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_async_test.go

@@ -0,0 +1,120 @@
+package pipeline
+
+import (
+	"context"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"evobgp/internal/store"
+)
+
+type blockingPrune struct {
+	store.Backend
+	calls   atomic.Int32
+	started chan struct{}
+	release chan struct{}
+	once    sync.Once
+}
+
+func (b *blockingPrune) PruneRevisionsBefore(ctx context.Context, tenantID string, cutoff time.Time) (int, error) {
+	n := b.calls.Add(1)
+	if n == 1 {
+		b.once.Do(func() { close(b.started) })
+		select {
+		case <-b.release:
+		case <-ctx.Done():
+			return 0, ctx.Err()
+		}
+	}
+	return b.Backend.PruneRevisionsBefore(ctx, tenantID, cutoff)
+}
+
+func TestRenderDoesNotWaitForRevisionPrune(t *testing.T) {
+	m := store.NewMemory()
+	m.SeedDemo()
+	tenant, _, modIP, _, _ := m.DemoIDs()
+	release := make(chan struct{})
+	var releaseOnce sync.Once
+	closeRelease := func() { releaseOnce.Do(func() { close(release) }) }
+	t.Cleanup(closeRelease)
+	st := &blockingPrune{Backend: m, started: make(chan struct{}), release: release}
+	var shift atomic.Int64
+	revisionPrunes.mu.Lock()
+	prevNow := revisionPrunes.now
+	revisionPrunes.now = func() time.Time {
+		return time.Now().Add(time.Duration(shift.Load()) * time.Minute)
+	}
+	revisionPrunes.mu.Unlock()
+	t.Cleanup(func() {
+		revisionPrunes.mu.Lock()
+		revisionPrunes.now = prevNow
+		revisionPrunes.mu.Unlock()
+	})
+
+	rows := []store.PrefixRow{{Prefix: "203.0.113.55/32", Source: "ip_range"}}
+	renderDone := make(chan error, 1)
+	go func() {
+		_, err := RenderTenantRevisionFromPrefixes(context.Background(), st, nil, tenant, modIP, rows)
+		renderDone <- err
+	}()
+	select {
+	case err := <-renderDone:
+		if err != nil {
+			t.Fatal(err)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("render waited for revision prune")
+	}
+	select {
+	case <-st.started:
+	case <-time.After(5 * time.Second):
+		t.Fatal("prune did not start")
+	}
+	if st.calls.Load() != 1 {
+		t.Fatalf("prune calls=%d", st.calls.Load())
+	}
+	applyRevisionRetention(st, tenant)
+	if st.calls.Load() != 1 {
+		t.Fatalf("overlapping render stacked prune calls=%d", st.calls.Load())
+	}
+	closeRelease()
+	waitCtx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+	defer cancel()
+	if err := waitRevisionPrune(waitCtx, tenant); err != nil {
+		t.Fatal(err)
+	}
+	shift.Store(10)
+	applyRevisionRetention(st, tenant)
+	waitCtx2, cancel2 := context.WithTimeout(context.Background(), 2*time.Second)
+	defer cancel2()
+	if err := waitRevisionPrune(waitCtx2, tenant); err != nil {
+		t.Fatal(err)
+	}
+	if st.calls.Load() != 2 {
+		t.Fatalf("prune after interval calls=%d", st.calls.Load())
+	}
+}
+
+func TestPruneTenantRevisionsDirect(t *testing.T) {
+	m := store.NewMemory()
+	m.SeedDemo()
+	tenant, _, _, _, _ := m.DemoIDs()
+	var calls atomic.Int32
+	st := pruneCountBackend{Backend: m, calls: &calls}
+	pruneTenantRevisions(context.Background(), st, tenant)
+	if calls.Load() != 1 {
+		t.Fatalf("calls=%d", calls.Load())
+	}
+}
+
+type pruneCountBackend struct {
+	store.Backend
+	calls *atomic.Int32
+}
+
+func (p pruneCountBackend) PruneRevisionsBefore(ctx context.Context, tenantID string, cutoff time.Time) (int, error) {
+	p.calls.Add(1)
+	return p.Backend.PruneRevisionsBefore(ctx, tenantID, cutoff)
+}