Completa il hardening autoscaler (Fase 4): budget, kill-switch e runbook.

Drain sicuro in scale-in, alert Ops su overflow e controlli admin senza abilitare il deploy in produzione.

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/spec/services/ops/health_checks_spec.rb

@@ -25,6 +25,7 @@ RSpec.describe Ops::HealthChecks do
       %i[
         check_recordings_size check_postgres check_redis check_mediamtx check_garage
         check_sidekiq_heartbeat check_sidekiq_dead check_http_rails check_rails_latency
+        check_stream_overflow
       ].each do |method|
         allow_any_instance_of(described_class).to receive(method).and_return(
           described_class::Finding.new(
@@ -52,6 +53,7 @@ RSpec.describe Ops::HealthChecks do
       %i[
         check_recordings_size check_postgres check_redis check_mediamtx check_garage
         check_sidekiq_heartbeat check_sidekiq_dead check_http_rails check_rails_latency
+        check_stream_overflow
       ].each do |method|
         allow_any_instance_of(described_class).to receive(method).and_return(
           described_class::Finding.new(
@@ -65,7 +67,7 @@ RSpec.describe Ops::HealthChecks do
       summary = described_class.new.summary
 
       expect(summary[:status]).to eq("ok")
-      expect(summary[:checks].size).to eq(10)
+      expect(summary[:checks].size).to eq(11)
     end
 
     it "degraded quando la latenza p95 supera la soglia warning" do