server: fix egress-MTU probe — connect the socket before reading IP_MTU

IP_MTU getsockopt returns ENOTCONN on an unconnected socket; the v0.3.4
probe set IP_MTU_DISCOVER and Sendto but never Connect'd, so every probe
errored. UDP-connect (no handshake) pins the route so IP_MTU reflects the
path; switched to Write (two return values). Sysctl audit already flagged
the four real fmr issues in v0.3.4; this makes the MTU proof report.

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

server/README.md

@@ -53,6 +53,26 @@ All configurable via `ECHOLOT_*_LISTEN`. Plus:
 Deliberately out of scope here: an echo listener on 443 (to detect port-based egress filtering)
 — that genuinely needs 443 and belongs on a dedicated IP, not on a host running a reverse proxy.
 
+## Host tuning (measurement fidelity)
+
+A measurement server must not let the kernel distort what clients observe. Apply the
+recommended sysctls and the daemon will confirm the host is clean:
+
+```sh
+sudo cp deploy/99-echolot-sysctl.conf /etc/sysctl.d/ && sudo sysctl --system
+```
+
+The daemon **self-tests at startup and via `GET /admin/selftest`** (localhost):
+- **sysctl audit** — flags settings that would distort results (RA acceptance on a static host,
+  ICMP redirects, ICMP rate-limiting of the server's own errors, disabled TCP options).
+- **egress-MTU self-proof** — DF-probes external anchors (`ECHOLOT_MTU_PROBE_TARGETS`,
+  default 1.1.1.1 + a v6 anchor) and reads the discovered path MTU. If the server's *own* uplink
+  can't carry 1500, client MTU results would measure this server, not the client — so the profile
+  exposes `server_selftest.mtu_ok` and the log warns loudly.
+
+Both signals ride in `GET /v1/profile` as `server_selftest` so a client can trust — or skip —
+MTU testing accordingly.
+
 ## Run in Docker (config via env)
 
 ```sh

server/cmd/echolot-server/main.go

@@ -18,6 +18,7 @@ import (
 	"crypto/tls"
 	"crypto/x509"
 	"crypto/x509/pkix"
+	"encoding/json"
 	"encoding/pem"
 	"errors"
 	"fmt"
@@ -30,6 +31,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"strconv"
+	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -37,6 +39,7 @@ import (
 	"echo-lot.app/server/internal/config"
 	"echo-lot.app/server/internal/control"
 	"echo-lot.app/server/internal/dataplane"
+	"echo-lot.app/server/internal/selftest"
 	"echo-lot.app/server/internal/selfupdate"
 	"echo-lot.app/server/internal/session"
 	"echo-lot.app/server/internal/store"
@@ -138,11 +141,42 @@ func serve(cfg *config.Config) error {
 		}(addr, ln)
 	}
 
+	// Self-test: prove the host is a clean measurement target. Sysctl audit is
+	// instant; the egress-MTU proof does network round trips, so publish the
+	// sysctl-only report immediately and swap in the full one when it lands.
+	var selftestPtr atomic.Pointer[selftest.Report]
+	initial := selftest.Report{Sysctls: selftest.Sysctls()}
+	selftestPtr.Store(&initial)
+	for _, c := range initial.Sysctls {
+		if c.Severity == selftest.Warn {
+			slog.Warn("sysctl not measurement-clean", "sysctl", c.Name, "got", c.Got, "want", c.Want, "why", c.Why)
+		}
+	}
+	go func() {
+		r := selftest.Run(config.Addrs(cfg.MTUProbeTargets))
+		selftestPtr.Store(&r)
+		for _, m := range r.EgressMTU {
+			if !m.FullMTU {
+				slog.Warn("egress MTU below 1500 — client MTU results measure THIS server, not the client",
+					"target", m.Target, "discovered_mtu", m.DiscoveredMTU, "err", m.Err)
+			}
+		}
+		slog.Info("self-test complete", "sysctl_ok", r.SysctlOK, "mtu_ok", r.MTUOK)
+	}()
+	ctl.ProvenGood = func() (mtuOK, sysctlOK bool) {
+		r := selftestPtr.Load()
+		return r.MTUOK, r.SysctlOK
+	}
+
 	// Admin/health (plain HTTP, localhost by default; spec §7)
 	admin := http.NewServeMux()
 	admin.HandleFunc("GET /healthz", func(w http.ResponseWriter, _ *http.Request) {
 		fmt.Fprintf(w, `{"ok":true,"version":%q}`, Version)
 	})
+	admin.HandleFunc("GET /admin/selftest", func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(selftestPtr.Load())
+	})
 	// TODO(spec §7): enrollment token management + device list. Until the
 	// admin UI exists, mint tokens with: echolot-admin (or curl on this
 	// listener once the endpoint lands).

