diff --git a/internal/tigerfs/fuse/adapter.go b/internal/tigerfs/fuse/adapter.go
index ca7e5df..a891954 100644
--- a/internal/tigerfs/fuse/adapter.go
+++ b/internal/tigerfs/fuse/adapter.go
@@ -224,9 +224,13 @@ func (a *FSAdapter) EntriesToDirEntries(entries []tigerfs.Entry) []gofuse.DirEnt
 // which fails in-flight queries. We only lose the per-request
 // cancellation channel, which the bug shows we don't actually want.
 //
-// Write paths intentionally do NOT use this -- partial writes have
-// real consistency implications when interrupted, and the request-ctx
-// cancellation surface is the right tool there.
+// Applies to both read and write paths. Each tigerfs write op runs as a
+// single auto-committed statement (or a single transaction wrapping a
+// few atomic statements -- e.g., DeleteAndUpdate for POSIX rename-as-
+// replace). Either the whole statement/transaction commits or none of
+// it commits; there is no partial state to leak when we ignore the
+// kernel's transient cancellation. The kernel's FUSE_INTERRUPT means
+// "you may stop early if you want," not "abort, the syscall is dying."
 func decoupleFromRequestCancel(ctx context.Context) context.Context {
 	return context.WithoutCancel(ctx)
 }
@@ -305,6 +309,7 @@ func (a *FSAdapter) ReadFile(ctx context.Context, path string) ([]byte, syscall.
 //
 // Returns errno (0 on success).
 func (a *FSAdapter) WriteFile(ctx context.Context, path string, data []byte) syscall.Errno {
+	ctx = decoupleFromRequestCancel(ctx)
 	fsErr := a.ops.WriteFile(ctx, path, data)
 	return a.ErrorToErrno(fsErr)
 }
@@ -319,6 +324,7 @@ func (a *FSAdapter) WriteFile(ctx context.Context, path string, data []byte) sys
 //
 // Returns errno (0 on success).
 func (a *FSAdapter) Delete(ctx context.Context, path string) syscall.Errno {
+	ctx = decoupleFromRequestCancel(ctx)
 	fsErr := a.ops.Delete(ctx, path)
 	return a.ErrorToErrno(fsErr)
 }
@@ -334,6 +340,7 @@ func (a *FSAdapter) Delete(ctx context.Context, path string) syscall.Errno {
 //
 // Returns errno (0 on success).
 func (a *FSAdapter) Mkdir(ctx context.Context, path string) syscall.Errno {
+	ctx = decoupleFromRequestCancel(ctx)
 	fsErr := a.ops.Mkdir(ctx, path)
 	return a.ErrorToErrno(fsErr)
 }
@@ -350,6 +357,7 @@ func (a *FSAdapter) Mkdir(ctx context.Context, path string) syscall.Errno {
 //
 // Returns errno (0 on success).
 func (a *FSAdapter) Rename(ctx context.Context, oldPath, newPath string) syscall.Errno {
+	ctx = decoupleFromRequestCancel(ctx)
 	fsErr := a.ops.Rename(ctx, oldPath, newPath)
 	return a.ErrorToErrno(fsErr)
 }
diff --git a/internal/tigerfs/fuse/adapter_ctx_test.go b/internal/tigerfs/fuse/adapter_ctx_test.go
index 2b4141d..8a7c151 100644
--- a/internal/tigerfs/fuse/adapter_ctx_test.go
+++ b/internal/tigerfs/fuse/adapter_ctx_test.go
@@ -88,6 +88,87 @@ func TestFSAdapter_ReadFile_DecouplesRequestCancel(t *testing.T) {
 	}
 }
 
+// Write-path coverage. Symmetric to the read-path tests: each FSAdapter
+// write method must also decouple the request ctx so FUSE_INTERRUPT
+// (typically from Go SIGURG goroutine preemption) doesn't surface as
+// close-time EIO / rename ENOENT / delete ENOENT on otherwise valid ops.
+
+func TestFSAdapter_WriteFile_DecouplesRequestCancel(t *testing.T) {
+	capturedCtx := captureCtxViaWriteFile(t, "/public/anything")
+	if capturedCtx == nil {
+		t.Fatal("Operations was not reached; FSAdapter short-circuited (regression)")
+	}
+	if capturedCtx.Err() != nil {
+		t.Errorf("FSAdapter forwarded cancelled ctx to DB layer: %v", capturedCtx.Err())
+	}
+}
+
+func TestFSAdapter_Delete_DecouplesRequestCancel(t *testing.T) {
+	capturedCtx := captureCtxViaDelete(t, "/public/anything")
+	if capturedCtx == nil {
+		t.Fatal("Operations was not reached; FSAdapter short-circuited (regression)")
+	}
+	if capturedCtx.Err() != nil {
+		t.Errorf("FSAdapter forwarded cancelled ctx to DB layer: %v", capturedCtx.Err())
+	}
+}
+
+func TestFSAdapter_Mkdir_DecouplesRequestCancel(t *testing.T) {
+	capturedCtx := captureCtxViaMkdir(t, "/public/anything")
+	if capturedCtx == nil {
+		t.Fatal("Operations was not reached; FSAdapter short-circuited (regression)")
+	}
+	if capturedCtx.Err() != nil {
+		t.Errorf("FSAdapter forwarded cancelled ctx to DB layer: %v", capturedCtx.Err())
+	}
+}
+
+func TestFSAdapter_Rename_DecouplesRequestCancel(t *testing.T) {
+	capturedCtx := captureCtxViaRename(t, "/public/old", "/public/new")
+	if capturedCtx == nil {
+		t.Fatal("Operations was not reached; FSAdapter short-circuited (regression)")
+	}
+	if capturedCtx.Err() != nil {
+		t.Errorf("FSAdapter forwarded cancelled ctx to DB layer: %v", capturedCtx.Err())
+	}
+}
+
+func captureCtxViaWriteFile(t *testing.T, path string) context.Context {
+	t.Helper()
+	adapter, captured := newAdapterWithCtxCapture()
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	_ = adapter.WriteFile(ctx, path, []byte("data"))
+	return *captured
+}
+
+func captureCtxViaDelete(t *testing.T, path string) context.Context {
+	t.Helper()
+	adapter, captured := newAdapterWithCtxCapture()
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	_ = adapter.Delete(ctx, path)
+	return *captured
+}
+
+func captureCtxViaMkdir(t *testing.T, path string) context.Context {
+	t.Helper()
+	adapter, captured := newAdapterWithCtxCapture()
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	_ = adapter.Mkdir(ctx, path)
+	return *captured
+}
+
+func captureCtxViaRename(t *testing.T, oldPath, newPath string) context.Context {
+	t.Helper()
+	adapter, captured := newAdapterWithCtxCapture()
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	_ = adapter.Rename(ctx, oldPath, newPath)
+	return *captured
+}
+
 // captureCtxViaReadDir constructs an FSAdapter wired to a MockDBClient
 // that records the ctx given to GetCurrentSchema (the earliest DB call
 // Operations.ReadDir makes), invokes ReadDir with an already-cancelled