feat(monitoring): add LLM Grafana dashboard, Prometheus scraping, and alerts

Wire mana-llm service into the monitoring stack:

Prometheus (docker/prometheus/prometheus.yml):
- Add mana-llm scrape job (port 3025, 15s interval)
- Include mana-llm in ServiceDown alert expression

Alerts (docker/prometheus/alerts.yml):
- New llm_alerts group with 4 rules:
  - LLMServiceDown: mana-llm down > 1 min (critical)
  - LLMHighErrorRate: > 10% errors for 5 min (warning)
  - OllamaProviderDown: > 50% requests via Google fallback (warning)
  - LLMSlowResponses: p95 > 30s for 5 min (warning)

Grafana Dashboard (docker/grafana/dashboards/mana-llm.json):
- 6 stat panels: status, req/min, error rate, fallback rate, latency, tokens/min
- Requests by Provider (stacked area: Ollama vs Google vs OpenRouter)
- Tokens by Type (prompt vs completion)
- Latency Percentiles (p50, p90, p99)
- Latency by Provider comparison
- Requests by Model breakdown
- Errors by Type
- Google Fallback Rate over time (with threshold coloring)
- Provider Distribution pie chart (24h)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docker/prometheus/alerts.yml

@@ -3,7 +3,7 @@ groups:
     rules:
       # Service Down Alert
       - alert: ServiceDown
-        expr: up{job=~"mana-core-auth|.*-backend|mana-search|mana-media|synapse"} == 0
+        expr: up{job=~"mana-core-auth|.*-backend|mana-search|mana-media|mana-llm|synapse"} == 0
         for: 1m
         labels:
           severity: critical
@@ -356,3 +356,50 @@ groups:
         annotations:
           summary: "OIDC token endpoint errors"
           description: "OIDC token endpoint is returning 5xx errors. SSO may be affected."
+
+  - name: llm_alerts
+    rules:
+      # mana-llm Down
+      - alert: LLMServiceDown
+        expr: up{job="mana-llm"} == 0
+        for: 1m
+        labels:
+          severity: critical
+        annotations:
+          summary: "mana-llm service is down"
+          description: "mana-llm has been down for more than 1 minute. All AI features will fail."
+
+      # High LLM Error Rate (> 10%)
+      - alert: LLMHighErrorRate
+        expr: |
+          sum(rate(mana_llm_llm_errors_total[5m]))
+          / (sum(rate(mana_llm_llm_requests_total[5m])) + 0.001) > 0.1
+        for: 5m
+        labels:
+          severity: warning
+        annotations:
+          summary: "High LLM error rate"
+          description: "{{ $value | humanizePercentage }} of LLM requests are failing."
+
+      # Ollama Provider Down (all requests going to fallback)
+      - alert: OllamaProviderDown
+        expr: |
+          sum(rate(mana_llm_llm_requests_total{provider="google"}[5m]))
+          / (sum(rate(mana_llm_llm_requests_total[5m])) + 0.001) > 0.5
+        for: 10m
+        labels:
+          severity: warning
+        annotations:
+          summary: "Ollama appears down — most requests going to Google fallback"
+          description: "{{ $value | humanizePercentage }} of LLM requests are using Google Gemini fallback."
+
+      # LLM Slow Responses (p95 > 30s)
+      - alert: LLMSlowResponses
+        expr: |
+          histogram_quantile(0.95, sum(rate(mana_llm_llm_latency_seconds_bucket[5m])) by (le)) > 30
+        for: 5m
+        labels:
+          severity: warning
+        annotations:
+          summary: "LLM responses are slow"
+          description: "LLM p95 latency is {{ $value | humanizeDuration }}."

docker/prometheus/prometheus.yml

@@ -158,6 +158,13 @@ scrape_configs:
   # Core Services
   # ============================================
 
+  # Mana LLM Gateway (Ollama + Google Fallback)
+  - job_name: 'mana-llm'
+    static_configs:
+      - targets: ['mana-llm:3025']
+    metrics_path: '/metrics'
+    scrape_interval: 15s
+
   # Mana Search Service
   - job_name: 'mana-search'
     static_configs: