fix(ai): isolate heavy Ollama workloads from GCP alert lane

apps/api/src/services/ollama_endpoint_resolver.py

@@ -16,27 +16,37 @@ from src.core.config import settings
 OllamaWorkloadType = Literal[
     "interactive",
     "healthcheck",
+    "alert_fast",
     "batch",
     "embedding",
     "rag",
     "code_review",
     "shadow",
     "canary",
+    "deep_rca",
+    "image_analysis",
+    "hermes",
     "local_required",
     "privacy_sensitive",
     "dr",
 ]
 
-_GCP_B_PREFERRED_WORKLOADS = {
+_GCP_A_PREFERRED_WORKLOADS = {
+    "interactive",
+    "healthcheck",
+    "alert_fast",
+}
+
+_LOCAL_PREFERRED_WORKLOADS = {
     "batch",
     "embedding",
     "rag",
     "code_review",
     "shadow",
     "canary",
-}
-
-_LOCAL_PREFERRED_WORKLOADS = {
+    "deep_rca",
+    "image_analysis",
+    "hermes",
     "local_required",
     "privacy_sensitive",
     "dr",
@@ -68,20 +78,28 @@ def resolve_ollama_selection(
     secondary = cfg.OLLAMA_SECONDARY_URL
     fallback = cfg.OLLAMA_FALLBACK_URL
 
-    if workload_type in _GCP_B_PREFERRED_WORKLOADS and secondary:
+    if workload_type in _LOCAL_PREFERRED_WORKLOADS:
+        if fallback:
+            return OllamaEndpointSelection(
+                url=fallback,
+                provider_name="ollama_local",
+                workload_type=workload_type,
+                reason="local_heavy_or_privacy_lane",
+            )
+        if secondary:
+            return OllamaEndpointSelection(
+                url=secondary,
+                provider_name="ollama_gcp_b",
+                workload_type=workload_type,
+                reason="local_missing_gcp_b_fallback",
+            )
+
+    if workload_type not in _GCP_A_PREFERRED_WORKLOADS and secondary:
         return OllamaEndpointSelection(
             url=secondary,
             provider_name="ollama_gcp_b",
             workload_type=workload_type,
-            reason="gcp_b_batch_lane",
-        )
-
-    if workload_type in _LOCAL_PREFERRED_WORKLOADS and fallback:
-        return OllamaEndpointSelection(
-            url=fallback,
-            provider_name="ollama_local",
-            workload_type=workload_type,
-            reason="local_privacy_or_dr_lane",
+            reason="gcp_b_default_non_alert_lane",
         )
 
     return OllamaEndpointSelection(