feat(httpapi): harden control plane and speed up prefix refresh

Очередь job получила lease и fencing, телеметрия и прогресс deploy сохраняются, ревизия пишет manifest входов. Обновление префиксов не переписывает неизменённый снимок, большие наборы идут через COPY, а устаревшая ревизия не попадает в deploy. Бандл ноды кэшируется, prune и постановка job больше не держат горячий путь.

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

internal/pipeline/revision_prune.go

@@ -0,0 +1,108 @@
+package pipeline
+
+import (
+	"context"
+	"strings"
+	"sync"
+	"time"
+
+	"evobgp/internal/store"
+)
+
+const (
+	revisionPruneMinInterval = 5 * time.Minute
+	revisionPruneTimeout     = 2 * time.Minute
+)
+
+// revisionPruneGate runs at most one prune per tenant per interval and drops
+// overlapping attempts so renders do not stack prune transactions.
+type revisionPruneGate struct {
+	mu       sync.Mutex
+	last     map[string]time.Time
+	inflight map[string]chan struct{}
+	now      func() time.Time
+}
+
+var revisionPrunes revisionPruneGate
+
+func (g *revisionPruneGate) start(tenantID string) bool {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	now := time.Now()
+	if g.now != nil {
+		now = g.now()
+	}
+	if g.last == nil {
+		g.last = make(map[string]time.Time)
+		g.inflight = make(map[string]chan struct{})
+	}
+	if _, running := g.inflight[tenantID]; running {
+		return false
+	}
+	if last, ok := g.last[tenantID]; ok && now.Sub(last) < revisionPruneMinInterval {
+		return false
+	}
+	g.last[tenantID] = now
+	g.inflight[tenantID] = make(chan struct{})
+	return true
+}
+
+func (g *revisionPruneGate) done(tenantID string) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	ch, ok := g.inflight[tenantID]
+	if !ok {
+		return
+	}
+	close(ch)
+	delete(g.inflight, tenantID)
+}
+
+// applyRevisionRetention schedules retention for tenantID and returns without
+// waiting. Prune uses its own short context, at most once per tenant per five
+// minutes, and overlapping renders do not start another transaction.
+func applyRevisionRetention(st store.Backend, tenantID string) {
+	tenantID = strings.TrimSpace(tenantID)
+	if st == nil || tenantID == "" {
+		return
+	}
+	if !revisionPrunes.start(tenantID) {
+		return
+	}
+	go func() {
+		defer revisionPrunes.done(tenantID)
+		ctx, cancel := context.WithTimeout(context.Background(), revisionPruneTimeout)
+		defer cancel()
+		pruneTenantRevisions(ctx, st, tenantID)
+	}()
+}
+
+// pruneTenantRevisions applies the tenant retention cutoff immediately.
+func pruneTenantRevisions(ctx context.Context, st store.Backend, tenantID string) {
+	if st == nil || ctx == nil {
+		return
+	}
+	if err := ctx.Err(); err != nil {
+		return
+	}
+	settings, err := st.ListGlobalSettings(tenantID)
+	if err != nil {
+		return
+	}
+	_, _ = st.PruneRevisionsBefore(ctx, tenantID, RevisionRetentionCutoff(settings))
+}
+
+func waitRevisionPrune(ctx context.Context, tenantID string) error {
+	revisionPrunes.mu.Lock()
+	ch := revisionPrunes.inflight[strings.TrimSpace(tenantID)]
+	revisionPrunes.mu.Unlock()
+	if ch == nil {
+		return nil
+	}
+	select {
+	case <-ch:
+		return nil
+	case <-ctx.Done():
+		return ctx.Err()
+	}
+}