fips/docs/design/fips-transport-layer.md

# FIPS Transport Layer

<!-- markdownlint-disable MD024 -->

The transport layer is the bottom of the FIPS protocol stack. It delivers
datagrams between transport-specific endpoints over arbitrary physical or
logical media. Everything above — peer authentication, routing, encryption,
session management — is built on the services the transport layer provides.

## Role

A **transport** is a driver for a particular communication medium: a UDP
socket, an Ethernet interface, a serial line, a Tor circuit, a radio modem.
The transport layer's job is simple: accept a datagram and a transport
address, deliver the datagram to that address, and push inbound datagrams up
to the FIPS Mesh Protocol (FMP) above.

The transport layer deals exclusively in **transport addresses** — IP:port
or hostname:port addresses, MAC addresses, .onion identifiers, radio device addresses. These are
opaque to every layer above FMP. The mapping from transport address to FIPS
identity happens at the link layer after the Noise IK link handshake completes.
The word "peer" belongs to the link layer and above; the transport layer
knows only about remote endpoints identified by transport addresses.

A single transport instance can serve multiple remote endpoints
simultaneously — a UDP socket exchanges datagrams with many remote
addresses, an Ethernet interface communicates with many MAC addresses on the
same segment. Each endpoint may become a separate FMP link, but the
transport layer itself maintains no per-endpoint state.

## Services Provided to FMP

The transport layer provides four services to the FIPS Mesh Protocol above:

### Datagram Delivery

Send and receive datagrams to/from transport addresses. The transport
handles all medium-specific details: socket management, framing for stream
transports, radio configuration. FMP sees only "send bytes to address" and
"bytes arrived from address."

Inbound datagrams are pushed to FMP through a channel. The transport spawns
a receive task that pushes arriving datagrams (along with the source
transport address and transport identifier) onto a bounded channel. FMP
reads from this channel and dispatches based on the source address and
packet content.

### MTU Reporting

Report the maximum datagram size for a given link. FMP needs this to
determine how much payload can fit in a single packet after link-layer
encryption overhead.

MTU is fundamentally a per-link property. A transport with a fixed MTU
(Ethernet effective 1497, UDP default 1280) returns the same value for every
link — this is the degenerate case. Transports that negotiate MTU
per-connection (e.g., the BLE L2CAP CoC MTU) report the negotiated value
for each link individually.

The transport trait exposes two MTU methods:

- `fn mtu(&self) -> u16` — Transport-wide default MTU
- `fn link_mtu(&self, addr: &TransportAddr) -> u16` — Per-link MTU for a
  specific remote address. The default implementation falls back to
  `mtu()`, so transports with uniform MTU (like UDP) need not override it.

FMP uses `link_mtu()` when computing path MTU for SessionDatagram
forwarding and LookupResponse transit annotation.

### Connection Lifecycle

For connection-oriented transports, manage the underlying connection: TCP
handshake, Tor circuit establishment, BLE pairing. FMP cannot begin
the Noise IK link handshake until the transport-layer connection is
established.

Connection-oriented transports expose a non-blocking connect interface.
`connect(addr)` initiates the connection in a background task and returns
immediately. `connection_state(addr)` reports the current status:

```text
ConnectionState {
    None        No connection attempt in progress
    Connecting  Background task running
    Connected   Ready for send()
    Failed(msg) Error message from failed attempt
}
```

Connectionless transports (UDP, raw Ethernet) return `Connected`
immediately — no async work needed.

At the node level, `PendingConnect` entries track links waiting for
transport connection. `poll_pending_connects()` runs each tick, checks
`connection_state()`, and calls `start_handshake()` on success or
`schedule_retry()` on failure. This decouples transport-layer connection
(which may take seconds for Tor circuits) from the FMP event loop.

### Discovery (Optional)

Notify FMP when FIPS-capable endpoints are discovered on the local medium.
This is an optional capability — transports that don't support it simply
don't provide discovery events.

See [Discovery](#discovery) below for details.

## Transport Properties

Transports vary widely in their characteristics. FIPS operates over all of
them because the transport interface abstracts these differences behind a
uniform datagram service.

### Transport Categories

**Overlay transports** tunnel FIPS over an existing network layer, typically
for internet connectivity:

| Transport | Addressing | MTU | Reliability | Notes |
| --------- | ---------- | --- | ----------- | ----- |
| UDP/IP | host:port | 1280–1472 | Unreliable | Primary internet transport |
| TCP/IP | host:port | Stream | Reliable | Requires length-prefix framing |
| Tor | .onion | Stream | Reliable | High latency, strong anonymity |
| Nym | host:port | Stream | Reliable | Mixnet, outbound-only, strong anonymity |

**Shared medium transports** operate over broadcast- or multicast-capable
media:

| Transport | Addressing | MTU | Reliability | Notes |
| --------- | ---------- | --- | ----------- | ----- |
| Ethernet | MAC | 1500 | Unreliable | Raw AF_PACKET frames |
| WiFi | MAC | 1500 | Unreliable | Infrastructure mode = Ethernet |
| BLE | BD_ADDR | 2048 default | Reliable | Per-connection L2CAP CoC MTU |
| Radio | Device addr | 51–222 | Unreliable | Low bandwidth, long range |

**Point-to-point transports** connect exactly two endpoints:

| Transport | Addressing | MTU | Reliability | Notes |
| --------- | ---------- | --- | ----------- | ----- |
| Serial | None (P2P) | 256–1500 | Reliable | SLIP/COBS framing |
| Dialup | None (P2P) | 1500 | Reliable | PPP framing |

### Properties That Matter to FMP

**MTU**: Determines how much data FMP can pack into a single datagram after
accounting for link encryption overhead. Heterogeneous MTUs across the mesh
are normal — the IPv6 minimum (1280 bytes) is the safe baseline for FIPS
packet sizing.

**Reliability**: Whether the transport guarantees delivery. FIPS prefers
unreliable transports because running TCP application traffic over a reliable
transport creates TCP-over-TCP, where retransmission and congestion control
at both layers interact adversely. FIPS tolerates packet loss, reordering,
and duplication at the routing layer.

**Connection model**: Connectionless transports (UDP, raw Ethernet) allow
immediate datagram exchange. Connection-oriented transports (TCP, Tor, BLE)
require connection setup before FMP can begin the Noise IK link handshake,
adding startup latency.

**Stream vs. datagram**: Datagram transports have natural packet boundaries.
Stream transports (TCP, Tor) require framing to delineate FIPS packets
within the byte stream. The FMP common prefix includes a payload length
field that provides this framing directly, replacing the need for a
separate length-prefix layer.

**Addressing opacity**: Transport addresses are opaque byte vectors. FMP
doesn't interpret them — it just passes them back to the transport when
sending. This means adding a new transport type with a novel address format
requires no changes to FMP or FSP.

## Connection Model

### Connectionless Transports

Datagrams can be sent to any reachable address without prior setup. Links
are lightweight — a transport address is sufficient to begin communication.

| Transport | Notes |
| --------- | ----- |
| UDP/IP | Stateless datagrams; NAT state is implicit |
| Ethernet | Send to MAC address directly |
| Radio | Raw packets to device address |

### Connection-Oriented Transports

Explicit connection setup is required before FIPS traffic can flow. The link
must complete transport-layer connection before FMP authentication can
proceed.

| Transport | Connection Setup |
| --------- | ---------------- |
| TCP/IP | TCP three-way handshake |
| Tor | Circuit establishment (typically 10–60s, default timeout 120s) |
| Nym | SOCKS5 connect through mixnet (minutes possible, default timeout 300s) |
| BLE | L2CAP CoC connection |
| Serial | Physical connection (static) |

### Implications

**Link lifecycle**: Connectionless transports use a trivial link model.
Connection-oriented transports need a real state machine: Connecting →
Connected → Disconnected. Failure can occur during connection setup, adding
error handling paths that connectionless transports don't have.

**Startup latency**: Connection-oriented transports add delay before a peer
becomes usable. This ranges from milliseconds (TCP) to tens of seconds
(Tor circuit). Peer timeout configuration must account for
transport-specific setup times.

