison, threshold_db, duration_s, clear_margin_db,
  schedule, channels, recipients}.
- alerts.py: per-(unit,rule) state machine IDLE->ACTIVE->IDLE with duration
  debounce (both edges) + clear_margin hysteresis; onset/clear are distinct
  events; optional nighttime schedule; rule cache w/ invalidation. The
  state-machine core (_evaluate_step) is pure (no DB/clock) for testing.
- Dispatch is a server log (POC); _dispatch() is the seam for a Terra-View
  webhook (email/SMS) later.
- CRUD: POST/GET/PUT/DELETE /{unit}/alerts/rules, GET /{unit}/alerts/events,
  POST /{unit}/alerts/events/{id}/ack.
- test_alert_evaluator.py: synthetic level series proves onset debounce,
  spike rejection, hysteresis hold, and below-comparison (4/4 pass, no device).

Source-agnostic: the same rules transfer unchanged if a unit's feed is later
sourced from FTP intervals instead of the DOD monitor.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app/alerts.py

@@ -0,0 +1,244 @@
+"""
+Threshold alert engine.
+
+Each unit can have any number of AlertRules. A rule is evaluated against the
+unit's live monitor snapshots via a small per-(unit, rule) state machine:
+
+    IDLE  --(metric exceeds threshold for duration_s)-->  ACTIVE   (fire ONSET)
+    ACTIVE --(metric recovers past hysteresis for duration_s)--> IDLE (fire CLEAR)
+
+duration_s debounces both edges; clear_margin_db adds hysteresis so a level
+hovering at the threshold doesn't flap. Onset and clear are distinct events.
+
+The state-machine logic (`_evaluate_step`) is intentionally pure — no DB, no
+real clock — so it can be unit-tested with a synthetic level series and a fake
+clock. The AlertEvaluator wraps it with rule loading, scheduling, persistence,
+and dispatch. Dispatch is a server log for now (POC); the seam to POST events to
+a Terra-View webhook (email/SMS) is _dispatch().
+"""
+
+import asyncio
+import logging
+import os
+from dataclasses import dataclass
+from datetime import datetime, timedelta
+from typing import Dict, List, Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+# Local timezone offset for schedule windows (same env var services.py uses).
+_TZ_OFFSET_HOURS = float(os.getenv("TIMEZONE_OFFSET", "-5"))
+
+# How long to cache a unit's rules before re-querying the DB (rules change rarely).
+_RULE_CACHE_TTL_S = 15.0
+
+
+@dataclass
+class RuleState:
+    """In-memory runtime state for one (unit, rule)."""
+    phase: str = "idle"                 # "idle" | "active"
+    edge_since: Optional[float] = None  # when the current edge condition began (clock time)
+    peak: float = 0.0
+    event_id: Optional[int] = None      # the open AlertEvent row (for the clear update)
+
+
+def _exceeds(value: float, rule) -> bool:
+    if rule.comparison == "below":
+        return value < rule.threshold_db
+    return value > rule.threshold_db
+
+
+def _recovered(value: float, rule) -> bool:
+    margin = rule.clear_margin_db or 0.0
+    if rule.comparison == "below":
+        return value > rule.threshold_db + margin
+    return value < rule.threshold_db - margin
+
+
+def _evaluate_step(state: RuleState, value: float, now: float, rule) -> Optional[str]:
+    """Advance the state machine by one reading.
+
+    Pure: mutates `state`, returns 'onset' | 'clear' | None. `now` is injected so
+    tests can drive a fake clock.
