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/store/memory_job_queue_test.go

@@ -0,0 +1,122 @@
+package store
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestEnqueueJobTenantsDoNotShareLock(t *testing.T) {
+	m := NewMemory()
+	ctx := context.Background()
+	unlock := m.lockJobEnqueueTenant("tenant-a")
+	var once sync.Once
+	release := func() { once.Do(unlock) }
+	t.Cleanup(release)
+
+	otherDone := make(chan error, 1)
+	go func() {
+		_, _, err := m.EnqueueJob(ctx, JobEnqueue{TenantID: "tenant-b", Kind: "test"})
+		otherDone <- err
+	}()
+	select {
+	case err := <-otherDone:
+		if err != nil {
+			t.Fatal(err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("other tenant waited on the held tenant lock")
+	}
+
+	sameDone := make(chan struct{})
+	go func() {
+		_, _, err := m.EnqueueJob(ctx, JobEnqueue{TenantID: "tenant-a", Kind: "test"})
+		if err != nil {
+			t.Error(err)
+		}
+		close(sameDone)
+	}()
+	select {
+	case <-sameDone:
+		t.Fatal("same tenant did not wait on its lock")
+	case <-time.After(200 * time.Millisecond):
+	}
+	release()
+	select {
+	case <-sameDone:
+	case <-time.After(2 * time.Second):
+		t.Fatal("same tenant did not finish after unlock")
+	}
+}
+
+func TestEnqueueJobIdempotencyKeyConflicts(t *testing.T) {
+	m := NewMemory()
+	ctx := context.Background()
+	key := "same-key"
+	in := JobEnqueue{TenantID: "tenant-a", Kind: "refresh", IdempotencyKey: &key, Meta: map[string]any{"module": "a"}}
+	first, created, err := m.EnqueueJob(ctx, in)
+	if err != nil || !created || first.ID == "" {
+		t.Fatalf("first=%+v created=%v err=%v", first, created, err)
+	}
+	in.Meta = map[string]any{"module": "b"}
+	if _, _, err = m.EnqueueJob(ctx, in); !errors.Is(err, ErrJobConflict) {
+		t.Fatalf("same key accepted: %v", err)
+	}
+	other := JobEnqueue{TenantID: "tenant-b", Kind: "refresh", IdempotencyKey: &key, Meta: map[string]any{"module": "b"}}
+	if _, created, err = m.EnqueueJob(ctx, other); err != nil || !created {
+		t.Fatalf("other tenant created=%v err=%v", created, err)
+	}
+}
+
+func TestMemoryJobQueueEqualTimestampOrdering(t *testing.T) {
+	m := NewMemory()
+	ctx := context.Background()
+	stamp := time.Date(2026, 10, 26, 0, 0, 0, 0, time.UTC)
+	var ids []string
+	for range 32 {
+		j, _, err := m.EnqueueJob(ctx, JobEnqueue{TenantID: "tenant", Kind: "test"})
+		if err != nil {
+			t.Fatal(err)
+		}
+		ids = append(ids, j.ID)
+		m.jobMu.Lock()
+		saved := m.jobs[j.ID]
+		saved.CreatedAt = stamp
+		m.jobs[j.ID] = saved
+		m.jobMu.Unlock()
+	}
+	// Repeated offset pages must use the same total order despite map iteration.
+	for range 5 {
+		cursor := ""
+		var listed []string
+		for {
+			page, next, more, err := m.ListJobs(ctx, "tenant", "", "", cursor, 3)
+			if err != nil {
+				t.Fatal(err)
+			}
+			for _, j := range page {
+				listed = append(listed, j.ID)
+			}
+			if !more {
+				break
+			}
+			cursor = next
+		}
+		if len(listed) != len(ids) {
+			t.Fatalf("listed %d jobs", len(listed))
+		}
+		for i, id := range listed {
+			if id != ids[len(ids)-1-i] {
+				t.Fatalf("list[%d]=%s expected reverse enqueue sequence %s", i, id, ids[len(ids)-1-i])
+			}
+		}
+	}
+	for i, want := range ids {
+		claimed, err := m.ClaimJobs(ctx, "owner", 1, time.Minute)
+		if err != nil || len(claimed) != 1 || claimed[0].ID != want {
+			t.Fatalf("claim[%d]=%+v %v want %s", i, claimed, err, want)
+		}
+	}
+}