server/deploy/99-echolot-sysctl.conf

@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: 2026 Echolot contributors
+# SPDX-License-Identifier: GPL-3.0-or-later
+#
+# Recommended sysctls for an Echolot probe-server host: keep the kernel from
+# silently altering what clients measure. Install with:
+#   sudo cp 99-echolot-sysctl.conf /etc/sysctl.d/
+#   sudo sysctl --system
+# The daemon audits these at startup and via GET /admin/selftest; anything not
+# set here shows up as a "not measurement-clean" warning.
+
+# Static-addressed host: never let a Router Advertisement mutate our routing.
+# (Echolot's whole job is detecting broken RAs — the server must be immune.)
+net.ipv6.conf.all.accept_ra = 0
+net.ipv6.conf.default.accept_ra = 0
+
+# Don't let ICMP redirects rewrite our routing mid-measurement, and don't
+# emit redirects (we're an endpoint, not a router).
+net.ipv4.conf.all.accept_redirects = 0
+net.ipv4.conf.default.accept_redirects = 0
+net.ipv6.conf.all.accept_redirects = 0
+net.ipv4.conf.all.send_redirects = 0
+net.ipv4.conf.default.send_redirects = 0
+
+# Don't throttle the server's own ICMP errors (dest-unreachable/frag-needed/
+# time-exceeded) — throttling produces false loss/black-hole readings when
+# clients probe toward this server.
+net.ipv4.icmp_ratelimit = 0
+
+# These are usually already correct; pinned so the server can honestly
+# negotiate/reflect them (a missing option in a client's evidence is then the
+# path's fault, not ours).
+net.ipv4.tcp_sack = 1
+net.ipv4.tcp_timestamps = 1
+net.ipv4.tcp_window_scaling = 1
+net.ipv4.ip_no_pmtu_disc = 0
+net.ipv4.icmp_echo_ignore_all = 0
+
+# Loose reverse-path filtering suits a multi-IP measurement host (strict mode
+# can drop alt-address / asymmetric replies used by STUN 5780).
+net.ipv4.conf.all.rp_filter = 2

server/internal/config/config.go

@@ -32,6 +32,8 @@ type Config struct {
 	// Optional cleartext HTTP-echo listener (spec §4 plaintext-path test).
 	// Default empty = off; it exposes only POST /v1/echo, no auth, no secrets.
 	HTTPEchoListen string // ECHOLOT_HTTP_ECHO_LISTEN / --http-echo-listen
+	// Comma-separated anchors for the egress-MTU self-proof (host or ip).
+	MTUProbeTargets string // ECHOLOT_MTU_PROBE_TARGETS / --mtu-probe-targets
 
 	// Admin UI / health listener (spec §7: localhost-only by default)
 	AdminListen string // ECHOLOT_ADMIN_LISTEN / --admin-listen
@@ -75,6 +77,7 @@ func Load(args []string) (*Config, *Actions, error) {
 	fs.StringVar(&c.DNSListen, "dns-listen", envOr("DNS_LISTEN", ""), "canary-DNS listen address(es) udp+tcp/53, comma-separated; empty disables (spec §6.1)")
 	fs.StringVar(&c.CanaryZone, "canary-zone", envOr("CANARY_ZONE", ""), "authoritative canary zone, e.g. c.echo-lot.app")
 	fs.StringVar(&c.HTTPEchoListen, "http-echo-listen", envOr("HTTP_ECHO_LISTEN", ""), "optional CLEARTEXT http-echo listen address(es); empty disables (spec §4)")
+	fs.StringVar(&c.MTUProbeTargets, "mtu-probe-targets", envOr("MTU_PROBE_TARGETS", "1.1.1.1,2606:4700:4700::1111"), "egress-MTU self-proof anchors, comma-separated")
 	fs.StringVar(&c.AdminListen, "admin-listen", envOr("ADMIN_LISTEN", "127.0.0.1:8444"), "admin/health listen address (keep localhost)")
 	fs.StringVar(&c.StateDir, "state-dir", envOr("STATE_DIR", defaultStateDir()), "state directory (device store, generated TLS)")
 	fs.StringVar(&c.Name, "name", envOr("NAME", "echolot"), "server profile name")

server/internal/control/control.go

@@ -53,6 +53,11 @@ type Server struct {
 	CanaryQueries func(sessionPrefix string) any
 	// CanaryZone is surfaced in the profile so the app knows what to query.
