diff --git a/pkg/dind/bindstage.go b/pkg/dind/bindstage.go
index df089ad..116dbef 100644
--- a/pkg/dind/bindstage.go
+++ b/pkg/dind/bindstage.go
@@ -75,3 +75,33 @@ func SweepStagedBinds(dataDir string, log *slog.Logger) {
 	}
 	sweepStagedBinds(stagingRootDir(dataDir), log)
 }
+
+// SweepStagedBindsForJob removes the bind staging mounts and directory of ONE
+// job. Safe to call while other jobs are running, which is what separates it
+// from SweepStagedBinds.
+//
+// This exists because #187's periodic dead-container reaper called the
+// whole-tree SweepStagedBinds on its sweeper tick. That function is documented
+// STARTUP ONLY for a precise reason, and the reaper violated it: sweeping the
+// root unmounts and deletes the staging directory of every LIVE job too. The
+// running job's already-started containers keep working (their mounts are
+// already in their namespaces), so nothing fails loudly — but the job's NEXT
+// `docker run -v` cannot create its mountpoint, because ensureDirLocked has
+// already latched m.ready and will not recreate the parent:
+//
+//	docker: Error response from daemon: bind mount ... rejected:
+//	  creating bind staging mountpoint <data>/dind-binds/<job>/2:
+//	  mkdir ...: no such file or directory
+//
+// Observed on the fleet's Linux amd64 nodes against ephpm's release workflow,
+// which runs several `docker run -v "$PWD":/w` steps per job — binds 0 and 1
+// stage fine, a reaper tick lands, and bind 2 fails. Scoping the sweep to the
+// container actually being reaped keeps the benefit (releasing the mounts that
+// pin that container's rootfs snapshot, which is why the reaper swept at all)
+// without touching anyone else's.
+func SweepStagedBindsForJob(dataDir, jobID string, log *slog.Logger) {
+	if dataDir == "" || jobID == "" {
+		return
+	}
+	sweepStagedBindsForJob(stagingRootDir(dataDir), jobID, log)
+}
diff --git a/pkg/dind/bindstage_linux.go b/pkg/dind/bindstage_linux.go
index 81475f4..c249606 100644
--- a/pkg/dind/bindstage_linux.go
+++ b/pkg/dind/bindstage_linux.go
@@ -232,6 +232,25 @@ func stagingRootParent(jobDir string) string { return filepath.Dir(jobDir) }
 // sweepStagedBinds is the startup half of the lifecycle: it unmounts and
 // removes staging directories left by a previous ephemerd process. See
 // SweepStagedBinds for why leaked staging mounts are worse than untidy.
+// sweepStagedBindsForJob unmounts and removes exactly one job's staging
+// directory. Unlike sweepStagedBinds it never reads the root, so a live job's
+// directory sitting next to this one is untouched.
+func sweepStagedBindsForJob(root, jobID string, log *slog.Logger) {
+	dir := filepath.Join(root, jobID)
+	if _, err := os.Lstat(dir); err != nil {
+		return // this job never staged a bind, or it is already gone
+	}
+	if err := unmountTreeAndRemove(dir); err != nil {
+		if log != nil {
+			log.Warn("failed to sweep dind bind staging dir for reaped job", "path", dir, "error", err)
+		}
+		return
+	}
+	if log != nil {
+		log.Info("swept dind bind staging dir for reaped job", "path", dir)
+	}
+}
+
 func sweepStagedBinds(root string, log *slog.Logger) {
 	entries, err := os.ReadDir(root)
 	if err != nil {
diff --git a/pkg/dind/bindstage_linux_test.go b/pkg/dind/bindstage_linux_test.go
index 7657b37..83dcbe3 100644
--- a/pkg/dind/bindstage_linux_test.go
+++ b/pkg/dind/bindstage_linux_test.go
@@ -726,3 +726,50 @@ func TestBindStaging_MountpointTypeMatchesSource(t *testing.T) {
 		t.Errorf("staged file reads %q (err %v), want %q", body, err, "body")
 	}
 }
+
+// TestSweepStagedBindsForJob_SparesOtherJobs pins the fix for the regression
+// #187 introduced: its periodic reaper called the whole-tree
+// SweepStagedBinds, which is documented STARTUP ONLY because it cannot tell a
+// live job's staging dir from a dead one. On the fleet that deleted running
+// jobs' directories, and their next `docker run -v` failed with
+// "creating bind staging mountpoint <data>/dind-binds/<job>/2:
+//
+//	mkdir ...: no such file or directory"
+//
+// because ensureDirLocked had already latched m.ready and would not recreate
+// the parent. No mounts are needed to prove the scoping: plain directories
+// exercise the same path selection.
+func TestSweepStagedBindsForJob_SparesOtherJobs(t *testing.T) {
+	data := t.TempDir()
+	root := stagingRootDir(data)
+
+	dead := filepath.Join(root, "job-dead")
+	live := filepath.Join(root, "job-live")
+	for _, d := range []string{filepath.Join(dead, "0"), filepath.Join(live, "0")} {
+		if err := os.MkdirAll(d, 0o700); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	SweepStagedBindsForJob(data, "job-dead", discardLog())
+
+	if _, err := os.Lstat(dead); !os.IsNotExist(err) {
+		t.Errorf("reaped job's staging dir still present (err=%v), want removed", err)
+	}
+	// The whole point: a live job sitting next to the reaped one must survive,
+	// mountpoint subdirectory and all, or its next docker run cannot stage.
+	if _, err := os.Lstat(filepath.Join(live, "0")); err != nil {
+		t.Errorf("LIVE job's staging dir was destroyed by a scoped sweep: %v — "+
+			"this is the #187 regression: the next `docker run -v` in that job "+
+			"fails with mkdir ENOENT", err)
+	}
+}
+
+// TestSweepStagedBindsForJob_ToleratesMissing: a job that never staged a bind
+// (no -v in the whole job) must not make the reaper log or fail.
+func TestSweepStagedBindsForJob_ToleratesMissing(t *testing.T) {
+	data := t.TempDir()
+	SweepStagedBindsForJob(data, "never-staged", discardLog())
+	SweepStagedBindsForJob("", "job", discardLog())
+	SweepStagedBindsForJob(data, "", discardLog())
+}
diff --git a/pkg/dind/bindstage_other.go b/pkg/dind/bindstage_other.go
index 7895c67..dc05b25 100644
--- a/pkg/dind/bindstage_other.go
+++ b/pkg/dind/bindstage_other.go
@@ -5,6 +5,7 @@ package dind
 import (
 	"log/slog"
 	"os"
+	"path/filepath"
 )
 
