rightsize: trim over-provisioned memory (batch 2/N)

claude-agent-service 12Gi->3Gi (peak 585Mi — the single biggest fat, ~9Gi of limit-overcommit removed), job-hunter 1280->768Mi (kept chromium headroom; 30d peak 118Mi), fire-planner 1024->320Mi, wealthfolio 1Gi->512Mi (kept history-growth headroom). Burstable, limits kept >= generous peak headroom, never below peak. ~10.7Gi of limit overcommit removed. paperless-ai intentionally LEFT at 4Gi (documented in-process RAG model load).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

stacks/fire-planner/main.tf

@@ -318,10 +318,10 @@ resource "kubernetes_deployment" "fire_planner" {
           resources {
             requests = {
               cpu    = "100m"
-              memory = "512Mi"
+              memory = "192Mi"
             }
             limits = {
-              memory = "1024Mi"
+              memory = "320Mi"
             }
           }
         }