+    """
+    duration = rule.duration_s or 0
+
+    if state.phase == "idle":
+        if _exceeds(value, rule):
+            if state.edge_since is None:
+                state.edge_since = now
+            if now - state.edge_since >= duration:
+                state.phase = "active"
+                state.edge_since = None
+                state.peak = value
+                return "onset"
+        else:
+            state.edge_since = None
+        return None
+
+    # active
+    if rule.comparison == "below":
+        state.peak = min(state.peak, value)
+    else:
+        state.peak = max(state.peak, value)
+
+    if _recovered(value, rule):
+        if state.edge_since is None:
+            state.edge_since = now
+        if now - state.edge_since >= duration:
+            state.phase = "idle"
+            state.edge_since = None
+            return "clear"
+    else:
+        state.edge_since = None
+    return None
+
+
+def _in_window(now_minutes: int, start: str, end: str) -> bool:
+    """Is now_minutes (minutes since local midnight) within [start, end)?
+    Handles wraparound windows like 22:00–07:00."""
+    def _m(s: str) -> int:
+        h, m = s.split(":")
+        return int(h) * 60 + int(m)
+    s, e = _m(start), _m(end)
+    if s == e:
+        return True
+    if s < e:
+        return s <= now_minutes < e
+    return now_minutes >= s or now_minutes < e  # wraparound
+
+
+class AlertEvaluator:
+    def __init__(self):
+        self._states: Dict[Tuple[str, int], RuleState] = {}
+        self._rule_cache: Dict[str, Tuple[float, list]] = {}  # unit_id -> (fetched_at, rules)
+        logger.info("[ALERT] rule-based evaluator ready")
+
+    async def evaluate(self, unit_id: str, snap) -> None:
+        """Evaluate every enabled rule for this unit against one snapshot."""
+        rules = self._get_rules(unit_id)
+        if not rules:
+            return
+        now = asyncio.get_running_loop().time()
+        for rule in rules:
+            if not self._in_schedule(rule):
+                continue
+            raw = getattr(snap, rule.metric, None)
+            try:
+                value = float(raw)
+            except (TypeError, ValueError):
+                continue  # missing / non-numeric ("-.-")
+            state = self._states.setdefault((unit_id, rule.id), RuleState())
+            action = _evaluate_step(state, value, now, rule)
+            if action == "onset":
+                await self._on_onset(unit_id, rule, value, state)
+            elif action == "clear":
+                await self._on_clear(unit_id, rule, value, state)
+
+    # -- rule loading (cached) ----------------------------------------------
+
+    def _get_rules(self, unit_id: str) -> list:
+        loop_now = asyncio.get_running_loop().time()
+        cached = self._rule_cache.get(unit_id)
+        if cached and loop_now - cached[0] < _RULE_CACHE_TTL_S:
+            return cached[1]
+        rules = self._load_rules(unit_id)
+        self._rule_cache[unit_id] = (loop_now, rules)
+        return rules
+
+    def _load_rules(self, unit_id: str) -> list:
+        from app.database import SessionLocal
+        from app.models import AlertRule
+        db = SessionLocal()
+        try:
+            return db.query(AlertRule).filter_by(unit_id=unit_id, enabled=True).all()
+        except Exception as e:
+            logger.warning(f"[ALERT] failed to load rules for {unit_id}: {e}")
+            return []
+        finally:
+            db.close()
+
+    def invalidate(self, unit_id: Optional[str] = None) -> None:
+        """Drop cached rules so a change is picked up immediately."""
