feat(phase-c): drift-adaptive anomaly detection + smarter eviction scoring

serve_v2: Welford online stats per node (hot_pct, hit_rate, fs_pct,
chunk_miss_rate) updated on every poller tick. Anomaly score = sum of z²;
warn at 9, alert at 16. Exposed via GET /api/v2/anomalies. Score embedded
in MetricSample so sparklines can also surface it.

hot: Replace pure-LRU gc_by_space with size×age eviction scoring.
Candidates are ranked by ln(size) × ln(age_secs); pinned projects immune.
Evicts the most space with the least cost rather than just the oldest entry.

dashboard: Fleet anomaly banner in CommandCenter (red/amber) + per-node
"anomaly"/"drift" badge on NodeCard. Anomalies polled every 10s alongside
the fleet endpoint.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

claw-store/src/hot.rs

@@ -55,25 +55,42 @@ pub fn gc_stale_targets(manifest: &Manifest, stale_hours: u64) -> Result<Vec<Str
     Ok(evicted)
 }
 
+/// Eviction priority score: larger and older = higher = evict first.
+/// log-scaled so a 10GB project doesn't completely dominate a 1GB project
+/// and a week-old entry doesn't mask a 3-day-old one of similar value.
+fn eviction_score(size_bytes: u64, secs_since_active: u64) -> f64 {
+    let size_factor = (size_bytes as f64 + 1.0).ln();
+    let age_factor = (secs_since_active as f64 + 1.0).ln();
+    size_factor * age_factor
+}
+
 pub fn gc_by_space(manifest: &mut Manifest, max_gb: f64) -> Result<Vec<String>> {
+    let now = Utc::now();
     let mut evicted = Vec::new();
     loop {
         let used = total_used_gb(manifest)?;
         if used <= max_gb {
             break;
         }
-        // LRU eviction skips pinned projects — they may NEVER be evicted
+        // Score-based eviction: prefer removing large, stale projects over
-        // for space pressure. The trade-off: if every non-pinned project
+        // arbitrarily picking the oldest-active one. Pinned projects are
-        // is gone and we're still over `max_gb`, we stop and log; better
+        // never evicted regardless of score — operator intent is absolute.
-        // to over-allocate hot than to violate operator intent.
+        let best = manifest
-        let lru_name = manifest
             .projects
             .iter()
-            .filter(|p| !p.pinned)
+            .filter(|p| !p.pinned && p.hot_target_path.exists())
-            .filter(|p| p.hot_target_path.exists())
+            .map(|p| {
-            .min_by_key(|p| p.last_active)
+                let size = target_size_bytes(&p.hot_target_path).unwrap_or(0);
-            .map(|p| p.name.clone());
+                let age_secs = p
-        match lru_name {
+                    .last_active
+                    .map(|t| (now - t).num_seconds().max(0) as u64)
+                    .unwrap_or(0);
+                (p.name.clone(), eviction_score(size, age_secs))
+            })
+            .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
+            .map(|(name, _)| name);
+
+        match best {
             None => {
                 tracing::warn!(
                     used_gb = used,

claw-store/src/serve_v2.rs

@@ -34,9 +34,99 @@ use crate::cluster::transport::{NodeIdentity, QuicClient};
 use crate::config::{Config, PeerEntry, TokenEntry};
 use crate::sessions::{LeasedTag, Session, SessionStore};
 
-// ── metrics history ring buffer ──────────────────────────────────
+// ── metrics history ring buffer + drift-adaptive anomaly detection ──
 
