# QUIC edge v1 wire format

## Scope and topology

QUIC edge v1 carries two functions over one authenticated QUIC connection:

- a reliable bidirectional control stream for node registration, heartbeat,
  current local IP addresses, the advertised gRPC endpoint and media metadata;
- QUIC DATAGRAM payloads for lossy, real-time audio and video.

Robot arm, AGV and other device-control APIs remain on the existing cmvr-es
gRPC server. The QUIC node descriptor advertises that server's current endpoint;
it does not move robot-control RPCs onto QUIC.

```text
cmvr-es -- custom QUIC edge v1 --> Java receiver / edge gateway
                                               |
                                               +-- WebTransport/HTTP3 --> browser
```

A browser cannot connect to this custom QUIC ALPN directly. The gateway must
terminate v1, register the node, reassemble media fragments and expose
WebTransport (or another browser-supported media API). Browser sessions do not
participate in edge-node registration or heartbeat.

## Build and host-safe defaults

cmvr-es pins MsQuic v2.5.9 and builds it from source into the repository
dependency tree. It is not installed system-wide:

```bash
script/build_msquic.sh \
  --arch x86 \
  --version 2.5.9 \
  --jobs 2 \
  --clean

cmake -S . -B build-quic \
  -DCMVR_ARCH=x86 \
  -DBUILD_TESTING=ON \
  -DCMVR_INSTALL_DEFAULT_RUNTIME_ASSETS=OFF
cmake --build build-quic -j2
ctest --test-dir build-quic --output-on-failure
cmake --install build-quic
```

The source helper clones the official `v2.5.9` tag, verifies its reviewed
source commit, and initializes its QuicTLS submodule
into `build/third_party/`, then installs the relocatable public result under
`dependency/x86/third_party/msquic/v2.5.9`. Its first run therefore requires
network access, but neither root privileges nor a system MsQuic package. The
result uses a statically linked QuicTLS backend; `libmsquic.so` is copied to
`output/lib` when cmvr-es is installed.

The repository dependency is declared in `request.txt`, which adds its include
and library directories and installs `libmsquic.so*` with the other bundled
runtime libraries. MsQuic is therefore a required build dependency for
cmvr-es. The repository runtime configuration keeps `quic_edge` disabled, and
the existing inbound gRPC server remains enabled independently.

`CMVR_INSTALL_DEFAULT_RUNTIME_ASSETS=OFF` preserves an existing
`output/bin/config` and `output/bin/model` while updating the binary and
runtime libraries. Leave its default value `ON` when a clean copy of the
repository defaults is desired.

## Local real-QUIC verification

The development-only server in `test/quic_gateway` listens through real
MsQuic/TLS/UDP and implements the v1 control and DATAGRAM receiver. With the
strict build above, CTest registers:

- `cmvr_quic_msquic_e2e_test`, which connects the production edge service and
  feeds synthetic H.264 video plus AAC audio through `MediaSourceHub`, then
  verifies registration, heartbeat, descriptors, DATAGRAMs and frame
  reassembly;
- `cmvr_es_quic_process_smoke_test`, which starts the actual `cmvr_es`
  executable with a temporary no-device configuration and verifies
  registration and heartbeat ACKs against the local gateway.

Both tests use a short-lived loopback certificate generated under the build
tree. They do not modify `output/bin/config` or require physical devices. See
[`test/e2e/README.md`](../../../../test/e2e/README.md) and
[`test/quic_gateway/README.md`](../../../../test/quic_gateway/README.md) for
their exact scope and manual gateway options.

To enable node presence without media:

1. Configure the gateway address, ALPN `cmvr-quic-edge/1`, TLS trust, node ID,
   advertised gRPC endpoint and heartbeat values in
   `cmvr-es/config/tasks/quic_edge_task/quic_edge_task.pb.txt`.
   Certificate paths are relative to the cmvr-es configuration root, for
   example `certs/quic_gateway_ca.pem`.
2. Leave `tracks` empty, set `quic_edge.enable: true`, and enable the
   `quic_edge` TaskManager entry.