+        if unit_id is None:
+            self._rule_cache.clear()
+        else:
+            self._rule_cache.pop(unit_id, None)
+
+    # -- scheduling ----------------------------------------------------------
+
+    def _in_schedule(self, rule) -> bool:
+        if not rule.schedule_start or not rule.schedule_end:
+            day_ok = self._day_ok(rule)
+            return day_ok
+        local = datetime.utcnow() + timedelta(hours=_TZ_OFFSET_HOURS)
+        if not self._day_ok(rule, local):
+            return False
+        return _in_window(local.hour * 60 + local.minute, rule.schedule_start, rule.schedule_end)
+
+    @staticmethod
+    def _day_ok(rule, local: Optional[datetime] = None) -> bool:
+        if not rule.schedule_days:
+            return True
+        if local is None:
+            local = datetime.utcnow() + timedelta(hours=_TZ_OFFSET_HOURS)
+        allowed = {int(d) for d in str(rule.schedule_days).split(",") if d.strip() != ""}
+        return local.weekday() in allowed  # Mon=0
+
+    # -- event persistence + dispatch ---------------------------------------
+
+    async def _on_onset(self, unit_id: str, rule, value: float, state: RuleState) -> None:
+        from app.database import SessionLocal
+        from app.models import AlertEvent
+        db = SessionLocal()
+        try:
+            evt = AlertEvent(
+                rule_id=rule.id, unit_id=unit_id, rule_name=rule.name,
+                metric=rule.metric, threshold_db=rule.threshold_db,
+                onset_value=value, peak_value=value, status="active",
+            )
+            db.add(evt)
+            db.commit()
+            db.refresh(evt)
+            state.event_id = evt.id
+        except Exception as e:
+            logger.warning(f"[ALERT] failed to record onset for {unit_id}: {e}")
+        finally:
+            db.close()
+        await self._dispatch(
+            "ONSET", unit_id, rule,
+            f"{rule.metric.upper()}={value:.1f} dB "
+            f"{'<' if rule.comparison == 'below' else '>'} {rule.threshold_db:.1f} dB"
+            f"{f' for {rule.duration_s}s' if rule.duration_s else ''}",
+        )
+
+    async def _on_clear(self, unit_id: str, rule, value: float, state: RuleState) -> None:
+        peak = state.peak
+        from app.database import SessionLocal
+        from app.models import AlertEvent
+        db = SessionLocal()
+        try:
+            if state.event_id is not None:
+                evt = db.query(AlertEvent).filter_by(id=state.event_id).first()
+                if evt:
+                    evt.clear_at = datetime.utcnow()
+                    evt.peak_value = peak
+                    evt.status = "cleared"
+                    db.commit()
+        except Exception as e:
+            logger.warning(f"[ALERT] failed to record clear for {unit_id}: {e}")
+        finally:
+            db.close()
+        state.event_id = None
+        await self._dispatch(
+            "CLEAR", unit_id, rule,
+            f"recovered to {value:.1f} dB (peak {peak:.1f} dB)",
+        )
+
+    async def _dispatch(self, kind: str, unit_id: str, rule, detail: str) -> None:
+        """POC dispatch: server log. Swap in a Terra-View webhook (email/SMS) here."""
+        logger.warning(f"[ALERT:{kind}] {unit_id} '{rule.name}': {detail}")
+
+
+# Module-level singleton (the monitor calls alert_evaluator.evaluate per snapshot)
+alert_evaluator = AlertEvaluator()

app/models.py

@@ -1,4 +1,4 @@
-from sqlalchemy import Column, String, DateTime, Boolean, Integer, Text, func
+from sqlalchemy import Column, String, DateTime, Boolean, Integer, Float, Text, func
 from app.database import Base
 
 
@@ -74,3 +74,53 @@ class DeviceLog(Base):
     level = Column(String, default="INFO")  # DEBUG, INFO, WARNING, ERROR
     category = Column(String, default="GENERAL")  # TCP, FTP, POLL, COMMAND, STATE, SYNC
     message = Column(Text, nullable=False)
+
+
+class AlertRule(Base):
+    """A threshold-alert rule evaluated against a unit's live monitor feed.
+
+    Source-agnostic: today it runs over the DOD monitor; the same rule transfers
+    unchanged if a unit's feed is later sourced from FTP intervals.
+    """
+
+    __tablename__ = "alert_rules"
+
+    id = Column(Integer, primary_key=True, autoincrement=True)
+    unit_id = Column(String, index=True, nullable=False)
+    name = Column(String, nullable=False, default="Alert")
+    metric = Column(String, nullable=False, default="lp")  # lp/leq/lmax/lmin/lpeak/ln1/ln2
+    comparison = Column(String, nullable=False, default="above")  # above | below
+    threshold_db = Column(Float, nullable=False)
+    duration_s = Column(Integer, nullable=False, default=0)       # sustained seconds (0 = instant)
+    clear_margin_db = Column(Float, nullable=False, default=2.0)  # hysteresis band
+    cooldown_s = Column(Integer, nullable=False, default=300)     # min seconds between onsets
+    # Optional time-of-day scoping (local time). schedule_start/end as "HH:MM";
+    # null = always active. schedule_days = CSV of 0-6 (Mon=0); null = every day.