 const HISTORY_MAX_SAMPLES: usize = 1440; // 24h at 1-min resolution
+/// Require at least this many samples before reporting an anomaly score.
+/// Protects against false positives on a freshly-started daemon.
+const ANOMALY_MIN_SAMPLES: u64 = 10;
+/// Sum-of-z² above this value triggers a "warn" level anomaly.
+/// Interpretation: average ~1.5σ deviation across 4 tracked metrics.
+const ANOMALY_WARN_THRESHOLD: f64 = 9.0;
+/// Sum-of-z² above this value triggers "alert".
+const ANOMALY_ALERT_THRESHOLD: f64 = 16.0;
+
+/// Welford's online algorithm for a running mean + variance.
+/// Numerically stable, O(1) per update.
+#[derive(Default, Clone)]
+struct Welford {
+    n: u64,
+    mean: f64,
+    m2: f64,
+}
+
+impl Welford {
+    fn update(&mut self, x: f64) {
+        self.n += 1;
+        let delta = x - self.mean;
+        self.mean += delta / self.n as f64;
+        self.m2 += delta * (x - self.mean);
+    }
+
+    fn stddev(&self) -> f64 {
+        if self.n < 2 { return 1.0; }
+        (self.m2 / (self.n - 1) as f64).sqrt().max(1e-9)
+    }
+
+    /// Squared z-score for a new observation (non-destructive).
+    fn z_sq(&self, x: f64) -> f64 {
+        if self.n < ANOMALY_MIN_SAMPLES { return 0.0; }
+        let z = (x - self.mean) / self.stddev();
+        z * z
+    }
+}
+
+/// Per-node rolling stats for drift-adaptive anomaly detection.
+#[derive(Default, Clone)]
+struct NodeAnomalyStats {
+    hot_pct: Welford,
+    hit_rate: Welford,
+    fs_pct: Welford,
+    chunk_miss_rate: Welford,
+    /// Most recent anomaly score (sum of z²); 0 until ANOMALY_MIN_SAMPLES.
+    pub last_score: f64,
+    /// Total samples seen (mirrors hot_pct.n for convenience).
+    pub n: u64,
+}
+
+impl NodeAnomalyStats {
+    fn update(&mut self, s: &MetricSample) {
+        let hot_pct = if s.hot_max_bytes > 0 {
+            s.hot_used_bytes as f64 / s.hot_max_bytes as f64
+        } else {
+            0.0
+        };
+        let fs_pct = if s.fs_total_bytes > 0 {
+            s.fs_used_bytes as f64 / s.fs_total_bytes as f64
+        } else {
+            0.0
+        };
+        let total_chunks = s.has_chunk_hits + s.has_chunk_misses;
+        let chunk_miss_rate = if total_chunks > 0 {
+            s.has_chunk_misses as f64 / total_chunks as f64
+        } else {
+            0.0
+        };
+
+        self.last_score = self.hot_pct.z_sq(hot_pct)
+            + self.hit_rate.z_sq(s.cache_hit_rate)
+            + self.fs_pct.z_sq(fs_pct)
+            + self.chunk_miss_rate.z_sq(chunk_miss_rate);
+
+        self.hot_pct.update(hot_pct);
+        self.hit_rate.update(s.cache_hit_rate);
+        self.fs_pct.update(fs_pct);
+        self.chunk_miss_rate.update(chunk_miss_rate);
+        self.n = self.hot_pct.n;
+    }
+
+    fn level(&self) -> &'static str {
+        if self.n < ANOMALY_MIN_SAMPLES { return "ok"; }
+        if self.last_score >= ANOMALY_ALERT_THRESHOLD { "alert" }
+        else if self.last_score >= ANOMALY_WARN_THRESHOLD { "warn" }
+        else { "ok" }
+    }
+}
 
 /// One time-series sample snapshotted from a peer's DashboardStatus RPC.
 #[derive(Debug, Clone, Serialize)]
@@ -52,18 +142,41 @@ pub struct MetricSample {
     pub fs_used_bytes: u64,
     pub fs_total_bytes: u64,
     pub fs_available_bytes: u64,
+    /// Drift-adaptive anomaly score (sum of z² across 4 metrics).
+    /// `None` until the node has collected ANOMALY_MIN_SAMPLES.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub anomaly_score: Option<f64>,
+}
+
+#[derive(Serialize)]
+pub struct AnomalyStatus {
+    pub node: String,
+    pub score: f64,
+    pub level: &'static str,
+    pub samples_used: u64,
 }
 
 pub struct MetricsHistory {
     samples: HashMap<String, VecDeque<MetricSample>>,
+    anomaly: HashMap<String, NodeAnomalyStats>,
 }
 
 impl MetricsHistory {
     fn new() -> Self {
-        Self { samples: HashMap::new() }
+        Self {
+            samples: HashMap::new(),
+            anomaly: HashMap::new(),
+        }
     }
 