**Framing**: Stream transports must delimit FIPS packets within the byte
stream. The FMP common prefix includes a payload length field that provides
integrated framing. Datagram transports preserve packet boundaries naturally.

## UDP/IP: The Primary Internet Transport

For internet-connected nodes, UDP/IP is the recommended transport:

- **No TCP-over-TCP**: UDP's unreliable delivery avoids the adverse
  interaction between application-layer TCP retransmission and transport-layer
  TCP retransmission
- **NAT traversal**: UDP hole punching enables peer connections through NAT
  without relay infrastructure
- **Low overhead**: 8-byte UDP header, no connection state
- **Matches FIPS model**: FIPS is datagram-oriented; UDP preserves this
  naturally without framing

Raw IP with a custom protocol number would be simpler but is blocked by most
NAT devices and firewalls, limiting deployment to networks without NAT.

### Socket Buffer Sizing

The default Linux UDP receive buffer (`net.core.rmem_default`,
typically 212 KB) is insufficient for high-throughput forwarding. At
~85 MB/s, a 212 KB buffer fills in ~2.5 ms; any stall in the async
receive loop (decryption, routing, forwarding overhead) causes the
kernel to silently drop incoming datagrams.

FIPS uses `socket2::Socket` wrapped in `tokio::io::unix::AsyncFd` for
the UDP receive path. This replaces `tokio::UdpSocket` and enables
direct `libc::recvmsg()` calls with ancillary data parsing —
specifically the `SO_RXQ_OVFL` socket option, which delivers a
cumulative kernel receive buffer drop counter on every received
packet. The drop counter feeds into the ECN congestion detection
system (see [fips-mmp.md](fips-mmp.md#ecn-congestion-signaling)).

Socket buffers (`recv_buf_size`, `send_buf_size`) are configured at
bind time via `socket2`. Linux internally doubles the requested value
(to account for kernel bookkeeping overhead) and silently clamps to
`net.core.rmem_max` / `net.core.wmem_max` if the request exceeds the
host kernel limits. The full UDP transport configuration is in
[../reference/configuration.md](../reference/configuration.md). The
host-side sysctl requirements and how to set them persistently live
in
[../how-to/tune-udp-buffers.md](../how-to/tune-udp-buffers.md).

## Ethernet: The Local Network Transport

For nodes on the same LAN segment, raw Ethernet provides a direct transport
without IP/UDP overhead — 25 bytes more FIPS payload per frame compared to
UDP (1497 vs 1472 MTU).

- **No IP dependency**: Operates below the IP layer. Nodes on the same
  Ethernet segment can communicate without IP addresses or routing
  infrastructure
- **Broadcast neighbor detection**: Nodes discover each other via periodic beacon
  broadcasts on the shared medium, with no static peer configuration required
- **Higher MTU**: Standard Ethernet frames carry 1500 bytes of payload,
  yielding an effective FIPS MTU of 1497 after the 3-byte frame header
- **Matches FIPS model**: Like UDP, Ethernet is connectionless and
  unreliable — datagrams flow immediately to any MAC address on the segment

### Implementation

The Ethernet transport uses Linux AF_PACKET sockets in SOCK_DGRAM mode with
EtherType 0x2121, and BPF devices (`/dev/bpf*`) on macOS. SOCK_DGRAM mode
lets the kernel handle Ethernet header construction and parsing — the
transport deals only with payloads and MAC addresses; the macOS BPF backend
presents the same API and handles the 14-byte Ethernet header itself.

Data frames use a 3-byte header: a 1-byte frame type (`0x00`) followed by
a 2-byte little-endian payload length. The length field allows the receiver
to trim Ethernet minimum-frame padding that would otherwise corrupt AEAD
verification. Beacon frames (`0x01`) use only the 1-byte type prefix
(fixed 34-byte payload). Beacons and data share the same EtherType and
socket.

| Property | Value |
| -------- | ----- |
| EtherType | 0x2121 |
| Socket type | AF_PACKET SOCK_DGRAM |
| Data frame header | `[type:1][length:2 LE][payload]` |
| Beacon frame header | `[type:1][payload]` (fixed 34 bytes) |
| Effective MTU | Interface MTU - 3 (typically 1497) |
| Addressing | 6-byte MAC address |
| Platform | Linux (AF_PACKET, `CAP_NET_RAW` required) and macOS (BPF `/dev/bpf*`) |

### Neighbor Beacons

Ethernet nodes discover peers via broadcast beacons sent to
ff:ff:ff:ff:ff:ff. Each beacon is a 34-byte frame containing the sender's
x-only public key. Receiving nodes extract the MAC source address from the
frame and the public key from the payload, then report the discovered peer
to FMP.

Four configuration flags control neighbor behavior — `listen`
(listen for beacons), `announce` (broadcast beacons), `auto_connect`
(initiate handshakes to discovered peers), and `accept_connections`
(accept inbound handshakes). The flag table and per-flag defaults
live in [../reference/configuration.md](../reference/configuration.md)
under `transports.ethernet.*`.

A typical discoverable node sets `announce`, `auto_connect`, and
`accept_connections` all true. A passive listener uses just
`listen: true` to observe the network without announcing itself.

### WiFi Compatibility

WiFi interfaces in infrastructure (managed) mode work transparently for
unicast — the mac80211 subsystem handles frame translation between 802.11
and 802.3. Broadcast neighbor detection is unreliable in managed mode because
access points commonly isolate clients from each other's broadcast traffic.

Startup logging:

```text
Ethernet transport started name=eth0 interface=eth0 mac=aa:bb:cc:dd:ee:ff mtu=1497 if_mtu=1500
```

## TCP/IP: Transport for UDP-Filtered Networks

For peers whose networks filter outbound UDP, the TCP transport
provides an alternative datagram path between public endpoints. TCP
is not a NAT-traversal mechanism — there is no `tcp:nat` analogue to
the UDP hole-punch flow.

FIPS protocols (FMP, FSP, MMP) are all unreliable datagrams. Running them
over TCP introduces head-of-line blocking, which adds latency jitter. MMP
correctly measures this jitter, and cost-based parent selection naturally
penalizes TCP links (higher SRTT leads to higher link cost). ETX will be
1.0 over TCP since TCP handles retransmission.

### Architecture

Unlike UDP (one socket serves all peers), TCP requires one `TcpStream` per
peer. The transport maintains two pools: a `ConnectingPool` for background
connection attempts in progress, and an established connection pool
(`HashMap<TransportAddr, TcpConnection>`) for active connections, plus an
optional `TcpListener` for inbound connections.

| Property | Value |
| -------- | ----- |
| Addressing | host:port — IP address or DNS hostname |
| Default MTU | 1400 bytes |
| Per-link MTU | Derived from `TCP_MAXSEG` socket option |
| Framing | FMP header-based (zero overhead) |
| Connection model | Non-blocking connect, connect-on-send fallback, optional listener |
| Platform | Cross-platform (no `#[cfg]` gates) |

### FMP Header-Based Framing

TCP is a byte stream; FIPS packets need delineation. Rather than adding a
separate length-prefix layer, the TCP transport uses the existing 4-byte
FMP common prefix `[ver+phase:1][flags:1][payload_len:2 LE]` to determine
packet boundaries:

- **Phase 0x0 (established)**: remaining = 12 + payload_len + 16 (header + AEAD tag)
- **Phase 0x1 (msg1)**: remaining = payload_len (fixed at 110, total 114 bytes)
- **Phase 0x2 (msg2)**: remaining = payload_len (fixed at 65, total 69 bytes)
- **Unknown phase**: close connection (protocol error)

This provides zero framing overhead and built-in phase validation. The
stream reader is implemented in a separate module (`stream.rs`) for reuse
by the Tor transport.

### Connection Establishment

TCP connections use a non-blocking connect model. When FMP needs to reach
a configured peer address, the node calls `connect(addr)` on the transport,
which spawns a background tokio task to perform the TCP handshake and socket
configuration (TCP_NODELAY, keepalive, buffer sizes, TCP_MAXSEG query). The
call returns immediately without blocking the event loop.

The node tracks each pending connection in a `PendingConnect` entry. On
every tick, `poll_pending_connects()` calls `connection_state(addr)` to
check progress. When the transport reports `Connected`, the completed
connection is promoted to the established pool (stream split into
read/write halves, per-connection receive task spawned), and the node
initiates the Noise IK link handshake. If the transport reports `Failed`,
the node schedules a retry with exponential backoff.

As a fallback, `send(addr, data)` still performs synchronous
connect-on-send if no connection exists — this handles the case where a
send arrives before the node-level connect path runs. The non-blocking
path is the primary mechanism for configured peers.

### Session Independence

TCP connection loss does **not** tear down the FIPS peer. Noise keys, MMP
state, and FSP sessions are bound to the peer's npub, not the TCP
connection. The transport reconnects transparently via the non-blocking
connect path or connect-on-send fallback. MMP liveness timeout is the sole
authority for peer death.

### Connection Deduplication

Simultaneous outbound connections from both sides are resolved by the
existing cross-connection tie-breaker in `promote_connection`. The losing
TCP connection is closed via `Transport::close_connection(addr)`, which
removes it from the pool and aborts its receive task.

### Configuration

The TCP transport configuration block (`transports.tcp.*` — bind
address, MTU, connect timeout, TCP_NODELAY, keepalive, socket buffer
sizes, max inbound connections) is documented in
[../reference/configuration.md](../reference/configuration.md). If
`bind_addr` is configured, the transport accepts inbound connections;
without it, the transport operates in outbound-only mode (no listener
socket is created).

## Tor: The Anonymity Transport

The Tor transport routes FIPS traffic through the Tor network, hiding
a node's IP address from its peers. A node behind Tor connects outbound
through a local Tor SOCKS5 proxy; the remote peer sees the Tor exit
node's IP, not the initiator's. After the Noise IK handshake, the remote
peer knows the initiator's FIPS identity (npub) but not its network
location.

Like TCP, Tor is connection-oriented and reliable. The same TCP-over-TCP
considerations apply — MMP correctly measures the elevated latency and
cost-based parent selection naturally deprioritizes Tor links.

### Architecture

The Tor transport is a separate `TorTransport` implementation, not a TCP
variant, because it manages SOCKS5 proxy negotiation, has different
address semantics (.onion vs IP:port), and has significantly different
latency characteristics. It reuses the FMP header-based stream reader
(`tcp/stream.rs`) for packet framing on the underlying TCP connection.

The transport maintains two pools (same pattern as TCP): a
`ConnectingPool` for background SOCKS5 connection attempts, and an
established pool of `TorConnection` entries. Each `TorConnection` holds
a write half, a per-connection receive task, the negotiated MTU, and
a connection timestamp.

| Property | Value |
| -------- | ----- |
| Addressing | .onion:port or IP:port |
| Default MTU | 1400 bytes |
| Framing | FMP header-based (shared with TCP) |
| Connection model | Non-blocking connect, outbound SOCKS5 + inbound via onion service |
| Platform | Cross-platform (requires external Tor daemon) |

### Address Types

The Tor transport accepts three address formats, parsed into a `TorAddr`
enum:

- **Onion**: `.onion:port` — connects to a Tor hidden service. Both
  sides anonymous. (e.g., `abcdef...xyz.onion:8443`)
- **Clearnet IP**: `IP:port` — connects through a Tor exit node to a
  remote TCP listener. Hides the initiator's IP; the remote peer sees
  the exit node's IP.
- **Clearnet Hostname**: `hostname:port` — hostname is passed through
  SOCKS5 for Tor-side DNS resolution, avoiding local DNS leaks. Compatible
  with SafeSocks 1. (e.g., `fips.example.com:8443`)

All address types are routed through the same SOCKS5 proxy.

### Connection Establishment

Connection setup follows the same non-blocking pattern as TCP. When FMP
needs to reach a peer, the node calls `connect(addr)` on the transport.
The transport spawns a background tokio task that:

1. Opens a SOCKS5 connection through the local Tor proxy
2. Configures the socket: `TCP_NODELAY`, keepalive (30s)
3. Returns the connected stream

The call returns immediately. `connection_state(addr)` reports progress.
Tor circuit establishment typically takes 10–60 seconds (vs milliseconds
for TCP), making non-blocking connect essential — a blocking connect
would stall the entire FMP event loop.

The connect timeout defaults to 120 seconds (vs 5 seconds for TCP),
accounting for Tor circuit setup time. As a fallback, `send(addr, data)`
performs synchronous connect-on-send if no connection exists.

### Inbound via Onion Service (Directory Mode)

In `directory` mode (recommended for production), Tor manages the onion
service via `HiddenServiceDir` in `torrc`. FIPS reads the `.onion` address
from the hostname file at startup and binds a local TCP listener that the
Tor daemon forwards inbound connections to.

This mode enables Tor's `Sandbox 1` (seccomp-bpf) — the strongest single
hardening option — because no control port interaction is required for
onion service management. Tor handles key generation and persistence
directly through the `HiddenServiceDir`.

The inbound accept loop mirrors the TCP transport's pattern: accept
connection, configure socket (TCP_NODELAY, keepalive), spawn a
per-connection receive loop using the shared FMP stream reader. Inbound
connections arrive from `127.0.0.1` (Tor daemon's local forwarding); peer
identity is resolved during the Noise IK handshake, not from the transport
address.

Configuration requires coordinating `torrc` and `fips.yaml`. The
operator setup — torrc directives, `fips.yaml` `tor` section,
HiddenServiceDir permissions, and `Sandbox 1` notes — is in
[../how-to/deploy-tor-onion.md](../how-to/deploy-tor-onion.md). In
brief: the `HiddenServicePort` external port is what peers connect
to, and `tor.directory_service.bind_addr` must match the
`HiddenServicePort` target address.

### Session Independence

Same as TCP: Tor connection loss does **not** tear down the FIPS peer.
Noise keys, MMP state, and FSP sessions survive reconnection.

### Bridge Node Pattern

A node running both Tor and UDP transports acts as a bridge between
anonymous and clearnet portions of the mesh:

```text
[Anonymous node] --tor--> [Bridge node] --udp--> [Clearnet node]
```

No special code is needed — FIPS multi-transport routing handles it.
Anonymous nodes connect to the bridge via Tor; the bridge forwards
traffic to clearnet peers over UDP. Clearnet peers never see the
anonymous node's IP.

### Latency Characteristics

Tor adds 200ms–2s RTT per circuit. MMP measures this elevated latency,
and cost-based parent selection penalizes Tor links (high SRTT → high
link cost). ETX is 1.0 since TCP handles retransmission.

Tor throughput is typically 1–5 Mbps — adequate for control plane and
moderate data transfer, not for bulk transfer.

### Monitoring

In `control_port` mode and optionally in `directory` mode (when
`control_addr` is configured), the transport spawns a background
monitoring task that polls the Tor daemon every 10 seconds via the
control port. The cached monitoring data is exposed through the
`show_transports` control socket query and displayed in fipstop.

Monitoring data includes:

- **Bootstrap progress** (0–100%) with INFO logging at milestones
  (25/50/75/100%) and WARN if stalled >60s
- **Circuit status** (whether Tor has a working circuit)
- **Network liveness** (up/down) with WARN on transitions
- **Dormant mode** detection with WARN on entry
- **Tor daemon version** and **traffic counters** (bytes read/written)

The control port connection uses cookie authentication by default
(reading from `/var/run/tor/control.authcookie`). Unix socket
connections (`/run/tor/control`) are preferred over TCP for security.