+    schedule_start = Column(String, nullable=True)
+    schedule_end = Column(String, nullable=True)
+    schedule_days = Column(String, nullable=True)
+    channels = Column(String, nullable=False, default="log")  # CSV: log,email,sms
+    recipients = Column(Text, nullable=True)                  # CSV of emails/phones
+    enabled = Column(Boolean, default=True)
+    created_at = Column(DateTime, default=func.now())
+
+
+class AlertEvent(Base):
+    """A fired alert (onset → clear), for history / inbox / acknowledgement."""
+
+    __tablename__ = "alert_events"
+
+    id = Column(Integer, primary_key=True, autoincrement=True)
+    rule_id = Column(Integer, index=True, nullable=False)
+    unit_id = Column(String, index=True, nullable=False)
+    rule_name = Column(String, nullable=True)
+    metric = Column(String, nullable=False)
+    threshold_db = Column(Float, nullable=False)
+    onset_at = Column(DateTime, default=func.now(), index=True)
+    onset_value = Column(Float, nullable=True)
+    peak_value = Column(Float, nullable=True)
+    clear_at = Column(DateTime, nullable=True)
+    status = Column(String, default="active")  # active | cleared
+    acknowledged_at = Column(DateTime, nullable=True)
+    acknowledged_by = Column(String, nullable=True)
+    notes = Column(Text, nullable=True)

app/monitor.py

@@ -0,0 +1,176 @@
+"""
+Per-device live monitor (fan-out hub).
+
+ONE DOD poll loop per device, broadcast to many subscribers:
+- browser WebSocket clients (live view) — they no longer each open their own
+  device stream, so the NL43's single-connection limit stops causing the
+  "second viewer sees nothing" contention.
+- the alert evaluator (threshold alerts), which can keep a device's feed running
+  even with no browser attached.
+- persistence (each snapshot is written to NL43Status, like the poller does).
+
+The device's one TCP connection is respected: every poll goes through the same
+per-device lock + connection pool in services.py, so the monitor, the background
+poller, and on-demand commands all serialize safely.
+"""
+
+import asyncio
+import logging
+import os
+from datetime import datetime
+from typing import Dict, Optional, Set
+
+from app.database import SessionLocal
+from app.models import NL43Config, NL43Status
+from app.services import NL43Client, persist_snapshot
+from app.alerts import alert_evaluator
+
+logger = logging.getLogger(__name__)
+
+# Sleep between DOD polls. Note the 1s device rate-limit (and DOD?+Measure? per
+# poll) already paces the effective rate to a few seconds; this is the extra idle.
+MONITOR_POLL_INTERVAL = float(os.getenv("MONITOR_POLL_INTERVAL", "1.0"))
+
+
+def _snapshot_payload(snap, unit_id: str, measurement_start_time) -> dict:
+    """Build the broadcast payload — same shape as the DRD stream, but DOD-sourced
+    so it carries ln1/ln2 (which DRD cannot)."""
+    return {
+        "unit_id": unit_id,
+        "timestamp": datetime.utcnow().isoformat(),
+        "measurement_state": snap.measurement_state,
+        "measurement_start_time": measurement_start_time,
+        "counter": snap.counter,
+        "lp": snap.lp,
+        "leq": snap.leq,
+        "lmax": snap.lmax,
+        "lmin": snap.lmin,
+        "lpeak": snap.lpeak,
+        "ln1": snap.ln1,
+        "ln2": snap.ln2,
+        "raw_payload": snap.raw_payload,
+    }
+
+
+class DeviceMonitor:
+    """Owns a single DOD poll loop for one device and fans each snapshot out to
+    all subscribers. Runs while it has at least one browser subscriber OR the
+    server-side keep-alive (alerting) flag is set."""
