server: upstream trains, observed TTL/DSCP/ECN, rate limits, action ids

Types 0x03/0x04/0x05 land with a bounded columnar train buffer (head kept,
truncation declared) and grant-free multi-part reports - a report row is
smaller than the packet it answers, so $3.4 holds without a grant. The
read loop now collects TTL/TOS cmsgs on Linux, replacing the 0xFF stubs in
the observation block with what the kernel saw; downtrain gained a dscp
parameter, so DSCP survival is measurable in both directions.

Rate limiting ($2.5) exists now: per-credential AND per-source buckets,
429 on the control plane, silent drop on the data plane after the HMAC
gate and before the replay window. UDP ceilings default above the largest
legitimate run - a limit that clips a real measurement produces a
confidently wrong number.

Every granted packet carries its action_id at payload[8:16]; overlapping
actions were unattributable before. Canary DNS logs now honor the stated
24h privacy default. /admin/enroll-tokens answers the spec's JSON shape.
protocol_version 1.0.1 (additive).

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

server/internal/dataplane/frag_linux.go

@@ -104,8 +104,8 @@ func (s *Server) FragSend(
 		return res, fmt.Errorf("session has no recorded local address")
 	}
 
-	if sizeBytes < HeaderSize+8 {
-		sizeBytes = HeaderSize + 8
+	if sizeBytes < HeaderSize+32 {
+		sizeBytes = HeaderSize + 32
 	}
 	if sizeBytes > 8000 {
 		sizeBytes = 8000
@@ -117,7 +117,10 @@ func (s *Server) FragSend(
 	// The ELT1 packet, signed exactly as any other, then wrapped in UDP.
 	payload := make([]byte, sizeBytes-HeaderSize)
 	binary.BigEndian.PutUint32(payload[0:4], uint32(sizeBytes))
-	copy(payload[4:], mode)
+	// [8:16]: the action id, as on every granted packet (spec §5 correlation); the mode string
+	// sits past the reserved slot.
+	putActionID(payload, g.ActionID)
+	copy(payload[16:], mode)
 	elt := s.buildPacket(sess, TypeFragData, 0, payload)
 
 	udp := buildUDP(local, target, elt)