To add a media uplink, also enable the camera/microphone in DeviceManager and
add a matching `tracks` entry. A zero-track configuration is valid: unavailable
media must not suppress node registration, IP reporting or heartbeat.
`maximum_frame_bytes` must fit in one atomically admitted DATAGRAM batch; reduce
it or increase `datagram_send_queue_depth` when changing the DATAGRAM size.

## Reliable control stream

The edge opens one bidirectional reliable stream after the QUIC handshake. Both
directions carry a sequence of length-prefixed protobuf messages:

```text
uint32_be protobuf_length
protobuf EdgeControlEnvelope
```

The length excludes the four-byte prefix. Each receiver must reject a frame
larger than its configured control-frame limit. `protocol_version` is `1`.
`message_sequence` increases independently in each direction; gaps are allowed,
but duplicates and reordering are rejected. The edge resets its outbound
control sequence on every connection, so its first `NodeRegisterRequest`
currently has `message_sequence=0`; gateways must not assume the sequence starts
at 1. Heartbeat uses a separate sequence so its acknowledgement remains
explicit.

The legal session order is:

1. The edge sends `NodeRegisterRequest` as its first application message after
   every QUIC connect or reconnect. It includes `node_id`, the immutable
   `robot_id`, `boot_id`, software version, the current IPv4/IPv6 interface
   snapshot and advertised gRPC endpoint.
2. The gateway replies with `NodeRegisterResponse`. Media and heartbeat must not
   start until `accepted=true` and a non-empty `session_id` are received.
3. The edge sends `NodeHeartbeat` at the negotiated interval. The gateway returns
   `NodeHeartbeatAck` with the same `session_id` and exact
   `acknowledged_sequence`. A missing, rejected or mismatched ACK causes the edge
   to reconnect and register again after its configured backoff. Every heartbeat
   also carries a fresh `DeviceManagerSnapshot`; its `devices` list contains
   only enabled devices, including enabled devices whose construction,
   initialization or start failed.
4. If media tracks are available, the edge sends `MediaSessionOpen`, binding its
   `session_epoch` to the accepted `session_id`, followed by track descriptors.
   The edge calls reliable descriptor send before DATAGRAM send for a new
   generation. QUIC does not guarantee arrival order across a reliable stream
   and DATAGRAMs, so a gateway must boundedly buffer or drop media whose session
   or generation metadata has not arrived yet.
5. The gateway may send `ProtocolError`; a fatal error closes the connection.
   The current edge receive path otherwise accepts only `NodeRegisterResponse`
   and `NodeHeartbeatAck`. Although `MediaSessionClose` is present in the v1
   schema, the current edge neither sends it nor accepts it inbound. Gateways
   must not depend on that message until both sides implement and test it.

The edge sends a fresh local-interface snapshot in every heartbeat, so an IP
change is reported without a second protocol. These addresses are edge claims.
`observed_source_ip` is authoritative for the public/NAT-facing address and must
be derived by the gateway from the authenticated QUIC peer, never copied from a
client field.

The heartbeat's `device_manager.devices` contains only entries with
`enabled=true`. Disabled configuration entries remain available in the edge's
local DeviceManager snapshot but are not transmitted. An enabled entry that
fails creation, initialization or start remains visible with
`MANAGED_DEVICE_STATE_ERROR`, `has_error=true` and an error detail. Its
independent health value may still be `UNSPECIFIED` when no trustworthy device
probe exists. `kind` is the stable category used for machine decisions, while
`type_name` is a concrete implementation name when the object exists and
otherwise a category label; it is intended only for display and diagnostics.
Senders order entries by `device_id` to make
snapshots deterministic. `sampled_at_unix_ms` is the snapshot time, whereas
`status_updated_at_unix_ms` records when DeviceManager last changed that row's
lifecycle/error record. The enclosing snapshot time is the freshness timestamp
for the health observation.