### Configuration

The Tor transport block (`transports.tor.*`) is documented in
[../reference/configuration.md](../reference/configuration.md). Three
modes are available:

- **`socks5`** (default): Outbound-only through a SOCKS5 proxy. No
  control port, no inbound connections.
- **`control_port`**: Outbound via SOCKS5 plus control port connection
  for Tor daemon monitoring. No inbound connections.
- **`directory`** (recommended for inbound): Outbound via SOCKS5 plus
  inbound via Tor-managed `HiddenServiceDir` onion service.
  Optionally connects to the control port for monitoring when
  `control_addr` is set. Enables Tor's `Sandbox 1` for maximum
  security.

The Tor transport requires an external Tor daemon. Named instances
are supported for multiple proxy endpoints.

### Implementation Roadmap

- Outbound SOCKS5 connections to .onion, clearnet IP, and clearnet
  hostname addresses *(implemented)*
- Inbound connections via Tor onion service using `HiddenServiceDir`
  directory mode *(implemented)*
- Operator visibility: cached monitoring snapshot, control socket
  exposure, fipstop display, bootstrap/liveness logging *(implemented)*
- Embedded `arti` (Rust Tor implementation) for self-contained operation
  without an external Tor daemon *(future)*

### Statistics

The Tor transport exposes per-instance counters covering successful
send/receive, send/receive errors, connection establishment,
SOCKS5-level errors, MTU rejections, accepted/rejected inbound
connections, and Tor control-port errors. The full counter table
lives in [../reference/transports.md](../reference/transports.md).

## Nym: The Mixnet Transport

The Nym transport routes FIPS traffic through the Nym mixnet, providing
network-level anonymity via Sphinx packet routing and timing
obfuscation. It uses the "mixnet-as-proxy" pattern: a node connects
outbound through a local `nym-socks5-client` SOCKS5 proxy, which carries
the traffic into the mixnet. The `nym-socks5-client` runs as a separate
process alongside the fips daemon and must be started independently.

Like Tor, Nym is a privacy-oriented deployment mode chosen for the
anonymity properties of the mixnet, not a failover for other transports.
Like TCP and Tor, it is connection-oriented and reliable; the same
TCP-over-TCP considerations apply, and cost-based parent selection
naturally deprioritizes the high-latency Nym links.

### Architecture

The Nym transport is a separate `NymTransport` implementation. It reuses
the FMP header-based stream reader (`tcp/stream.rs`) for packet framing
on the underlying byte stream, and follows the same connection-pool
pattern as the TCP and Tor transports.

It maintains two pools: a `ConnectingPool` for background SOCKS5
connection attempts, and an established pool of `NymConnection` entries.
Each `NymConnection` holds a write half, a per-connection receive task,
the configured MTU, and a connection timestamp.

| Property | Value |
| -------- | ----- |
| Addressing | IP:port or hostname:port |
| Default MTU | 1400 bytes |
| Framing | FMP header-based (shared with TCP) |
| Connection model | Outbound-only, non-blocking connect through SOCKS5 |
| Platform | Cross-platform (requires external nym-socks5-client) |

### Outbound-Only

The Nym transport is strictly outbound. It supports no inbound service:
`accept_connections()` returns `false` and `discover()` returns no
peers. A node using the Nym transport can initiate links to remote peers
through the mixnet, but cannot accept inbound connections over Nym. (A
node can still accept inbound links over other transports it runs.)

### Address Types

The Nym transport accepts two address formats, parsed into an internal
target address:

- **IP:port** — a numeric IP and port, sent to the SOCKS5 proxy as a
  numeric target.
- **Hostname:port** — the hostname is passed through SOCKS5 so it is
  resolved on the exit side rather than locally.

Both forms are routed through the same SOCKS5 proxy.

### Connection Establishment

Connection setup follows the same non-blocking pattern as the TCP and
Tor transports. When FMP needs to reach a peer, the node initiates a
background connect (`connect_async`). The transport spawns a background
tokio task that opens a SOCKS5 connection through the local
`nym-socks5-client`, configures the socket (including TCP keepalive),
splits the stream, and spawns a per-connection receive loop using the
shared FMP stream reader. The call returns immediately while the connect
proceeds in the background.

SOCKS5 connection setup through the mixnet can take much longer than a
direct TCP connection because each connection traverses multiple mix
nodes with timing obfuscation. Accordingly the connect timeout defaults
to 300 seconds (`connect_timeout_ms`). Non-blocking connect is essential
here — a blocking connect would stall the FMP event loop for the
duration of mixnet setup. As a fallback, `send_async(addr, data)`
performs a connect-on-send if no connection to the address yet exists.

Each outbound packet is checked against the configured MTU before being
written; an oversized packet is rejected with an MTU-exceeded error
rather than being sent.

### Startup Readiness

At startup the transport validates the configured `socks5_addr` and then
probes the SOCKS5 port to wait for `nym-socks5-client` to become ready,
using exponential backoff (starting at 1 second, capped at 10 seconds
between attempts) up to `startup_timeout_secs` (default 120 seconds). If
the proxy does not become reachable within that window, the transport
logs a warning and starts anyway; outbound connections then fail until
the `nym-socks5-client` becomes available.

### Session Independence

Same as TCP and Tor: loss of a Nym connection does **not** tear down the
FIPS peer. Noise keys, MMP state, and FSP sessions survive reconnection.

### Configuration

The Nym transport block (`transports.nym.*`) has the following fields:

| Field | Default | Description |
| ----- | ------- | ----------- |
| `socks5_addr` | `127.0.0.1:1080` | Address (host:port) of the local nym-socks5-client SOCKS5 proxy |
| `connect_timeout_ms` | `300000` | Outbound SOCKS5 connect timeout in milliseconds (300s) |
| `mtu` | `1400` | Maximum FIPS packet size for Nym connections, in bytes |
| `startup_timeout_secs` | `120` | Seconds to wait for nym-socks5-client to become ready at startup |

The Nym transport requires an external `nym-socks5-client`. Named
instances are supported for multiple proxy endpoints. Unknown
configuration keys are rejected.

### Statistics

The Nym transport exposes per-instance counters covering successful
send/receive, send/receive errors, connection establishment, SOCKS5-level
errors, connect timeouts, and MTU rejections.

## BLE: The Local Radio Transport

The BLE transport peers two nodes over Bluetooth Low Energy with no IP
network between them, using an L2CAP connection-oriented channel as the
byte pipe. It is the only transport whose reach is a radio horizon
rather than a route, which makes it the fallback when there is no
infrastructure at all: two phones in a room, a node and a handset, a
mesh with its uplink cut.

Like TCP, Tor and Nym it is connection-oriented and reliable, so the
same TCP-over-TCP considerations apply. Unlike them, its peer set is
discovered rather than configured, and the addresses it discovers are
not stable.

### Architecture

Nothing above the radio has a platform dependency. `BleTransport<I>` is
generic over a `BleIo` seam (`ble/io.rs`) that covers listening,
connecting, advertising, scanning and the stream I/O itself; the
connection pool, the PSM wire format, the stream framer and the
scan/probe loop are shared by every backend.

The backends live one per file and are selected by a cascade in
`ble/mod.rs`: `BluerIo` (`io_linux.rs`) talks to BlueZ over D-Bus;
`RadioIo` (`io_radio.rs`) drives a radio operated through commands and
callbacks — on Android the embedding application's, on macOS the
in-process CoreBluetooth radio in `io_macos/`; and `MockBleIo` (`io.rs`)
is an in-memory double compiled only under `cfg(test)`. A build that
matches none of the arms fails with a `compile_error!` rather than
silently selecting the mock.

That failure is deliberate. An earlier arrangement wrote the mock arm as
"anything that is not BlueZ", which meant a new platform got a transport
that compiled, started, reported itself Up and never peered, with no
error anywhere to find it.

### Backend Availability

`build.rs` sets `ble_available` for glibc Linux, Android or macOS, which is the
set of platforms with a concrete backend rather than the set that could
plausibly have Bluetooth. `bluer_available`, the BlueZ sub-condition, is
glibc Linux alone: musl cannot satisfy `libdbus-sys`'s pkg-config
cross-compile requirement, and musl router targets do not run BlueZ by
default. FreeBSD and Windows have no backend and so have no BLE transport
at all.

On glibc Linux the build needs `libdbus-1-dev` and `pkg-config`; the
BlueZ daemon itself is a runtime dependency. On Android the radio is
supplied by the application: scanning, advertising, L2CAP listen and
connect all sit behind Java APIs held under a permission and
foreground-service model that only the app can satisfy, so the embedder
implements `AndroidRadio` and installs it into a per-node slot which the
backend resolves per operation.

On macOS the radio is CoreBluetooth, reached through `objc2`. It is
delegate-driven and moves channel bytes through Foundation streams,
which is the same command-out, callback-in shape as the Android
embedder's radio, so the macOS backend reuses that machinery rather than
implementing `BleIo` itself: `MacRadio` (`io_macos/corebluetooth.rs`)
implements the radio trait and installs a fresh bridge into its slot
each time Bluetooth comes up. Both CoreBluetooth managers deliver their
callbacks on a private serial dispatch queue, and each channel's streams
are pumped by two threads doing blocking I/O, so nothing needs the main
thread's run loop. CoreBluetooth is privacy-gated: the `fips` binary
embeds an `Info.plist` carrying `NSBluetoothAlwaysUsageDescription`, and
the process needs Bluetooth permission — which macOS never grants a root
launchd daemon. So when CoreBluetooth refuses the daemon, it listens on
`/var/run/fips/ble-agent.sock` instead, and a per-user LaunchAgent runs
`fips-ble-agent --ble-agent` (a hard link to `fips`) in the login
session: the agent owns the radio and proxies it to the daemon
(`io_macos/agent*.rs`), carrying the radio commands, the bridge
callbacks and the channel bytes as length-prefixed frames. The agent
uses that fixed path, never the control socket's `/tmp` fallback, and
serves it only when `getpeereid()` says root answered. The daemon
installs the agent connection into its slot as one more radio, so the
transport above is unchanged; an agent that disconnects is a radio
switched off, and an agent whose own radio goes down disconnects. One
agent at a time: while one is connected, the daemon answers another's
hello with `Busy`, and that agent backs off. Outbound channel bytes are
credit-based per channel — the agent grants a window when the channel
opens and returns credit as its radio takes each packet, and the daemon
sends only while it holds credit — so a peer that stops granting L2CAP
credits stalls its own channel at the daemon rather than the agent's
session. CoreBluetooth also hides link
addresses; a peer's per-host identifier is hashed into the six bytes of
a `BleAddr`, so a Mac can dial only scan-discovered peers, not a
configured `hci0/AA:…` address.

### Framing

The channel is L2CAP CoC, not GATT, so there is no ATT_MTU to negotiate.
The per-connection CoC MTU applies, defaulting to 2048, and it overrides
the transport-wide default per link.

Packet boundaries are recovered from the byte stream rather than assumed
from the socket. BlueZ's `SOCK_SEQPACKET` preserves SDU boundaries, but
that is a property of one backend's socket type and not of L2CAP:
Android's `BluetoothSocket` input stream and macOS's `CBL2CAPChannel`
may return a fragment of a packet or several packets coalesced in one
read. FIPS packets are self-delimiting through the 4-byte FMP common
prefix, so `stream_read.rs` adapts the datagram-shaped stream into the
`AsyncRead` that `transport::framing::read_fmp_packet` already expects,
shared with every other stream-oriented transport.

### Discovery and the PSM

Discovery is an LE advertisement, received passively, carrying the
128-bit FIPS service UUID plus the listener's L2CAP PSM as service data.

The PSM has to ride the advertisement because it is not knowable any
other way. BlueZ lets an application choose the PSM it binds, and BlueZ
is the exception: Android's `listenUsingInsecureL2capChannel` and
macOS's `CBPeripheralManager.publishL2CAPChannel` both return an
OS-assigned PSM the application cannot request. A dialer cannot guess
it, and before a connection exists there is no channel on which to be
told. So `BleIo::listen` reports the PSM it actually bound,
`start_advertising` takes that PSM, and the scanner yields it alongside
the address.

macOS cannot advertise it: `CBPeripheralManager` advertises a local name
and service UUIDs and nothing else. A Mac therefore also serves its PSM
over GATT, as Apple's L2CAP PSM characteristic in a primary service
carrying the FIPS service UUID, and a Mac dialling a peer whose advert
carried no PSM reads that characteristic on the connection the channel
needs anyway, falling back to the requested PSM when the peer serves
none. The characteristic is specified in `ble/psm.rs` beside the
service-data layout. Only macOS serves or reads it today, so Linux and
Android cannot dial a Mac; with node-ordered link arbitration admitting a
lone link whichever side dialled it, the Mac's own dial is enough for
every pairing, Mac ↔ Mac included.

The wire layout is fixed by a byte budget and specified in `ble/psm.rs`.
A legacy advertising PDU carries 31 bytes of AD payload. Flags take 3
and the 128-bit service UUID list takes 18, so keying the service data
on the full 128-bit UUID would need 20 more and overrun by 10. Keying it
on the 16-bit UUID `0x9C90`, which is the leading 16 bits of the FIPS
service UUID expanded through the Bluetooth base UUID, takes 6 and fits
at 27. The budget is asserted at compile time. It leaves no room for a
local name, and it must ride the primary advertisement rather than the
scan response, because a scan response arrives only after an active-scan
round trip that drops asymmetrically across chipsets.

### Connection Establishment

A scan/probe loop dials discovered addresses, keeping the learned PSM
per address beside a probe-cooldown book and falling back to the
configured `DEFAULT_PSM` for a peer that advertises none.

A peer with a verified link is identified by node address, not by link
address. A link is verified once the node's FIPS session with that peer
uses it; the key a remote claims in the pre-handshake exchange proves
nothing on its own. A device using resolvable private addresses rotates
continually, and modern phones do so by default, so an address-keyed
pool sees every rotation as a new device and every already-connected
guard fails to fire.

A second link to an already-linked node is arbitrated, never simply
added, at the same address or another. A lone link is always admitted,
whichever side dialled it. Beside a verified link a second link is
declined: the verified one carries the node's session. Unverified links
still settle an honest simultaneous dial. When the second link completes
within one connect timeout of the first and runs in the opposite
direction, the two raced — typically both nodes dialled each other — and
both ends keep the link the node with the smaller address dialled. Each
end admits the two links in its own order, so "keep the first" can leave
each side holding the channel the other just closed; the node order is
the one input both ends share. A second link at the address of an
unverified one is declined unless it wins that race, since the pool holds
one link per address. Any other second link is admitted beside the
unverified one, because a claim that has not verified proves nothing,
and the extra link leaves by eviction when the pool needs the slot. So
address rotation and late redials cannot churn a verified link.

Failing addresses back off by powers of two up to
`MAX_PROBE_BACKOFF_SHIFT`, and the retry book is capped at
`MAX_PENDING_PROBES` so that rotating addresses cannot grow it without
bound. Both bounds matter more here than on other transports because BLE
hardware caps concurrent connections at roughly four to ten, so a
handful of unreachable addresses can starve discovery of everything
behind them.

Inbound connections are admitted off the accept loop, with
`INBOUND_HANDSHAKE_INFLIGHT` handshakes allowed at once and the oldest
aborted at the bound rather than the loop waiting for a slot.

## Discovery

Discovery determines that a FIPS-capable endpoint is reachable at a given
transport address. It is distinct from raw transport-level endpoint
detection — a new TCP connection or UDP packet from an unknown source is not
discovery; a FIPS-specific announcement or response is.

Discovery is an optional transport capability. Transports that don't support
it (configured UDP endpoints, TCP, Tor) simply don't provide discovery events.
FMP handles both cases uniformly: with discovery, it waits for events then
initiates link setup; without discovery, it initiates link setup directly to
configured addresses.

### Local/Medium Discovery

For transports where endpoints share a physical or link-layer medium — LAN
broadcast, radio, BLE — discovery uses beacon and query mechanisms:

- **Beacon**: A node periodically broadcasts its FIPS presence on the shared
  medium. Content is a FIPS-defined discovery frame carrying enough
  information to initiate a link. Non-FIPS endpoints ignore the frame.
- **Query**: A node broadcasts a one-shot solicitation. FIPS-capable nodes
  respond. Responses arrive on the same channel as beacon events.

