diff --git a/docs/plans/2026-02-13-centralized-log-collection-design.md b/docs/plans/2026-02-13-centralized-log-collection-design.md
new file mode 100644
index 00000000..60940345
--- /dev/null
+++ b/docs/plans/2026-02-13-centralized-log-collection-design.md
@@ -0,0 +1,140 @@
+# Centralized Log Collection Design
+
+## Date: 2026-02-13
+
+## Goal
+
+Centrally collect logs from all Kubernetes pods for monitoring and alerting. Minimize disk I/O by holding logs in memory for extended periods, flushing to NFS once daily. Alert on log patterns via existing Alertmanager pipeline.
+
+## Requirements
+
+- **Primary use case**: Monitoring and alerting (log-based alert rules evaluated in real-time)
+- **Retention**: 7 days on disk after flush
+- **Memory budget**: 4-8GB total (~6.6GB used)
+- **Disk strategy**: 24h in-memory chunks, WAL on tmpfs, single daily flush to NFS
+- **Crash policy**: Accept up to 24h log loss on pod/node crash (alerts still fire in real-time before flush)
+- **Alert delivery**: Loki Ruler -> existing Alertmanager -> Slack/email
+
+## Architecture
+
+```
+┌──────────────────┐     ┌──────────────────────┐     ┌──────────────┐
+│ Alloy DaemonSet  │     │ Loki SingleBinary     │     │ Grafana       │
+│ 5 pods, 128Mi ea │────>│ 1 pod, 6Gi RAM        │<────│ (existing)    │
+│ tails /var/log/  │     │                        │     │ + Loki        │
+│ pods on each node│     │ Ingester: 24h chunks   │     │   datasource  │
+└──────────────────┘     │ WAL: tmpfs (in-memory) │     └──────────────┘
+                         │ Storage: NFS 15Gi      │
+┌──────────────────┐     │ Ruler ──> Alertmanager │
+│ Sysctl DaemonSet │     └──────────────────────┘
+│ 5 pods (pause)   │
+│ sets inotify     │
+│ limits on nodes  │
+└──────────────────┘
+```
+
+## Components
+
+### 1. Sysctl DaemonSet
+
+Solves the `too many open files` / fsnotify watcher exhaustion problem that previously blocked Alloy.
+
+- Privileged init container runs `sysctl -w` on each node
+- Settings: `fs.inotify.max_user_watches=1048576`, `fs.inotify.max_user_instances=512`, `fs.inotify.max_queued_events=1048576`
+- Main container: `pause` image (near-zero resources)
+- Survives node reboots (DaemonSet recreates pod)
+- Namespace: `monitoring`
+
+### 2. Loki (Helm Release)
+
+Single-binary deployment. Existing Helm chart config in `loki.yaml`, updated with:
+
+**Ingester tuning (disk-friendly):**
+- `chunk_idle_period: 12h` — don't flush idle streams quickly
+- `max_chunk_age: 24h` — hold chunks in memory for full day
+- `chunk_retain_period: 1m` — brief retain after flush
+- `chunk_target_size: 1572864` (1.5MB) — larger chunks = fewer writes
+- WAL: tmpfs emptyDir (`medium: Memory`, 2Gi limit)
+
+**Retention:**
+- `retention_period: 168h` (7 days)
+- Compactor enabled for retention enforcement
+
+**Ruler:**
+- Evaluates LogQL alert rules in real-time (before chunk flush)
+- Fires to `http://alertmanager.monitoring.svc.cluster.local:9093`
+
+**Storage:**
+- NFS PV/PVC at `/mnt/main/loki/loki` (15Gi, existing)
+- TSDB index with 24h period
+
+**Resources:**
+- Memory: 6Gi limit
+- CPU: 500m limit
+
+### 3. Alloy (Helm Release)
+
+DaemonSet log collector. Existing config in `alloy.yaml` is complete:
+- Discovers pods via `discovery.kubernetes`
+- Labels: namespace, pod, container, app, job, container_runtime, cluster
+- Tails `/var/log/pods/` on each node
+- Forwards to `http://loki.monitoring.svc.cluster.local:3100/loki/api/v1/push`
+
+**Resources per pod:**
+- Memory: 128Mi limit
+- CPU: 100m limit
+
+### 4. Grafana Datasource
+
+ConfigMap with label `grafana_datasource: "1"` for sidecar auto-discovery:
+- Name: Loki
+- Type: loki
+- URL: `http://loki.monitoring.svc.cluster.local:3100`
+- Existing `loki.json` dashboard already in dashboards directory
+
+### 5. Starter Alert Rules
+
+Configured in Loki Ruler (evaluated in real-time, before disk flush):
+
+| Alert | LogQL Expression | Severity |
+|-------|-----------------|----------|
+| HighErrorRate | `sum(rate({namespace=~".+"} \|= "error" [5m])) by (namespace) > 10` | warning |
+| PodCrashLoopBackOff | `count_over_time({namespace=~".+"} \|= "CrashLoopBackOff" [5m]) > 0` | critical |
+| OOMKilled | `count_over_time({namespace=~".+"} \|= "OOMKilled" [5m]) > 0` | critical |
+
+## Memory Budget
+
+| Component | Per-pod | Pods | Total |
+|-----------|---------|------|-------|
+| Alloy | 128Mi | 5 | 640Mi |
+| Loki | 6Gi | 1 | 6Gi |
+| Sysctl DS | ~0 (pause) | 5 | ~0 |
+| **Total** | | | **~6.6 GB** |
+
+## Files to Change
+
+| File | Action |
+|------|--------|
+| `modules/kubernetes/monitoring/loki.tf` | Uncomment Loki + Alloy helm releases, add sysctl DaemonSet, add Grafana Loki datasource ConfigMap |
+| `modules/kubernetes/monitoring/loki.yaml` | Update with ingester tuning, ruler config, retention, resource limits |
+| `modules/kubernetes/monitoring/alloy.yaml` | Add resource limits in Helm values wrapper |
+| `secrets/nfs_directories.txt` | Ensure `/mnt/main/loki` entries exist |
+
+## Implementation Steps
+
+1. Add sysctl DaemonSet to `loki.tf`
+2. Update `loki.yaml` with disk-friendly tuning, ruler, retention, resources
+3. Update `alloy.yaml` with resource limits
+4. Uncomment Loki Helm release in `loki.tf`, wire up NFS PV/PVC
+5. Uncomment Alloy Helm release in `loki.tf`
+6. Add Grafana Loki datasource ConfigMap to `loki.tf`
+7. Add alert rules to Loki config
+8. Ensure NFS exports exist in `secrets/nfs_directories.txt`
+9. `terraform apply -target=module.kubernetes_cluster.module.monitoring`
+10. Verify: Grafana Explore -> Loki datasource -> query `{namespace="monitoring"}`
+
+## Risks
+
+- **24h data loss on crash**: Accepted trade-off. Alerts fire in real-time before flush, so alert coverage is not affected — only historical log browsing is at risk.
+- **Memory pressure**: 6Gi for Loki on a 16GB node is significant. Monitor with existing Prometheus memory alerts.
+- **Log volume spikes**: A chatty pod could cause Loki to OOM. Alloy can be configured with rate limiting if needed (future enhancement).