Tighten health stats

Author: jahooma

web/src/server/free-session/fireworks-health.ts

@@ -18,12 +18,12 @@ export type FireworksHealth = 'healthy' | 'degraded' | 'unhealthy'
 /** Degrade once median prefill-queue latency crosses this bound. Strict by
  *  design — a 1s queue on top of ~1s prefill already means users feel 2s+
  *  before first token. */
-export const PREFILL_QUEUE_DEGRADED_MS = 600
+export const PREFILL_QUEUE_DEGRADED_MS = 200
 
 /** Leading indicator of load — responds instantly to memory pressure, while
  *  prefill-queue p50 is a lagging window statistic. Degrading here lets us
  *  halt admission *before* users feel it. */
-export const KV_BLOCKS_DEGRADED_FRACTION = 0.9
+export const KV_BLOCKS_DEGRADED_FRACTION = 0.8
 
 /** Hard backstop: if KV block memory gets this full, evictions dominate and
  *  even the median request will start stalling. */