fix(alerts): bypass proxy timeout for GCP Ollama

docs/adr/ADR-125-gcp-ollama-private-mesh-inference-gateway.md

@@ -58,6 +58,14 @@ Ollama endpoints after cutover:
 The current `192.168.0.110:11435/11436` nginx proxy remains an emergency bridge
 only until the mesh cutover passes shadow and canary gates.
 
+2026-05-06 operational amendment: the active production env temporarily uses the
+GCP public Ollama IPs directly (`34.143.170.20:11434`,
+`34.21.145.224:11434`) because the 110 bridge has an older duplicate
+`/etc/nginx/conf.d/ollama-gcp-proxy.conf` with a 120s read timeout that returns
+504 before `qwen3:14b` can finish. This is a stopgap, not the target transport.
+The next durable transport remains WireGuard mesh plus the AwoooP Inference
+Gateway.
+
 ### D2 - Public Ollama exposure is forbidden after cutover
 
 After mesh cutover:
@@ -110,10 +118,13 @@ gateway explicitly opens a maintenance window.
 
 ## Migration Plan
 
-### Phase 0 - Current bridge
+### Phase 0 - Current bridge / direct stopgap
 
 - Keep `192.168.0.110:11435` and `192.168.0.110:11436` active.
-- Alert path uses `ALERT_OLLAMA_MODEL=gemma3:4b`.
+- Active production may use direct GCP IPs until the 110 bridge timeout conflict
+  is removed.
+- Alert path uses `ALERT_OLLAMA_MODEL=qwen3:14b` when the user prioritizes
+  problem resolution quality over fast card delivery.
 - Gemini remains paid emergency fallback only.
 
 ### Phase 1 - Mesh build in parallel
@@ -184,4 +195,3 @@ Paid provider fallback must remain budget-gated.
 - AwoooP can manage Ollama as a platform resource shared by all tenants.
 - CPU-only GCP performance remains a capacity constraint; routing must keep
   heavy jobs off the alert lane or use GPU-capable GCP nodes.
-