Both produce the same result: "FIPS endpoint available at transport address
X." FMP does not need to distinguish beacons from query responses.

| Transport | Discovery | Notes |
| --------- | --------- | ----- |
| UDP (LAN) | Broadcast/multicast | On local network segment |
| Ethernet | Broadcast | Custom EtherType, ff:ff:ff:ff:ff:ff |
| Radio | Beacon | Shared RF channel, natural fit |
| BLE | Advertising | LE advertisement: 128-bit FIPS service UUID plus service-data PSM |

### Nostr Relay Discovery

For internet-reachable transports, a node publishes a signed Nostr event
containing its FIPS discovery information — public key and reachable
transport endpoints (UDP host:port, TCP host:port, .onion address). Other FIPS
nodes subscribing on the same relays learn about available peers.

Nostr relay discovery is not a transport — it is a discovery service that
feeds addresses to other transports. A node discovers via Nostr that a peer
is reachable at UDP 1.2.3.4:9735, then establishes the link over the UDP
transport.

For NAT'd UDP endpoints, a node may advertise `addr: "nat"` instead of a
concrete address, signaling that peers should initiate STUN-assisted UDP
hole punching. Offer/answer exchange uses Nostr gift-wrap (NIP-59) events
on the configured DM relays; the resulting punched socket is adopted into
the standard UDP transport via the bootstrap handoff path.

Key properties:

- Identity is built in — Nostr events are signed, so discovery information
  is authenticated
- Relay selection acts as scoping — which relays a node publishes to and
  subscribes on determines its discovery neighborhood
- Can only advertise IP-reachable endpoints (not radio, BLE, serial)
- Higher latency than local discovery (relay propagation delays)

### Current State

> **Implemented**: UDP, TCP, Tor, Ethernet, and BLE peers can be configured
> statically via YAML. Ethernet peers can also be discovered via beacon
> broadcast and BLE peers via LE scanning — the `discover()` trait method
> returns newly seen endpoints, and per-transport `auto_connect()` /
> `accept_connections()` policies control whether discovered peers are
> connected automatically or require explicit configuration. TCP and Tor
> have no built-in discovery mechanism.
> Nostr relay discovery and STUN-assisted UDP hole punching are
> implemented and toggled via configuration; see
> [../reference/configuration.md](../reference/configuration.md) for the
> `node.rendezvous.nostr.*` configuration tree. LAN/mDNS peer rendezvous
> is implemented as a separate subsystem and documented in
> [fips-nostr-discovery.md](fips-nostr-discovery.md).

## Transport Interface

The transport interface defines what every transport driver must provide.

### Trait Surface

```text
transport_id()        → TransportId         Unique identifier for this transport instance
transport_type()      → &TransportType      Static metadata (name, connection-oriented, reliable)
name()                → Option<&str>        Instance name (for multi-instance transports)
state()               → TransportState      Current lifecycle state
mtu()                 → u16                 Transport-wide default MTU
link_mtu(addr)        → u16                 Per-link MTU (defaults to mtu())
start()               → lifecycle           Bring transport up (bind socket, open device)
stop()                → lifecycle           Bring transport down
send(addr, data)      → delivery            Send datagram to transport address
connect(addr)         → ()                  Initiate non-blocking connection (connection-oriented only)
connection_state(addr)→ ConnectionState     Poll connection status (None/Connecting/Connected/Failed)
close_connection(addr)→ ()                  Close a specific connection (no-op for connectionless)
congestion()          → TransportCongestion  Local congestion indicators (optional)
discover()            → Vec<DiscoveredPeer> Report discovered FIPS endpoints (optional)
auto_connect()        → bool                Auto-connect discovered peers (default: false)
accept_connections()  → bool                Accept inbound handshakes (default: true)
```

### Receive Path

Rather than a synchronous receive method, transports use a channel-push
model. Each transport takes a sender handle at construction and spawns an
internal receive loop that pushes inbound datagrams onto the channel. The
node's main event loop reads from the corresponding receiver, which
aggregates datagrams from all active transports into a single stream.

Each inbound datagram carries:

- **transport_id** — which transport it arrived on
- **remote_addr** — the transport address of the sender
- **data** — the raw datagram bytes
- **timestamp** — arrival time

### Transport Metadata

Transport types carry static metadata that FMP can query:

```text
TransportType {
    name              "udp", "ethernet", "tor", etc.
    connection_oriented   bool
    reliable              bool
}
```

Predefined types exist for UDP, TCP, Ethernet, WiFi, Tor, Nym, BLE, and
Serial.

### Congestion Reporting

Transports optionally report local congestion indicators via a
`TransportCongestion` struct, providing a transport-agnostic interface for
the node layer's ECN congestion detection:

```text
TransportCongestion {
    recv_drops: Option<u64>    Cumulative kernel-dropped packets (monotonic)
}
```

The node samples each transport's congestion state on a 1-second tick via
`sample_transport_congestion()`. `TransportDropState` tracks per-transport
drop deltas: when new drops appear (rising edge), the `dropping` flag is
set, and `detect_congestion()` in the forwarding path triggers CE marking
on all forwarded datagrams.

| Transport | Congestion Source | Mechanism |
| --------- | ----------------- | --------- |
| UDP | `SO_RXQ_OVFL` kernel drop counter | `recvmsg()` ancillary data on every packet |
| TCP | Not implemented | Returns `None` (TCP handles congestion internally) |
| Tor | Not implemented | Returns `None` (TCP handles congestion internally) |
| Nym | Not implemented | Returns `None` (TCP handles congestion internally) |
| Ethernet | Not implemented | Returns `None` |

### Transport Addresses

Transport addresses (`TransportAddr`) are opaque byte vectors. The transport
layer interprets them — e.g. UDP and TCP resolve `host:port` strings (IP
fast path, DNS fallback with a 60s cache on UDP). All layers above treat
them as opaque handles passed back to the transport for sending.

### Transport State Machine

```text
Configured → Starting → Up → Down
                         ↓
                       Failed
```

Transports begin in `Configured` state with all parameters set. `start()`
transitions through `Starting` to `Up` (operational). `stop()` moves to
`Down`. Transport failures move to `Failed`.

## Interface Presence

`Up` describes the *transport*, not the socket. An interface-bound transport
(today: Ethernet) carries a second, orthogonal state — whether it is bound
right now — and the two are independent: a transport is `Up` from the moment
it starts, whether or not the interface it names exists.

### The Gap This Closes

Three deployment scenarios exercise one missing mechanism:

- **Boot ordering.** On OpenWrt, procd starts `fips` before wifi has created
  `fips-mesh0` / `fips-ap0`. Both transports were skipped and never retried,
  while the 802.11s peer link formed anyway — that is mac80211, not the
  daemon — so the node looked healthy and reached nothing. The failure was
  expensive precisely because nothing an operator could see said the node was
  deaf.
- **Intermittent hardware.** A USB ethernet adapter named in `fips.yaml` is
  plugged in some days and not others. Its absence is normal and must be
  silent; its arrival must bind without operator action.
- **Mid-operation restart.** `wifi reload` for a channel change destroys and
  recreates the mesh interface within a couple of seconds. The socket dies,
  the receive loop spun on `Err` with no backoff and no exit, and nothing
  rebound.

These are not three features. They are one presence machine plus one policy
field. Before it existed, the first observation was final: an interface
missing at start was logged once and skipped for the life of the process, and
one that disappeared at runtime published a health change but was never
rebound.

