fix(xnet): allocate ephemeral ports sequentially, not randomly (#180)

Ephemeral ports were drawn at random from the 16384-port dynamic range on
every dial. Under connection-per-request churn the birthday paradox reuses
a recently-released port after only a few dozen dials (~1% per dial at 200
outstanding-in-teardown), and a reused 4-tuple lands on whatever state the
previous conversation left along the path, such as the peer's TIME-WAIT
socket or a NAT's flow-table entry, which silently swallows the new SYN.
Measured end to end: an HTTP client with keep-alives off against a macOS
peer hit a dead dial after ~286 connections and stayed dark for ~30
seconds, exactly one 2MSL TIME-WAIT expiry.

Allocate sequentially from a per-stack random starting offset instead: a
port is only revisited after the full 16384-port cycle, and the random
start keeps a rebooted node off the ports its previous life just used.

TestEphemeralPortSequence pins the full-cycle-no-reuse property.

Co-authored-by: Derek den Haas <d.haas@directcode.com>

x/xnet/stack-async.go

@@ -53,6 +53,10 @@ type StackAsync struct {
 	lookup  dns.Message
 	dnssv   netip.Addr
 
+	// ephPort drives sequential ephemeral-port allocation (see
+	// [StackAsync.ephemeralPort]); zero means not yet seeded.
+	ephPort uint32
+
 	ntpUDP internet.StackUDPPort
 	ntp    ntp.Client
 
@@ -352,6 +356,23 @@ func (s *StackAsync) Prand32() (randval uint32) {
 	return randval
 }
 
+// ephemeralPort returns the next port of the IANA dynamic range (49152-65535,
+// RFC 6335 §6), allocated sequentially from a random per-stack start so a port is
+// revisited only after the full 16384-port cycle. Random selection instead reuses
+// a recent port at birthday-paradox rates, and a reused 4-tuple can collide with
+// state the previous conversation left behind (a TIME-WAIT, a NAT flow entry)
+// which swallows the new SYN.
+func (s *StackAsync) ephemeralPort() uint16 {
+	s.mu.Lock()
+	if s.ephPort == 0 {
+		s.ephPort = s.prand32()%16384 | 1
+	}
+	port := 49152 + s.ephPort%16384
+	s.ephPort++
+	s.mu.Unlock()
+	return uint16(port)
+}
+
 func (s *StackAsync) prand32() uint32 {
 	/* Algorithm "xor" from p. 4 of Marsaglia, "Xorshift RNGs" */
 	seed := internal.Prand32(s.prng)