+
+    def __init__(self, unit_id: str):
+        self.unit_id = unit_id
+        self._subscribers: Set[asyncio.Queue] = set()
+        self._keepalive = False
+        self._task: Optional[asyncio.Task] = None
+        self._lock = asyncio.Lock()
+
+    @property
+    def running(self) -> bool:
+        return self._task is not None and not self._task.done()
+
+    def subscriber_count(self) -> int:
+        return len(self._subscribers)
+
+    def _has_demand(self) -> bool:
+        return bool(self._subscribers) or self._keepalive
+
+    def _ensure_task(self) -> None:
+        if self._task is None or self._task.done():
+            self._task = asyncio.create_task(self._run())
+
+    async def subscribe(self) -> asyncio.Queue:
+        q: asyncio.Queue = asyncio.Queue(maxsize=5)
+        async with self._lock:
+            self._subscribers.add(q)
+            self._ensure_task()
+        return q
+
+    async def unsubscribe(self, q: asyncio.Queue) -> None:
+        async with self._lock:
+            self._subscribers.discard(q)
+
+    async def set_keepalive(self, on: bool) -> None:
+        async with self._lock:
+            self._keepalive = on
+            if on:
+                self._ensure_task()
+
+    async def _run(self) -> None:
+        logger.info(f"[MONITOR] {self.unit_id}: feed started")
+        try:
+            while self._has_demand():
+                snap, mst = await self._poll_once()
+                if snap is not None:
+                    payload = _snapshot_payload(snap, self.unit_id, mst)
+                    self._broadcast(payload)
+                    try:
+                        await alert_evaluator.evaluate(self.unit_id, snap)
+                    except Exception as e:
+                        logger.warning(f"[MONITOR] {self.unit_id}: alert eval failed: {e}")
+                await asyncio.sleep(MONITOR_POLL_INTERVAL)
+        finally:
+            logger.info(f"[MONITOR] {self.unit_id}: feed stopped")
+
+    async def _poll_once(self):
+        """One DOD poll: read, persist, return (snapshot, measurement_start_iso)."""
+        db = SessionLocal()
+        try:
+            cfg = db.query(NL43Config).filter_by(unit_id=self.unit_id).first()
+            if not cfg or not cfg.tcp_enabled:
+                return None, None
+            client = NL43Client(
+                cfg.host, cfg.tcp_port,
+                ftp_username=cfg.ftp_username, ftp_password=cfg.ftp_password,
+                ftp_port=cfg.ftp_port or 21,
+            )
+            snap = await client.request_dod()
+            snap.unit_id = self.unit_id
+            persist_snapshot(snap, db)
+            db.commit()
+            status = db.query(NL43Status).filter_by(unit_id=self.unit_id).first()
+            mst = (status.measurement_start_time.isoformat()
+                   if status and status.measurement_start_time else None)
+            return snap, mst
+        except Exception as e:
+            logger.warning(f"[MONITOR] {self.unit_id}: poll failed: {e}")
+            return None, None
+        finally:
+            db.close()
+
+    def _broadcast(self, payload: dict) -> None:
+        for q in list(self._subscribers):
+            try:
+                q.put_nowait(payload)
+            except asyncio.QueueFull:
+                # Slow consumer — drop this frame rather than stall the whole feed.
+                pass
+
+
+class MonitorManager:
+    """Registry of per-device monitors (one per unit_id)."""