-    fn push(&mut self, node: &str, sample: MetricSample) {
+    fn push(&mut self, node: &str, mut sample: MetricSample) {
+        // Update Welford stats and embed anomaly score in the sample.
+        let stats = self.anomaly.entry(node.to_string()).or_default();
+        stats.update(&sample);
+        if stats.n >= ANOMALY_MIN_SAMPLES {
+            sample.anomaly_score = Some(stats.last_score);
+        }
+
         let deque = self.samples.entry(node.to_string()).or_default();
         deque.push_back(sample);
         while deque.len() > HISTORY_MAX_SAMPLES {
@@ -81,6 +194,21 @@ impl MetricsHistory {
             })
             .unwrap_or_default()
     }
+
+    pub fn anomaly_statuses(&self) -> Vec<AnomalyStatus> {
+        let mut out: Vec<AnomalyStatus> = self
+            .anomaly
+            .iter()
+            .map(|(node, s)| AnomalyStatus {
+                node: node.clone(),
+                score: s.last_score,
+                level: s.level(),
+                samples_used: s.n,
+            })
+            .collect();
+        out.sort_by(|a, b| a.node.cmp(&b.node));
+        out
+    }
 }
 
 // ─────────────────────────────────────────────────────────────────
@@ -1375,6 +1503,8 @@ async fn metrics_poller(state: Arc<V2State>) {
                         fs_used_bytes: r.filesystem.as_ref().map(|f| f.used_bytes).unwrap_or(0),
                         fs_total_bytes: r.filesystem.as_ref().map(|f| f.total_bytes).unwrap_or(0),
                         fs_available_bytes: r.filesystem.as_ref().map(|f| f.available_bytes).unwrap_or(0),
+                        // Filled in by MetricsHistory::push after Welford update.
+                        anomaly_score: None,
                     };
                     let mut hist = state.history.lock().await;
                     hist.push(&peer_name, sample);
@@ -1402,6 +1532,11 @@ async fn handle_metrics_history(
     Json(hist.get_last(&name, limit))
 }
 
+async fn handle_anomalies(State(s): State<Arc<V2State>>) -> Json<Vec<AnomalyStatus>> {
+    let hist = s.history.lock().await;
+    Json(hist.anomaly_statuses())
+}
+
 // ─────────────────────────────────────────────────────────────────
 
 pub fn build(state: Arc<V2State>) -> Router {
@@ -1426,6 +1561,7 @@ pub fn build(state: Arc<V2State>) -> Router {
         .route("/api/v2/fleet", get(handle_fleet))
         .route("/api/v2/node/:name/status", get(handle_node_status))
         .route("/api/v2/node/:name/metrics-history", get(handle_metrics_history))
+        .route("/api/v2/anomalies", get(handle_anomalies))
         .route("/api/v2/storage/blobs", get(handle_blobs))
         .route("/api/v2/storage/tags", get(handle_tags))
         .route("/api/v2/storage/refs", get(handle_refs))

dashboard-v2/src/components/NodeCard.tsx

@@ -5,12 +5,13 @@ import { NodeHistorySparklines } from './NodeHistorySparklines';
 
 interface Props {
   node: NodeStatusV2;
+  anomalyLevel?: 'ok' | 'warn' | 'alert';
 }
 
 /// Human-oriented node card for the FleetHealth landing.
 /// Shows: overall health traffic-light, storage bars, mount state,
 /// cache hit rate, next scheduled job. No hex, no primitives.
-export function NodeCard({ node }: Props) {
+export function NodeCard({ node, anomalyLevel }: Props) {
   const health = healthOf(node);
   const border = {
     ok: 'border-emerald-700 hover:border-emerald-500',
@@ -41,6 +42,16 @@ export function NodeCard({ node }: Props) {
             <span className="text-lg font-semibold text-slate-100">
               {node.node_name}
             </span>
+            {anomalyLevel === 'alert' && (
+              <span className="text-xs font-mono px-1.5 py-0.5 rounded bg-red-900/60 text-red-300 border border-red-800">
+                anomaly
+              </span>
+            )}
+            {anomalyLevel === 'warn' && (
+              <span className="text-xs font-mono px-1.5 py-0.5 rounded bg-amber-900/60 text-amber-300 border border-amber-800">
+                drift
+              </span>
+            )}
           </div>
           <span className="text-xs text-slate-500 font-mono">
             {node.zone || '—'}

dashboard-v2/src/lib/api.ts

@@ -155,6 +155,15 @@ export interface MetricSample {
   fs_used_bytes: number;
   fs_total_bytes: number;
   fs_available_bytes: number;
+  /** Sum of z² across 4 drift-adapted metrics. Absent until 10 samples collected. */
+  anomaly_score?: number;
+}
+
+export interface AnomalyStatus {
+  node: string;
+  score: number;
+  level: 'ok' | 'warn' | 'alert';
+  samples_used: number;
 }
 