 // newBindStager returns the off-Linux stager. There is no mount(2) here and,
@@ -35,3 +36,9 @@ func (passthroughStager) teardown() {}
 func sweepStagedBinds(root string, log *slog.Logger) {
 	_ = os.RemoveAll(root)
 }
+
+// sweepStagedBindsForJob is the one-job form; same "nothing to unmount"
+// reasoning as above.
+func sweepStagedBindsForJob(root, jobID string, log *slog.Logger) {
+	_ = os.RemoveAll(filepath.Join(root, jobID))
+}
diff --git a/pkg/runtime/runtime.go b/pkg/runtime/runtime.go
index a860683..65d25c4 100644
--- a/pkg/runtime/runtime.go
+++ b/pkg/runtime/runtime.go
@@ -563,12 +563,11 @@ func (r *Runtime) ReapDeadContainers(ctx context.Context) error {
 	}
 
 	now := time.Now()
-	// Release leaked dind bind-staging mounts lazily — only once, and only
-	// if we actually find something to reap. Each leaked staging mount pins
-	// the runner rootfs it was bound from, which makes WithSnapshotCleanup
-	// fail with "device or resource busy"; CleanOrphans sweeps them for the
-	// same reason. Skipped entirely on an idle node with nothing to reap.
-	stagedSwept := false
+	// Each leaked staging mount pins the runner rootfs it was bound from,
+	// which makes WithSnapshotCleanup fail with "device or resource busy",
+	// so a container being reaped needs its staging dir released first. That
+	// happens per-container below, NOT once per pass over the whole tree —
+	// see the comment at the call site.
 	for _, c := range containers {
 		id := c.ID()
 		log := r.cfg.Log.With("id", id)
@@ -592,9 +591,18 @@ func (r *Runtime) ReapDeadContainers(ctx context.Context) error {
 			continue
 		}
 
-		if !stagedSwept {
-			dind.SweepStagedBinds(r.cfg.DataDir, r.cfg.Log)
-			stagedSwept = true
+		// Sweep ONLY this container's staging dir. The whole-tree
+		// SweepStagedBinds used to run here once per pass, which broke live
+		// jobs: it is documented STARTUP ONLY precisely because it cannot
+		// tell a live job's staging dir from a dead one, and deleting a
+		// running job's dir makes its NEXT `docker run -v` fail with
+		// "creating bind staging mountpoint ...: no such file or directory"
+		// (ensureDirLocked has latched m.ready and will not recreate the
+		// parent). The reaping motive is unchanged — a leaked staged mount
+		// pins this container's rootfs so its snapshot cannot be deleted —
+		// and scoping to the container being reaped keeps that.
+		{
+			dind.SweepStagedBindsForJob(r.cfg.DataDir, id, r.cfg.Log)
 		}
 
 		log.Info("reaping dead job container", "reason", "task dead past grace, snapshot pinned", "grace", grace)