+
+    def __init__(self):
+        self._monitors: Dict[str, DeviceMonitor] = {}
+        self._lock = asyncio.Lock()
+
+    async def get(self, unit_id: str) -> DeviceMonitor:
+        async with self._lock:
+            m = self._monitors.get(unit_id)
+            if m is None:
+                m = DeviceMonitor(unit_id)
+                self._monitors[unit_id] = m
+            return m
+
+    def status(self) -> dict:
+        return {
+            uid: {
+                "running": m.running,
+                "subscribers": m.subscriber_count(),
+                "keepalive": m._keepalive,
+            }
+            for uid, m in self._monitors.items()
+        }
+
+
+# Module-level singleton
+monitor_manager = MonitorManager()

app/routers.py

@@ -11,7 +11,7 @@ import os
 import asyncio
 
 from app.database import get_db
-from app.models import NL43Config, NL43Status
+from app.models import NL43Config, NL43Status, AlertRule, AlertEvent
 from app.services import NL43Client, persist_snapshot
 
 logger = logging.getLogger(__name__)
@@ -246,6 +246,185 @@ async def system_resume():
     return {"status": "ok", "mode": "active", "message": "Polling resumed"}
 
 
+# ============================================================================
+# LIVE MONITOR (fan-out) — one DOD feed per device, broadcast to many clients
+# ============================================================================
+
+@router.websocket("/{unit_id}/monitor")
+async def monitor_stream(websocket: WebSocket, unit_id: str):
+    """Subscribe a browser to the device's shared 1 Hz DOD feed.
+
+    Any number of clients can attach without each opening its own device
+    connection (one poll loop per device, fanned out). Same JSON shape as the
+    DRD stream, but DOD-sourced so it includes ln1/ln2 (L1/L10).
+    """
+    await websocket.accept()
+    from app.monitor import monitor_manager
+
+    monitor = await monitor_manager.get(unit_id)
+    queue = await monitor.subscribe()
+    logger.info(f"Monitor subscriber attached for {unit_id} ({monitor.subscriber_count()} total)")
+
+    async def _watch_disconnect():
+        # Completes when the client disconnects, so an idle feed (no data) still
+        # detects the drop and we don't leak a subscription that keeps the device
+        # feed (and its connection) alive.
+        try:
+            while True:
+                msg = await websocket.receive()
+                if msg.get("type") == "websocket.disconnect":
+                    return
+        except Exception:
+            return
+
+    gone = asyncio.ensure_future(_watch_disconnect())
+    try:
+        while not gone.done():
+            try:
+                payload = await asyncio.wait_for(queue.get(), timeout=1.0)
+            except asyncio.TimeoutError:
+                continue  # re-check gone.done()
+            await websocket.send_json(payload)
+    except WebSocketDisconnect:
+        logger.info(f"Monitor subscriber disconnected for {unit_id}")
+    except Exception as e:
+        logger.warning(f"Monitor stream error for {unit_id}: {e}")
+    finally:
+        gone.cancel()
+        await monitor.unsubscribe(queue)
+
+
+@router.post("/{unit_id}/monitor/start")
+async def monitor_start(unit_id: str):
+    """Keep the device's feed running even with no browser attached, so alerting
+    evaluates continuously. Runtime-only (resets on restart)."""
+    from app.monitor import monitor_manager
+    monitor = await monitor_manager.get(unit_id)
+    await monitor.set_keepalive(True)
+    return {"status": "ok", "unit_id": unit_id, "running": monitor.running, "keepalive": True}
+
+
+@router.post("/{unit_id}/monitor/stop")
+async def monitor_stop(unit_id: str):
+    """Drop the keep-alive; the feed stops once no browser subscribers remain."""
+    from app.monitor import monitor_manager
+    monitor = await monitor_manager.get(unit_id)
+    await monitor.set_keepalive(False)
+    return {"status": "ok", "unit_id": unit_id, "keepalive": False}
+
+
+@router.get("/_monitor/status")
+async def monitor_status():
+    """Status of every device monitor (running, subscriber count, keep-alive)."""