 export const api = {
@@ -163,6 +172,7 @@ export const api = {
   nodeStatus: (name: string) => get<NodeStatusV2>(`/v2/node/${name}/status`),
   metricsHistory: (name: string, limit = 60) =>
     get<MetricSample[]>(`/v2/node/${name}/metrics-history?limit=${limit}`),
+  anomalies: () => get<AnomalyStatus[]>('/v2/anomalies'),
   blobs: (limit = 200, offset = 0) =>
     get<BlobSummary[]>(`/v2/storage/blobs?limit=${limit}&offset=${offset}`),
   tags: (prefix = '') =>

dashboard-v2/src/pages/CommandCenter.tsx

@@ -1,5 +1,5 @@
 import { useEffect, useState } from 'react';
-import { api, FleetSnapshot, fmtBytes, fmtAge } from '../lib/api';
+import { api, FleetSnapshot, AnomalyStatus, fmtBytes, fmtAge } from '../lib/api';
 import { NodeCard } from '../components/NodeCard';
 import { ProjectsPanel } from '../components/ProjectsPanel';
 
@@ -7,6 +7,7 @@ import { ProjectsPanel } from '../components/ProjectsPanel';
 // Polls the aggregator's /api/v2/fleet every 10 s.
 export function CommandCenter() {
   const [fleet, setFleet] = useState<FleetSnapshot | null>(null);
+  const [anomalies, setAnomalies] = useState<AnomalyStatus[]>([]);
   const [err, setErr] = useState<string | null>(null);
   const [tick, setTick] = useState(0);
 
@@ -18,6 +19,7 @@ export function CommandCenter() {
         setErr(null);
       })
       .catch((e) => setErr(String(e)));
+    api.anomalies().then(setAnomalies).catch(() => {});
   }, [tick]);
 
   useEffect(() => {
@@ -39,6 +41,13 @@ export function CommandCenter() {
       )
     : null;
 
+  // Build a node→level lookup for NodeCard props.
+  const anomalyMap = Object.fromEntries(
+    anomalies.map((a) => [a.node, a.level] as const)
+  );
+  const alertNodes = anomalies.filter((a) => a.level === 'alert');
+  const warnNodes = anomalies.filter((a) => a.level === 'warn');
+
   return (
     <div className="space-y-6">
       <div>
@@ -65,6 +74,31 @@ export function CommandCenter() {
         </div>
       )}
 
+      {/* Anomaly banners (only shown once the ring buffer has 10+ samples) */}
+      {alertNodes.length > 0 && (
+        <div className="rounded border border-red-800 bg-red-950/40 px-4 py-3 text-sm flex items-start gap-3">
+          <span className="text-red-400 font-bold mt-0.5">●</span>
+          <div>
+            <span className="text-red-300 font-semibold">Metric anomaly detected — </span>
+            <span className="text-red-200">
+              {alertNodes.map((a) => a.node).join(', ')} deviating &gt;2σ from baseline
+              {alertNodes.length === 1 && ` (score ${alertNodes[0].score.toFixed(1)})`}
+            </span>
+          </div>
+        </div>
+      )}
+      {alertNodes.length === 0 && warnNodes.length > 0 && (
+        <div className="rounded border border-amber-800 bg-amber-950/40 px-4 py-3 text-sm flex items-start gap-3">
+          <span className="text-amber-400 font-bold mt-0.5">●</span>
+          <div>
+            <span className="text-amber-300 font-semibold">Metric drift — </span>
+            <span className="text-amber-200">
+              {warnNodes.map((a) => a.node).join(', ')} showing unusual patterns
+            </span>
+          </div>
+        </div>
+      )}
+
       {totals && totals.diskTotal > 0 && (
         <div className="rounded-lg border border-slate-800 bg-slate-900/60 p-4 flex items-center justify-between">
           <div>
@@ -89,7 +123,11 @@ export function CommandCenter() {
         <h2 className="text-lg font-semibold text-slate-100 mb-3">Nodes</h2>
         <div className="grid grid-cols-1 sm:grid-cols-2 lg:grid-cols-3 gap-4">
           {fleet?.nodes.map((n) => (
-            <NodeCard key={n.node_name} node={n} />
+            <NodeCard
+              key={n.node_name}
+              node={n}
+              anomalyLevel={anomalyMap[n.node_name]}
+            />
           ))}
           {!fleet &&
             [1, 2, 3].map((i) => (