feat: production-grade routing cache + stale detection fixes

Three-tier routing cache (ESL global -> file -> HTTP) eliminates HTTP
from call path for 2500-5000 concurrent calls. Lua reads cached route
in ~0.1us instead of blocking on api:execute('curl', ...).

Engine fixes:
- recalcBestLocked now skips stale nodes (was 5 min bug -> now ~10-15s)
- PickNodeForCall action_type='node' checks staleness for consistency
- onMetric callback pushes cache on every metric (no ticker delay)

Config:
- stale_threshold_sec default 20 -> 10 (industry standard)
- contrib/ included in Makefile deploy target
- route.lua paths updated for /opt/pulse-lets-go

internal/engine/engine.go

@@ -456,12 +456,18 @@ func (e *Engine) scoreNodeLocked(ns *models.NodeState, firstMetric bool) {
 
 // recalcBestLocked пересчитывает кэш лучшей ноды.
 // Вызывающий держит write lock.
+// Проверяет staleness (now - ns.TS > staleThreshold) — исключает ноды без свежих метрик.
 func (e *Engine) recalcBestLocked() {
 	bestID := ""
 	bestScore := -999.0
 	anyHealthy := false
+	now := time.Now().Unix()
+	staleThreshold := int64(e.cfg.StaleThresholdSec)
 
 	for id, ns := range e.nodes {
+		if now-ns.TS > staleThreshold {
+			continue
+		}
 		if ns.Score >= 0 {
 			anyHealthy = true
 		}

internal/engine/router.go

@@ -318,8 +318,10 @@ func (e *Engine) PickNodeForCall(callerID, dest, ingress string) (nodeID string,
 		case "node":
 			e.mu.RLock()
 			ns, ok := e.nodes[cr.ActionNodeID]
+			now := time.Now().Unix()
+			staleThreshold := int64(e.cfg.StaleThresholdSec)
 			e.mu.RUnlock()
-			if ok && ns.Score >= 0 {
+			if ok && ns.Score >= 0 && now-ns.TS <= staleThreshold {
 				nodeID, score, matchedRule = cr.ActionNodeID, 100, cr.Name
 				resolved = true
 			}