+    from app.monitor import monitor_manager
+    return {"status": "ok", "monitors": monitor_manager.status()}
+
+
+# ============================================================================
+# ALERTS — threshold rules + fired events
+# ============================================================================
+
+class AlertRulePayload(BaseModel):
+    name: str = "Alert"
+    metric: str = "lp"            # lp/leq/lmax/lmin/lpeak/ln1/ln2
+    comparison: str = "above"     # above | below
+    threshold_db: float
+    duration_s: int = 0           # sustained seconds before firing (0 = instant)
+    clear_margin_db: float = 2.0  # hysteresis band
+    cooldown_s: int = 300
+    schedule_start: str | None = None  # "HH:MM" local; null = always
+    schedule_end: str | None = None
+    schedule_days: str | None = None   # CSV of 0-6 (Mon=0); null = every day
+    channels: str = "log"
+    recipients: str | None = None
+    enabled: bool = True
+
+
+def _rule_dict(r: AlertRule) -> dict:
+    return {
+        "id": r.id, "unit_id": r.unit_id, "name": r.name, "metric": r.metric,
+        "comparison": r.comparison, "threshold_db": r.threshold_db,
+        "duration_s": r.duration_s, "clear_margin_db": r.clear_margin_db,
+        "cooldown_s": r.cooldown_s, "schedule_start": r.schedule_start,
+        "schedule_end": r.schedule_end, "schedule_days": r.schedule_days,
+        "channels": r.channels, "recipients": r.recipients, "enabled": r.enabled,
+    }
+
+
+def _event_dict(e: AlertEvent) -> dict:
+    return {
+        "id": e.id, "rule_id": e.rule_id, "unit_id": e.unit_id,
+        "rule_name": e.rule_name, "metric": e.metric, "threshold_db": e.threshold_db,
+        "onset_at": e.onset_at.isoformat() if e.onset_at else None,
+        "onset_value": e.onset_value, "peak_value": e.peak_value,
+        "clear_at": e.clear_at.isoformat() if e.clear_at else None,
+        "status": e.status,
+        "acknowledged_at": e.acknowledged_at.isoformat() if e.acknowledged_at else None,
+        "acknowledged_by": e.acknowledged_by,
+    }
+
+
+@router.post("/{unit_id}/alerts/rules")
+def create_alert_rule(unit_id: str, payload: AlertRulePayload, db: Session = Depends(get_db)):
+    rule = AlertRule(unit_id=unit_id, **payload.model_dump())
+    db.add(rule)
+    db.commit()
+    db.refresh(rule)
+    from app.alerts import alert_evaluator
+    alert_evaluator.invalidate(unit_id)
+    return {"status": "ok", "rule": _rule_dict(rule)}
+
+
+@router.get("/{unit_id}/alerts/rules")
+def list_alert_rules(unit_id: str, db: Session = Depends(get_db)):
+    rules = db.query(AlertRule).filter_by(unit_id=unit_id).all()
+    return {"status": "ok", "rules": [_rule_dict(r) for r in rules]}
+
+
+@router.put("/{unit_id}/alerts/rules/{rule_id}")
+def update_alert_rule(unit_id: str, rule_id: int, payload: AlertRulePayload, db: Session = Depends(get_db)):
+    rule = db.query(AlertRule).filter_by(id=rule_id, unit_id=unit_id).first()
+    if not rule:
+        raise HTTPException(status_code=404, detail="Alert rule not found")
+    for field, value in payload.model_dump().items():
+        setattr(rule, field, value)
+    db.commit()
+    db.refresh(rule)
+    from app.alerts import alert_evaluator
+    alert_evaluator.invalidate(unit_id)
+    return {"status": "ok", "rule": _rule_dict(rule)}
+
+
+@router.delete("/{unit_id}/alerts/rules/{rule_id}")
+def delete_alert_rule(unit_id: str, rule_id: int, db: Session = Depends(get_db)):
+    rule = db.query(AlertRule).filter_by(id=rule_id, unit_id=unit_id).first()
+    if not rule:
+        raise HTTPException(status_code=404, detail="Alert rule not found")
+    db.delete(rule)
+    db.commit()
+    from app.alerts import alert_evaluator
+    alert_evaluator.invalidate(unit_id)
+    return {"status": "ok", "deleted": rule_id}
+
+
+@router.get("/{unit_id}/alerts/events")
+def list_alert_events(unit_id: str, limit: int = 50, db: Session = Depends(get_db)):
+    events = (db.query(AlertEvent).filter_by(unit_id=unit_id)
+              .order_by(AlertEvent.onset_at.desc()).limit(limit).all())
+    return {"status": "ok", "events": [_event_dict(e) for e in events]}
+
+
+@router.post("/{unit_id}/alerts/events/{event_id}/ack")
+def ack_alert_event(unit_id: str, event_id: int, by: str | None = None, db: Session = Depends(get_db)):
+    evt = db.query(AlertEvent).filter_by(id=event_id, unit_id=unit_id).first()
+    if not evt:
+        raise HTTPException(status_code=404, detail="Alert event not found")
+    evt.acknowledged_at = datetime.utcnow()
+    evt.acknowledged_by = by
+    db.commit()
+    return {"status": "ok", "acknowledged": event_id}
+
+
 # ============================================================================
 # GLOBAL POLLING STATUS ENDPOINT (must be before /{unit_id} routes)
 # ============================================================================

test_alert_evaluator.py

@@ -0,0 +1,68 @@
+"""
+Synthetic unit test for the alert state machine — no DB, no device.