### The Presence Machine

```text
Absent ──attach──> Binding ──ok──> Present
  ^                   │              │
  └──── fail/backoff ─┘              │
  └──────────── detach ──────────────┘
```

`start_async` binds if it can and otherwise returns `Ok` with the transport
`Up` and `Absent`; a per-transport binder task then binds when the interface
appears, tears the socket down when it goes away, and rebinds when it
returns. Two invariants do the work:

- **The transport object survives detach.** Config, `TransportId`,
  statistics and the neighbor buffer persist; only the file descriptor and
  its loops go. A transport is never destroyed because its interface went
  away.
- **Start-time absence and runtime detach are the same transition.** A node
  that boots before its wifi and a node whose wifi reloads at 03:00 take one
  code path. The old asymmetry — skip forever at start, publish health at
  runtime — is gone.

`TransportError::InterfaceUnavailable` is what makes absence branchable. A
missing interface and a typo'd interface name were the same flat
`StartFailed(String)`, so nothing downstream could tell a state from a fault.

### What Counts as Present

Presence means `IFF_UP` — the interface exists and the operator has enabled
it — and deliberately **not** `IFF_RUNNING`.

Carrier and bindability are different questions, and only the second belongs
in a bind gate. An `AF_PACKET` socket on a carrier-less bridge is valid and
starts carrying traffic the instant a member port comes up, with no rebind:
the socket outlives the carrier. Gating on `IFF_RUNNING` bought nothing and
cost three things:

- `br-lan` on a router with nothing in its LAN ports is `UP` with
  `NO-CARRIER`, so a healthy wifi-only router reported `Degraded` forever and
  errored for a fault it did not have;
- every carrier flap the socket would have survived became an unbind/rebind
  cycle — churn the presence machine then has to damp, a mechanism
  compensating for a policy error;
- an 802.11s interface that reports `RUNNING` only once it has peered cannot
  peer, because peering needs beacons, beacons need a bound socket, and the
  gate refuses to bind. A deadlock reachable on the hardware this mechanism
  was written for.

The signal `IFF_RUNNING` carries is not lost: `show_transports` reports
`interface.carrier` beside presence, so an operator can still tell a bound
transport carrying nothing from a working one. It is reported rather than
obeyed.

The probe is `getifaddrs` plus `ifa_flags` rather than an `SIOCGIFFLAGS`
ioctl: it needs no socket, so the watcher can probe before any file
descriptor exists, and it is spelled the same on Linux and the BSDs.

### Interface Identity

The configured name is the key, but a name is not a device. Both backends
bind by *device* — `AF_PACKET` stores `sll_ifindex`, a BPF descriptor follows
the interface it was attached to — so an interface deleted and recreated
under the same name leaves the socket attached to something that no longer
exists while the name resolves perfectly well.

Nothing else notices. A stale `AF_PACKET` socket never becomes readable, so
the receive loop neither errors nor exits, and send failures go to the caller
rather than to the binder. A listen-only node (`announce: false`, so no
beacon sender to fail) therefore sat `present` and deaf indefinitely after a
`wifi reload` — the original bug wearing a different hat. The bound index is
captured at bind and compared on every poll; a mismatch is a detach.

Hardware can also change underneath a name. If the name reappears with a MAC
other than the one last bound, that is a different device, so the cached
neighbor entries for that transport are dropped rather than resumed onto, and
the swap is logged at `warn`. Richer selectors (`match: { name | mac |
id_path }`) are deliberately deferred; the requirement here is only that FIPS
never silently resumes onto different hardware.

### Detection

| Platform | Source |
| -------- | ------ |
| Linux | netlink `RTNLGRP_LINK` (`RTM_NEWLINK` / `RTM_DELLINK`) |
| macOS, FreeBSD† | `PF_ROUTE` socket, `RTM_IFINFO` |
| Fallback | poll `getifaddrs` + flags, 1 s |

† Aspirational: the Ethernet transport is
`cfg(any(target_os = "linux", target_os = "macos"))`, so FreeBSD has no
interface-bound transport for a watcher to serve. The `PF_ROUTE` branch
compiles for the BSD family, but only macOS reaches it.

Where an event source exists, detection is sub-second. The poll stays
underneath as a backstop rather than as the mechanism, and must stay at ~1 s:
the probe is cheap, and letting the interval drift to tens of seconds
reintroduces exactly the latency the event source was added to remove.
Construction is best-effort — a kernel or sandbox that refuses the socket
yields a watcher that never fires, and the binder degrades to its poll.

Link-event payloads are **not parsed**. An event is a hint to re-run the
presence probe, which is cheap and authoritative; decoding
`nlmsghdr`/`ifinfomsg` to reach the same answer would add a parser whose bugs
would be presence bugs.

Two rate limits protect the binder from its own event source. Probes are
coalesced to ten a second, because `PF_ROUTE` has no group filter and
delivers every routing message on the host — route churn, ARP, DHCP renewals,
a VPN going up and down — each of which would otherwise drive a full
`getifaddrs` walk. And a persistently failing event source is counted, logged
once, backed off, and after five consecutive errors abandoned for the poll:
losing events is survivable because the poll is the backstop, but burning a
core on a socket that is readable-but-erroring is not.

Bind failures that are *not* absence back off 1 s → 30 s. Absence itself does
not back off; there is nothing to poll but the probe.

### Policy: `optional`

One field per transport, `transports.ethernet.*.optional`, default `false`:

| | absence | log | retries |
| --- | --- | --- | --- |
| `optional: false` (default) | node reports `Degraded` | `info` at boot / `warn` on a runtime detach, then `error` once if it lasts past 10 s | forever |
| `optional: true` | no health impact | `info`, and nothing after | forever |

Naming an interface in configuration is a statement that you expect it, so
the default is to complain; silence is opted into.

`optional` describes **the interface's presence, not the transport's
importance**. An optional interface that is present is used exactly as hard
as any other. No value of it makes a missing interface fatal at startup: the
only fatal case remains "no transports at all came up". If a deployment ever
needs absence to abort startup, that arrives as an explicit `on_absent: exit`
— never as a second meaning for `optional`.

A bind failure that is not absence — no `CAP_NET_RAW`, no readable
`/dev/bpf*`, a buffer the kernel refused — is a fault, not a state, and still
fails the daemon's start. Retrying those forever would convert a hard,
actionable deployment error into a daemon that retries a socket it can never
open behind a `Degraded` nobody is watching. Only absence is waited out at
start; a non-absence failure during a later *rebind* does back off, since by
then the node is serving and killing it would be the worse answer.

### Health

Node health is recomputed on every presence transition **in both
directions**, so a returning interface clears `Degraded` without a restart.
That makes `Degraded` a level rather than a latch: the supervisor's reason set
was monotonic, which was correct while no child could recover, but with
recovery it would have come to mean "something broke at some point since boot"
rather than "something is broken now". Absence lives in its own reversible
set, separate from the one-way `failed` set a start failure enters.

An absent transport still counts as *up*. It came up — `start_async` returned
`Ok` — so it does not push a single-transport node into the fatal
`NoTransports`, which would make a node that merely booted before its wifi
exit instead of waiting. Absence degrades; it never kills.

There is deliberately **no restart action in the supervisor FSM.** The
presence watcher and the rebind loop live inside the transport, next to the
file descriptor they manage, and once that exists a supervisor-authored retry
has nothing left to do — it would be a second mechanism racing the first for
the same socket. The supervisor learns about presence
(`Event::ChildAbsent` / `Event::ChildPresent`) and republishes health; it does
not drive rebinding.

Peer state gets no grace period, and needs none. A transport's detach edge
withdraws every peer whose active link runs over it, on the same path the
liveness reaper uses, so those peers and the routes through them are gone at
the edge rather than up to `link_dead_timeout_secs` later — there is nothing
left for a linger timer to bound. A send over an absent interface still
returns `InterfaceUnavailable`, and a *half-built* link is still held on that
error, because the binder is already working to bring the interface back and
the initiator's resend has somewhere to land; an established peer is not held.
The trade is deliberate: an absence shorter than the dead timeout that then
recovers now costs a re-peer where it previously cost nothing, and that is
accepted because black-holing is silent, poisons other nodes' routing and
takes the full timeout to clear, where a re-peer is bounded, visible and
self-healing. A recreated mesh interface comes back with the same MAC (it is
derived from the phy), so the local address peers hold is unchanged across the
rebind.

