AIOE updater characterization (pure simulation, no client data)
params: eta=0.1 beta=0.3 maxStep=0.02 floor=0.05 ceiling=0.4 minEvents=25 windowDays=90

regime=rewards-skills cycles=40 sum_always_1=true clamps_held=true converged=false
  clearance            0.2500 -> 0.3030  (min 0.2500 max 0.3030)
  skills               0.2500 -> 0.3970  (min 0.2500 max 0.3970)
  certifications       0.1500 -> 0.0999  (min 0.0999 max 0.1500)
  customer_experience  0.1500 -> 0.1001  (min 0.1001 max 0.1500)
  experience           0.1000 -> 0.0500  (min 0.0500 max 0.1000)
  location             0.1000 -> 0.0500  (min 0.0500 max 0.1000)

regime=penalizes-location cycles=40 sum_always_1=true clamps_held=true converged=false
  clearance            0.2500 -> 0.3276  (min 0.2500 max 0.3276)
  skills               0.2500 -> 0.3268  (min 0.2500 max 0.3268)
  certifications       0.1500 -> 0.1228  (min 0.1228 max 0.1529)
  customer_experience  0.1500 -> 0.1228  (min 0.1228 max 0.1529)
  experience           0.1000 -> 0.0500  (min 0.0500 max 0.1000)
  location             0.1000 -> 0.0500  (min 0.0500 max 0.1000)

regime=pure-noise cycles=40 sum_always_1=true clamps_held=true converged=true
  clearance            0.2500 -> 0.2500  (min 0.2500 max 0.2500)
  skills               0.2500 -> 0.2500  (min 0.2500 max 0.2500)
  certifications       0.1500 -> 0.1500  (min 0.1500 max 0.1500)
  customer_experience  0.1500 -> 0.1500  (min 0.1500 max 0.1500)
  experience           0.1000 -> 0.1000  (min 0.1000 max 0.1000)
  location             0.1000 -> 0.1000  (min 0.1000 max 0.1000)

regime=adversarial-flip cycles=40 sum_always_1=true clamps_held=true converged=false
  clearance            0.2500 -> 0.3033  (min 0.2500 max 0.3033)
  skills               0.2500 -> 0.3965  (min 0.2500 max 0.3965)
  certifications       0.1500 -> 0.1001  (min 0.1001 max 0.1500)
  customer_experience  0.1500 -> 0.1001  (min 0.1001 max 0.1500)
  experience           0.1000 -> 0.0500  (min 0.0500 max 0.1000)
  location             0.1000 -> 0.0500  (min 0.0500 max 0.1000)

RESULT: every regime kept the sum-to-one invariant and the clamps. The updater is bounded and stable.