 	CanaryZone string
+	// ProvenGood reports the server's self-test signal (may be nil). Surfaced
+	// in the profile so a client can trust — or skip — MTU tests: if the
+	// server's own egress isn't full-MTU, client MTU results measure the
+	// server, not the client.
+	ProvenGood func() (mtuOK, sysctlOK bool)
 }
 
 func (s *Server) Handler() http.Handler {
@@ -78,6 +83,16 @@ func (s *Server) EchoHandler() http.Handler {
 	return mux
 }
 
+// selftestSignal is the compact "server proven good" object for the profile.
+// mtu_ok=false tells a client its MTU results would measure this server.
+func selftestSignal(f func() (bool, bool)) map[string]any {
+	if f == nil {
+		return map[string]any{"mtu_ok": nil, "sysctl_ok": nil}
+	}
+	mtuOK, sysctlOK := f()
+	return map[string]any{"mtu_ok": mtuOK, "sysctl_ok": sysctlOK}
+}
+
 // sessionAuth resolves {id} and requires the bearer to be the owning device.
 func (s *Server) sessionAuth(w http.ResponseWriter, r *http.Request) *session.Session {
 	dev := s.Store.DeviceByCredential(bearer(r))
@@ -267,6 +282,7 @@ func (s *Server) profile(w http.ResponseWriter, r *http.Request) {
 		"pins":            []string{"pin-sha256:" + s.PinB64},
 		"next_pins":       []string{},
 		"canary_zone":     s.CanaryZone,
+		"server_selftest": selftestSignal(s.ProvenGood),
 		"limits":          map[string]any{"max_kbps": 50000, "max_session_s": 900},
 	})
 }

server/internal/selftest/mtu_linux.go

@@ -0,0 +1,112 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+//go:build linux
+
+package selftest
+
+import (
+	"net"
+	"net/netip"
+	"syscall"
+	"time"
+)
+
+// Linux IP-level constants for PMTU discovery. Not all are exported by the
+// stdlib syscall package across versions, so they are pinned here (stable
+// kernel ABI) — same rationale as the prober's OsAbi.
+const (
+	ipMTUDiscover   = 10 // IP_MTU_DISCOVER
+	ipMTU           = 14 // IP_MTU
+	ipPMTUDiscDo    = 2  // IP_PMTUDISC_DO (set DF, honor PMTU)
+	ipv6MTUDiscover = 23 // IPV6_MTU_DISCOVER
+	ipv6MTU         = 24 // IPV6_MTU
+	ipv6PMTUDiscDo  = 2  // IPV6_PMTUDISC_DO
+)
+
+// probeEgressMTU sends a DF-flagged full-size UDP datagram toward target and
+// reads back the kernel's discovered path MTU. A reduction below 1500 means
+// the SERVER's own uplink can't carry full-size packets — so client MTU
+// results would measure the server, not the client. No root, no raw socket:
+// IP_MTU_DISCOVER + a getsockopt on IP_MTU, mirroring the prober's approach.
