server: DF-mode big_send + uploaded-run storage with an operator policy

big_send now forces the Don't-Fragment bit for the whole burst by default, so
the largest size that arrives IS the downstream path MTU rather than "fragments
got through" — two different measurements the schema already separates. Sizes
above our own egress MTU (from the startup self-test) are refused up front and
reported as max_df_bytes, because absence caused by our kernel must not be read
as a limit of the client's path.

Uploads: one JSON file per run under the state dir, with the policy the operator
actually cares about — who may upload (off / anonymous / account), how large,
how long to keep, and the least anonymization accepted. The profile advertises
all of it so the app can present the switch honestly instead of discovering the
rules by failing. `account` refuses today rather than falling back to anonymous:
picking the strict setting before OIDC lands must not silently mean the loose one.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

server/internal/dataplane/granted.go

@@ -52,10 +52,25 @@ func (s *Server) DownTrain(sess *session.Session, g *session.Grant, count, sizeB
 	return sent, nil
 }
 
-// BigSend transmits one datagram per requested size, largest-first metadata intact, so the client
-// can see which sizes survive the *downstream* path — the mtu.pmtud_down / mtu.blackhole evidence.
-// The client cannot produce this itself: only the far end can emit a large packet toward it.
-func (s *Server) BigSend(sess *session.Session, g *session.Grant, sizes []int) ([]int, error) {
+// BigSendResult records what happened to one requested size. `Sent` false with an EMSGSIZE-ish
+// Err means *we* could not put it on the wire (the datagram exceeds our own egress MTU with DF
+// set) — the client must not read its absence as a path limit, so this is reported, not hidden.
+type BigSendResult struct {
+	SizeBytes int    `json:"size_bytes"`
+	Seq       int    `json:"seq"`
+	Sent      bool   `json:"sent"`
+	Err       string `json:"err,omitempty"`
+}
+
+// BigSend transmits one datagram per requested size so the client can see which sizes survive the
+// *downstream* path — the mtu.pmtud_down / mtu.frag_delivery evidence. The client cannot produce
+// this itself: only the far end can emit a large packet toward it.
+//
+// With df set, the DF bit is forced for the whole burst, so nothing fragments and the largest
+// size that arrives IS the downstream path MTU. Without it, the kernel fragments freely and the
+// result only says whether fragments get through — a different (also useful) measurement, and
+// the reason the two are separate test types.
+func (s *Server) BigSend(sess *session.Session, g *session.Grant, sizes []int, df bool) ([]BigSendResult, error) {
 	target := sess.DataSource()
 	if !target.IsValid() {
 		return nil, fmt.Errorf("no observed data-plane source")
@@ -64,24 +79,46 @@ func (s *Server) BigSend(sess *session.Session, g *session.Grant, sizes []int) (
 	if conn == nil {
 		return nil, fmt.Errorf("no data-plane socket matches target family")
 	}
-	attempted := make([]int, 0, len(sizes))
-	for i, size := range sizes {
-		if size < HeaderSize+8 {
-			size = HeaderSize + 8
+
+	results := make([]BigSendResult, 0, len(sizes))
+	burst := func() error {
+		for i, size := range sizes {
+			if size < HeaderSize+8 {
+				size = HeaderSize + 8
+			}
+			if size > 9000 { // jumbo ceiling; beyond this the kernel will refuse anyway
+				size = 9000
+			}
+			if !g.Allow(size) {
+				break
+			}
+			payload := make([]byte, size-HeaderSize)
+			// Echo the intended size into the payload so a truncated/fragmented arrival is
+			// still attributable to the size we meant to send.
+			binary.BigEndian.PutUint32(payload[0:4], uint32(size))
+			err := s.sendErr(conn, target, sess, TypeBigSend, uint32(i), payload)
+			results = append(results, BigSendResult{
+				SizeBytes: size, Seq: i, Sent: err == nil, Err: errString(err),
+			})
+			time.Sleep(20 * time.Millisecond) // keep bursts from being read as congestion loss
 		}
-		if size > 9000 { // jumbo ceiling; beyond this the kernel will refuse anyway
-			size = 9000
-		}
-		if !g.Allow(size) {
-			break
-		}
-		payload := make([]byte, size-HeaderSize)
-		// Echo the intended size into the payload so a truncated/fragmented arrival is
-		// still attributable to the size we meant to send.
-		binary.BigEndian.PutUint32(payload[0:4], uint32(size))
-		s.send(conn, target, sess, TypeBigSend, uint32(i), payload)
-		attempted = append(attempted, size)
-		time.Sleep(20 * time.Millisecond) // keep bursts from being read as congestion loss
+		return nil
 	}
-	return attempted, nil
+
+	if df && dfSupported {
+		s.dfMu.Lock()
+		defer s.dfMu.Unlock()
+		if err := withDF(conn, burst); err != nil {
+			return results, err
+		}
+		return results, nil
+	}
+	return results, burst()
+}
+
+func errString(err error) string {
+	if err == nil {
+		return ""
+	}
+	return err.Error()
 }