fix(deploy): size MCP reservation from observed CPU usage

Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a07ff8-19d0-7820-b4d0-1353833cb7fc

deploy/railiance/apps/helm/state-hub-values.yaml

@@ -17,6 +17,11 @@ ingress:
 # ClusterIP only — no ingress, same reason the API's is disabled.
 mcp:
   enabled: true
+  resources:
+    requests:
+      # 2026-09-10: 1,078 observed 5-minute samples, p99 2.35m, peak 30.45m.
+      # Preserve the 500m limit for bursts; free 40m for the API rollout surge.
+      cpu: 10m
 
 config:
   # This deployment is the authoritative hub. Callers verify this rather than

workplans/STATE-WP-0090-projection-task-field-convergence.md

@@ -74,3 +74,14 @@ Offline container verification matches the tested parser source and passes neste
 section/empty-note cases. Server-side Helm preview permits only image, version
 labels and their rendered checksum; configuration and credential references are
 unchanged, and there are no migration changes. Previous Helm revision: 60.
+
+Rollout revision 61 could not schedule the API surge pod: the 4,000m node had
+3,935m requested, leaving 65m against the required 100m. Atomic rollback restored
+healthy old-image services at revision 62. Available Prometheus history contains
+1,078 five-minute samples within the seven-day query window (partial coverage):
+MCP p99 2.35m, peak 30.45m. The reviewed retry lowers only the MCP CPU request
+from 50m to 10m, preserving one replica, its 500m CPU limit and memory settings.
+This reduces its guaranteed share under contention; bursts remain permitted.
+Nominal headroom becomes 105m, sufficient for this 100m API surge but too narrow
+to count as factory capacity admission. Broader capacity planning needs a live
+residual before closeout. No unrelated workload allocation is changed.