+func probeEgressMTU(target string) MTUResult {
+	res := MTUResult{Target: target}
+	addr, err := netip.ParseAddr(target)
+	if err != nil {
+		// allow "host" that resolves
+		ips, e := net.LookupIP(target)
+		if e != nil || len(ips) == 0 {
+			res.Err = "resolve: " + errStr(err)
+			return res
+		}
+		addr, _ = netip.AddrFromSlice(ips[0])
+	}
+	addr = addr.Unmap()
+
+	is4 := addr.Is4()
+	fam := syscall.AF_INET6
+	if is4 {
+		fam = syscall.AF_INET
+	}
+	fd, err := syscall.Socket(fam, syscall.SOCK_DGRAM, 0)
+	if err != nil {
+		res.Err = "socket: " + errStr(err)
+		return res
+	}
+	defer syscall.Close(fd)
+
+	if is4 {
+		_ = syscall.SetsockoptInt(fd, syscall.IPPROTO_IP, ipMTUDiscover, ipPMTUDiscDo)
+	} else {
+		_ = syscall.SetsockoptInt(fd, syscall.IPPROTO_IPV6, ipv6MTUDiscover, ipv6PMTUDiscDo)
+	}
+
+	// IP_MTU reflects the CONNECTED path's MTU, so the socket must be connected
+	// (an unconnected socket returns ENOTCONN). No handshake — UDP connect just
+	// pins the destination and resolves the route.
+	sa := sockaddr(addr, 33434)
+	if err := syscall.Connect(fd, sa); err != nil {
+		res.Err = "connect: " + errStr(err)
+		return res
+	}
+
+	// Full-size probe: 1500 total − IP/UDP headers (28 v4, 48 v6). A DF send
+	// larger than the local MTU fails immediately with EMSGSIZE; a path
+	// reduction updates IP_MTU after the ICMP frag-needed returns, so we send,
+	// briefly wait, and read the discovered MTU.
+	payload := 1472
+	if !is4 {
+		payload = 1452
+	}
+	probe := make([]byte, payload)
+	_, _ = syscall.Write(fd, probe)
+	time.Sleep(700 * time.Millisecond)
+	_, _ = syscall.Write(fd, probe) // second send observes any reduction
+
+	level, opt := syscall.IPPROTO_IP, ipMTU
+	if !is4 {
+		level, opt = syscall.IPPROTO_IPV6, ipv6MTU
+	}
+	mtu, err := syscall.GetsockoptInt(fd, level, opt)
+	if err != nil || mtu <= 0 {
+		res.Err = "getsockopt IP_MTU: " + errStr(err)
+		return res
+	}
+	res.DiscoveredMTU = mtu
+	res.FullMTU = mtu >= 1500
+	return res
+}
+
+func sockaddr(a netip.Addr, port int) syscall.Sockaddr {
+	if a.Is4() {
+		return &syscall.SockaddrInet4{Port: port, Addr: a.As4()}
+	}
+	return &syscall.SockaddrInet6{Port: port, Addr: a.As16()}
+}
+
+func errStr(err error) string {
+	if err == nil {
+		return "nil"
+	}
+	return err.Error()
+}

server/internal/selftest/mtu_other.go

@@ -0,0 +1,13 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+//go:build !linux
+
+package selftest
+
+// probeEgressMTU: PMTUD via IP_MTU_DISCOVER is Linux-specific. Off-Linux the
+// self-test reports MTU as unproven rather than guessing (the daemon runs on
+// Linux in production; this keeps dev builds compiling).
+func probeEgressMTU(target string) MTUResult {
+	return MTUResult{Target: target, Err: "egress MTU probe is Linux-only"}
+}

server/internal/selftest/selftest.go

@@ -0,0 +1,126 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+// Package selftest lets the daemon prove its own host is a clean measurement
+// target: the kernel isn't silently altering what clients measure, and the
+// server's own egress reaches full MTU. If the server side is already broken,
+// client-side results (especially MTU/PMTUD) measure the server, not the
+// client — so the daemon says so.
+package selftest
+
+import (
+	"os"
+	"strconv"
+	"strings"
+)
+
+// Severity of a check result.
+type Severity string
+
+const (
+	OK   Severity = "ok"
+	Warn Severity = "warn"
+)
+
+// Check is one sysctl (or derived) assertion.