### Logging

Edges, never attempts. A loop that logs per attempt reproduces the hot log
spin this mechanism removed, at 1–30 s intervals forever on any router with
an unplugged WAN — and operators learn to filter it, which is how the next
real failure gets missed.

The edge itself is not an error. An interface missing when the daemon starts
and bound a moment later is the ordinary case the mechanism exists to absorb,
so it is `info`; calling it an error at t=0 and "recovered" at t=0.2 s is the
cry-wolf failure this rule exists to prevent. A runtime detach is `warn` — a
link coming and going is ordinary weather for a mesh daemon.

There is exactly one deadline. Ten seconds is the window in which absence
could still be a race — a radio, a container, a veth arriving late. Past it a
**required** interface is a fault an operator has to fix, and it is reported
once at `error`. Start-time absence and a runtime detach share that deadline
rather than getting one each, for the same reason they share a code path
everywhere else here. An `optional` interface never reaches `error`; that is
what `optional` means.

Once, not repeated. This was a 1 m / 10 m / 1 h ladder that re-announced the
same fact at rising severity and then went permanently quiet after an hour,
which got both halves wrong: it used the log as a store for something already
published continuously as state, and it stopped mentioning a fault that was
still live. Duration belongs in `interface.since_secs` and in how long
`Degraded` has been held, where a monitor can threshold it per deployment
instead of the daemon compiling one in.

Node health does not wait for the deadline. `Degraded` publishes on the first
edge, which is the signal an operator actually watches.

Successful rebinds are damped. Backoff covers failed binds; the opposite and
nastier case is binds that keep *succeeding* into a socket that dies moments
later, which a receive loop giving up on a persistent error while the
interface stays `UP` produces once per second, forever. The binder counts
consecutive bindings that die inside ten seconds, backs off on the same
1 s → 30 s curve, and past three of them stops announcing each bind as a
recovery — holding health where it is until a binding lasts.

### Egress MTU

`transport_mtu()` is the minimum across *bound* transports — `is_bound()`,
not `is_operational()`, because an interface-bound transport is operational
from the moment it starts whether or not it holds a socket. Filtering on the
weaker predicate let a transport whose interface had never appeared set the
whole node's IPv6 MTU from hardware that was not present.

Since a transport can now bind long after start, that minimum moves at
runtime, and every consumer has to read it live. `show_status`, the
control-socket snapshot and the session-layer fragmentation check always did.
The TUN reader and writer did not: they were handed a `u16` at spawn, so a
narrow interface binding later never tightened the TCP MSS clamp and the node
reported one effective MTU while clamping to another. The ceiling is now
shared with those threads — an atomic beside the per-destination
`path_mtu_lookup` they already read on the same packet — and recomputed on
every change to the bound set.

Both directions, for the same reason `Degraded` is a level rather than a
latch: a narrow interface arriving must tighten the clamp or traffic
egressing over it is clamped too loose, and that interface leaving must
release it or unplugging a low-MTU adapter leaves the node over-clamped until
it restarts. MSS is negotiated per connection at SYN time, so a change binds
connections opened after it and leaves established ones alone.

### Observability

`show_transports` carries an `interface` block per interface-bound transport:
netdev name, `presence` (`absent` / `binding` / `present`), `carrier`,
`policy` (`required` / `optional`), `since_secs`, `binds` and
`failed_attempts`. The two counters separate an interface that is flapping
from one that is there and refusing to bind, and `since_secs` measures the
absence *episode* rather than the phase — a bind that fails walks
`Absent → Binding → Absent`, and restarting the clock on those edges would
report a permanently unbindable interface as one second old forever.

`fipstop`'s transports view names the netdev and the absence policy in their
own columns and shows presence in the State column for these transports,
because `state` reads `up` from the moment the transport starts and is
therefore precisely the wrong answer in the one case someone is scanning that
column for. Both render sites sort by ascending transport id — creation
order, and so grouped by transport type — rather than by `HashMap` iteration
order, which was arbitrary and differed on every daemon restart.

### What This Retires

- The `hotplug.d/net` rule that restarted the daemon when the FIPS radio
  interfaces appeared, and the `wifi down; wifi up; sleep` dance provisioning
  performed to sequence around the race.
- The YAML comment-toggling in `fips-mesh-setup` / `fips-ap-setup` — the mesh
  and AP blocks ship enabled with `optional: true` and simply wait.
- The ad-hoc ENXIO socket reopen in the beacon sender: beacons now pause while
  absent because the task does not exist then, and recovery is the presence
  machine's job. One mechanism for every cause rather than one hack per
  symptom.
- The start-time versus runtime asymmetry in the supervisor.

It also covers the case none of those workarounds did: an interface that flaps
while the daemon is running.

## Implementation Status

| Transport | Status | Notes |
| --------- | ------ | ----- |
| UDP/IP | **Implemented** | Primary transport, AsyncFd/recvmsg, SO_RXQ_OVFL kernel drop detection |
| TCP/IP | **Implemented** | FMP header-based framing, non-blocking connect, per-connection MSS MTU |
| Ethernet | **Implemented** | AF_PACKET SOCK_DGRAM, EtherType 0x2121, neighbor beacons; Linux (AF_PACKET) and macOS (BPF) |
| WiFi | **Implemented** (via Ethernet transport, infrastructure mode) | mac80211 translates 802.11↔802.3; broadcast beacons unreliable through APs |
| Tor | **Implemented** | Outbound SOCKS5, inbound via onion service, .onion and clearnet addressing |
| Nym | **Implemented** | Outbound-only SOCKS5 through nym-socks5-client, mixnet anonymity, IP/hostname addressing |
| BLE | **Implemented** (glibc Linux, Android and macOS; experimental) | L2CAP CoC, per-connection MTU (2048 default), per-link MTU; musl, FreeBSD and Windows have no backend |
| Radio | Future direction | Constrained MTU (51–222 bytes) |
| Serial | Future direction | SLIP/COBS framing, point-to-point |

## Design Considerations

### TCP-over-TCP Avoidance

Running TCP application traffic over a reliable transport (TCP, Tor)
creates a layering violation where retransmission and congestion control
operate at both levels. When the inner TCP detects loss (which may just be
transport-layer retransmission delay), it retransmits, creating more traffic
for the outer TCP, which may itself be retransmitting. This amplification
loop degrades performance severely under any packet loss.

FIPS prefers unreliable transports for this reason. When a reliable transport
must be used (e.g., Tor), applications should be aware of the performance
implications.

### Multi-Transport Operation

A node can run multiple transports simultaneously. Peers from all transports
feed into a single spanning tree and routing table. If one transport fails,
traffic automatically routes through alternatives. A node with both UDP and
Ethernet transports bridges between internet-connected and local-only
networks transparently.

Multiple links to the same peer over different transports are possible. FMP
manages these independently — each link has its own Noise session, its own
MTU, and its own liveness tracking.

### Transport Quality and Path Selection

Transport characteristics (latency, bandwidth, reliability) affect path
quality. The spanning tree parent selection factors in link quality through
cost-based effective depth (`effective_depth = depth + link_cost`), where
`link_cost` is derived from locally measured MMP metrics (ETX and SRTT).
This allows the tree to prefer lower-latency, lower-loss links when the
quality difference is significant. Link cost is also the primary key in
`find_next_hop()` candidate ranking for data forwarding, which orders
candidates by `(link_cost, distance_to_dest, node_addr)`.

## References

- [fips-concepts.md](fips-concepts.md) — Protocol overview
- [fips-architecture.md](fips-architecture.md) — Layer architecture
- [fips-mesh-layer.md](fips-mesh-layer.md) — FMP specification (the
  layer above)
- [fips-mtu.md](fips-mtu.md) — How transport-reported `link_mtu`
  feeds the unified path-MTU model
- [../reference/wire-formats.md](../reference/wire-formats.md) —
  Transport framing details
- [../reference/configuration.md](../reference/configuration.md) —
  Per-transport configuration blocks
- [../reference/transports.md](../reference/transports.md) —
  Per-transport statistics counter inventory