+
+Drives `_evaluate_step` with a fake clock + a level series and checks that
+onset/clear fire with the right debounce + hysteresis. Run:
+
+    docker compose exec -T slmm python3 test_alert_evaluator.py
+    # or, if app.alerts imports cleanly standalone:  python3 test_alert_evaluator.py
+"""
+
+from types import SimpleNamespace
+from app.alerts import RuleState, _evaluate_step
+
+
+def rule(**kw):
+    base = dict(threshold_db=85.0, duration_s=3, clear_margin_db=2.0, comparison="above")
+    base.update(kw)
+    return SimpleNamespace(**base)
+
+
+def run(series, r):
+    st = RuleState()
+    events = [(now, a) for value, now in series
+              if (a := _evaluate_step(st, value, now, r))]
+    return events, st
+
+
+def main():
+    failures = 0
+
+    def check(label, cond, detail=""):
+        nonlocal failures
+        print(("PASS" if cond else "FAIL"), label, detail)
+        if not cond:
+            failures += 1
+
+    # 1) sustained exceedance -> onset after duration; recovery -> clear after duration
+    r = rule(threshold_db=85, duration_s=3, clear_margin_db=2)
+    ev, _ = run([(80, 0), (86, 1), (87, 2), (88, 3), (88, 4),
+                 (88, 5), (82, 6), (82, 7), (82, 8), (82, 9)], r)
+    onsets = [t for t, a in ev if a == "onset"]
+    clears = [t for t, a in ev if a == "clear"]
+    check("1 sustained onset@4 / clear@9", onsets == [4] and clears == [9], str(ev))
+
+    # 2) brief spike under duration -> no onset (debounce)
+    ev, _ = run([(80, 0), (90, 1), (90, 2), (80, 3), (80, 4)], rule(duration_s=3))
+    check("2 brief spike debounced", ev == [], str(ev))
+
+    # 3) hysteresis: a dip into the margin (below threshold, above threshold-margin)
+    #    does NOT clear
+    r = rule(threshold_db=85, duration_s=0, clear_margin_db=3)
+    ev, st = run([(86, 0), (84, 1), (84, 2), (84, 3)], r)
+    check("3 hysteresis holds ACTIVE", ev == [(0, "onset")] and st.phase == "active",
+          f"{ev} phase={st.phase}")
+
+    # 4) 'below' comparison (device too quiet) -> onset when value < threshold
+    ev, _ = run([(30, 0), (15, 1)], rule(threshold_db=20, duration_s=0,
+                                         clear_margin_db=2, comparison="below"))
+    check("4 below-comparison onset@1", ev == [(1, "onset")], str(ev))
+
+    print()
+    print("ALL PASS" if failures == 0 else f"{failures} FAILURE(S)")
+    return failures
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(1 if main() else 0)