fix(upload): stage encrypted body in temp file for known Content-Length

MinIO (and stricter S3 endpoints) reject presigned PUTs sent with
chunked transfer-encoding, returning HTTP 411 'Length Required'. The
pipeline could not know the final encrypted size up-front so it
streamed the request body with ContentLength=-1.

Drain the encrypt stage into a temp file, then issue the PUT with an
explicit Content-Length. The dump → compress → encrypt goroutines
still overlap because the drain reads from the encrypt pipe; only the
upload itself is sequenced after encryption completes.

apps/agent/internal/pipeline/runner.go

@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"os"
 	"strings"
 	"time"
 
@@ -302,14 +303,30 @@ func (r *Runner) Run(ctx context.Context, req *backupv1.RunBackup) (completed *b
 		errs <- nil
 	}()
 
-	// Stage 4 — upload (blocking call on the calling goroutine). On
-	// failure we still need to wait for the three upstream goroutines
-	// to unwind so the function does not leak them; closing the pipe
-	// readers below makes their pending Write calls return promptly.
-	sha256hex, uploaded, uploadErr := r.uploader.Put(ctx, req.UploadCreds.PresignedPutUrl, encryptedPR, -1)
+	// Stage 4 — drain the encrypted pipe into a temp file, then PUT it
+	// with a known Content-Length. MinIO (and stricter S3 endpoints)
+	// reject chunked PUTs against presigned URLs with HTTP 411.
+	// Buffering on disk keeps memory flat while still allowing the
+	// dump → compress → encrypt goroutines to overlap with the drain.
+	stagedSize, stagedPath, stageErr := stageEncryptedBody(encryptedPR)
+	if stagedPath != "" {
+		defer func() { _ = os.Remove(stagedPath) }()
+	}
+	var sha256hex string
+	var uploaded int64
+	var uploadErr error
+	if stageErr != nil {
+		uploadErr = stageErr
+	} else {
+		stagedFile, openErr := os.Open(stagedPath)
+		if openErr != nil {
+			uploadErr = fmt.Errorf("open staged body: %w", openErr)
+		} else {
+			sha256hex, uploaded, uploadErr = r.uploader.Put(ctx, req.UploadCreds.PresignedPutUrl, stagedFile, stagedSize)
+			_ = stagedFile.Close()
+		}
+	}
 	if uploadErr != nil {
-		// Closing the readers signals every upstream Write to fail
-		// with io.ErrClosedPipe so the producer goroutines exit.
 		_ = encryptedPR.CloseWithError(uploadErr)
 		_ = compressedPR.CloseWithError(uploadErr)
 		_ = dumpPR.CloseWithError(uploadErr)
@@ -469,6 +486,29 @@ func (passthroughDEK) Unwrap(_ context.Context, in []byte) ([]byte, error) {
 	return out, nil
 }
 
+// stageEncryptedBody drains src into a fresh temp file and returns the
+// path + total size so the caller can issue a PUT with an explicit
+// Content-Length. MinIO and stricter S3 endpoints reject chunked
+// transfer-encoding against presigned URLs (HTTP 411). On error, the
+// caller is responsible for removing the (possibly partial) file at
+// the returned path.
+func stageEncryptedBody(src io.Reader) (int64, string, error) {
+	f, err := os.CreateTemp(os.TempDir(), "backupy-upload-*.bin")
+	if err != nil {
+		return 0, "", fmt.Errorf("stage upload: create temp: %w", err)
+	}
+	path := f.Name()
+	n, copyErr := io.Copy(f, src)
+	closeErr := f.Close()
+	if copyErr != nil {
+		return n, path, fmt.Errorf("stage upload: copy: %w", copyErr)
+	}
+	if closeErr != nil {
+		return n, path, fmt.Errorf("stage upload: close: %w", closeErr)
+	}
+	return n, path, nil
+}
+
 // wipe zeroes a byte slice. Best-effort — the Go runtime makes no
 // guarantee that the underlying memory pages aren't already swapped
 // out, but this still raises the bar for casual memory inspection.