fix: use per-repo mutex to serialize snapshot generation

Replace scheduler-based serialization with a per-repo sync.Mutex.
This prevents concurrent snapshot runs for the same repo without
blocking on the scheduler's global worker pool, so on-demand HTTP
requests are not delayed by unrelated repos' background jobs.

Author: nssherpa

internal/strategy/git/git.go

@@ -51,6 +51,7 @@ type Strategy struct {
 	spoolsMu     sync.Mutex
 	spools       map[string]*RepoSpools
 	tokenManager *githubapp.TokenManager
+	snapshotMu   sync.Map // keyed by upstream URL, values are *sync.Mutex
 }
 
 func New(

internal/strategy/git/snapshot.go

@@ -9,6 +9,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/alecthomas/errors"
@@ -47,6 +48,10 @@ func (s *Strategy) generateAndUploadSnapshot(ctx context.Context, repo *gitclone
 
 	logger.InfoContext(ctx, "Snapshot generation started", slog.String("upstream", upstream))
 
+	mu := s.snapshotMutexFor(upstream)
+	mu.Lock()
+	defer mu.Unlock()
+
 	mirrorRoot := s.cloneManager.Config().MirrorRoot
 	snapshotDir, err := snapshotDirForURL(mirrorRoot, upstream)
 	if err != nil {
@@ -106,6 +111,11 @@ func (s *Strategy) scheduleSnapshotJobs(repo *gitclone.Repository) {
 	})
 }
 
+func (s *Strategy) snapshotMutexFor(upstreamURL string) *sync.Mutex {
+	mu, _ := s.snapshotMu.LoadOrStore(upstreamURL, &sync.Mutex{})
+	return mu.(*sync.Mutex)
+}
+
 func (s *Strategy) handleSnapshotRequest(w http.ResponseWriter, r *http.Request, host, pathValue string) {
 	ctx := r.Context()
 	logger := logging.FromContext(ctx)