+type Check struct {
+	Name     string   `json:"name"`
+	Got      string   `json:"got"`
+	Want     string   `json:"want"`
+	Severity Severity `json:"severity"`
+	Why      string   `json:"why"`
+}
+
+// MTUResult is one egress path-MTU probe outcome.
+type MTUResult struct {
+	Target       string `json:"target"`
+	DiscoveredMTU int   `json:"discovered_mtu"`
+	FullMTU      bool   `json:"full_mtu"` // >= 1500
+	Err          string `json:"err,omitempty"`
+}
+
+// Report is the whole self-test.
+type Report struct {
+	Sysctls   []Check     `json:"sysctls"`
+	EgressMTU []MTUResult `json:"egress_mtu"`
+	// SysctlOK / MTUOK are the compact "server proven good" signals; the
+	// profile surfaces these so a client can skip MTU tests the server can't
+	// support honestly.
+	SysctlOK bool `json:"sysctl_ok"`
+	MTUOK    bool `json:"mtu_ok"`
+}
+
+// readSysctl reads /proc/sys/<dotted.name>. Empty string if unavailable.
+func readSysctl(name string) string {
+	p := "/proc/sys/" + strings.ReplaceAll(name, ".", "/")
+	b, err := os.ReadFile(p)
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(b))
+}
+
+// sysctlChecks are the measurement-fidelity assertions. Each closure returns
+// OK/Warn given the read value; a missing value (non-Linux / restricted) is
+// reported as Warn "unreadable" but never fatal.
+var sysctlChecks = []struct {
+	name string
+	want string
+	why  string
+	ok   func(v string) bool
+}{
+	{"net.ipv6.conf.all.accept_ra", "0", "static v6 host must not let RAs mutate routing (the very thing Echolot detects)", eq("0")},
+	{"net.ipv4.conf.all.accept_redirects", "0", "ICMP redirects could alter routing mid-measurement", eq("0")},
+	{"net.ipv4.conf.all.send_redirects", "0", "an endpoint should not emit ICMP redirects", eq("0")},
+	{"net.ipv4.icmp_echo_ignore_all", "0", "server must answer ping so clients can measure to it", eq("0")},
+	{"net.ipv4.ip_no_pmtu_disc", "0", "server must honor path MTU on its own sends", eq("0")},
+	{"net.ipv4.tcp_sack", "1", "so a missing SACK in mss_observed is the path's fault, not the server's", eq("1")},
+	{"net.ipv4.tcp_timestamps", "1", "so TCP-timestamp absence reflects the path, not the server", eq("1")},
+	{"net.ipv4.tcp_window_scaling", "1", "so wscale absence reflects the path, not the server", eq("1")},
+	{"net.ipv4.icmp_ratelimit", "0", "nonzero throttles the server's ICMP errors → false loss/black-hole readings", eq("0")},
+}
+
+func eq(want string) func(string) bool { return func(v string) bool { return v == want } }
+
+// Sysctls runs the sysctl audit.
+func Sysctls() []Check {
+	out := make([]Check, 0, len(sysctlChecks))
+	for _, c := range sysctlChecks {
+		got := readSysctl(c.name)
+		sev := Warn
+		switch {
+		case got == "":
+			got = "(unreadable)"
+		case c.ok(got):
+			sev = OK
+		}
+		out = append(out, Check{Name: c.name, Got: got, Want: c.want, Severity: sev, Why: c.why})
+	}
+	return out
+}
+
+// Run performs the full self-test: sysctl audit + egress MTU probes to the
+// given targets (each "host" — port is irrelevant for PMTUD).
+func Run(mtuTargets []string) Report {
+	r := Report{Sysctls: Sysctls()}
+	r.SysctlOK = true
+	for _, c := range r.Sysctls {
+		if c.Severity == Warn {
+			r.SysctlOK = false
+		}
+	}
+	r.MTUOK = true
+	for _, t := range mtuTargets {
+		res := probeEgressMTU(t)
+		r.EgressMTU = append(r.EgressMTU, res)
+		if !res.FullMTU {
+			r.MTUOK = false
+		}
+	}
+	if len(r.EgressMTU) == 0 {
+		r.MTUOK = false // couldn't prove it
+	}
+	return r
+}
+
+var _ = strconv.Atoi