Lifecycle and health are deliberately separate. In particular,
`DEVICE_HEALTH_STATUS_UNSPECIFIED` means that no trustworthy health observation
was available; it is not equivalent to `DEVICE_HEALTH_STATUS_HEALTHY`.
Similarly, `has_error=false` only means that no error is currently confirmed
and must not be used to turn unknown health into healthy health. A health probe
failure must degrade that row to an unknown or fault result without suppressing
the rest of the heartbeat.

The edge caps each diagnostic string at 512 bytes without splitting a UTF-8
code point. Device identifiers, implementation names and manager metadata are
not silently truncated because doing so would change identity. The deployment
must therefore size `maximum_control_frame_bytes` for its enabled inventory;
the sender and receiver both reject an oversized control frame. Very large
inventories require a future explicit pagination/truncation extension rather
than silently dropping rows from this enabled-device snapshot.

The locally configured `QuicEdgeConfig.heartbeat_interval_ms` controls the
reporting interval before registration. A gateway may negotiate a different
interval through `NodeRegisterResponse.heartbeat_interval_ms`: zero keeps the
locally configured value, while a non-zero value overrides it for the current
registered QUIC connection. The edge clamps the negotiated value to its
supported safety range and returns to the local value on reconnect until a new
registration response is accepted.

`NodeDescriptor.robot_id = 6`, `NodeHeartbeat.device_manager = 9` and
`NodeHeartbeat.robot_id = 10` are additive protobuf fields, so these extensions
remain QUIC edge protocol v1. Existing gateways ignore unknown fields. Updated
gateways must continue accepting older v1 messages where these fields are absent
and must not treat an absent device snapshot as an empty, healthy DeviceManager.
Enum values may only be appended; existing numeric meanings must never be
renumbered or reused.

DATAGRAM negotiation is required only when at least one media track is enabled.
The reliable registration and heartbeat path remains valid for a zero-track
node or while all media sources are unavailable.

Media source registration, device start/keyframe callbacks and DATAGRAM
packetization run on a dedicated media worker. Reliable registration and
heartbeat therefore continue while a media source is slow or unavailable;
the reliable send path serializes heartbeat and media metadata so envelope
sequence order remains strict.

`MediaSourceHub` is protocol-neutral and gives each adapter an independent,
single-consumer subscription cursor. Its source-start callback receives a
cancellation predicate and must check it around potentially blocking device
startup. QUIC reconnect/stop and gRPC client cancellation can therefore abandon
startup without blocking presence or process shutdown. Ring overflow keeps the
newest bounded set of immutable frames, advances a slow cursor to the oldest
retained frame, and reports the exact dropped count so protocol adapters can
mark a discontinuity and request a new video keyframe.

## Media DATAGRAM wire format

Every media DATAGRAM starts with this fixed 64-byte, network-byte-order header:

| Offset | Size | Field |
|---:|---:|---|
| 0 | 4 | magic `CMQD` (`0x434d5144`) |
| 4 | 1 | protocol version (`1`) |
| 5 | 1 | media kind (`1` video, `2` audio) |
| 6 | 2 | flags |
| 8 | 2 | header size (`64`) |
| 10 | 2 | fragment index |
| 12 | 2 | fragment count |
| 14 | 2 | payload size |
| 16 | 4 | track ID |
| 20 | 4 | codec generation token |
| 24 | 8 | media session epoch |
| 32 | 8 | DATAGRAM packet sequence |
| 40 | 8 | media frame sequence |
| 48 | 8 | monotonic capture timestamp in microseconds |
| 56 | 4 | complete frame size |
| 60 | 4 | fragment byte offset in the complete frame |

Flag bit 0 marks a video keyframe, bit 1 is reserved and must currently be
zero, and bit 2 marks a discontinuity. Codec initialization bytes are carried
reliably in `MediaTrackDescriptor`. QUIC already authenticates each DATAGRAM,
so the application header has no redundant checksum.

Receivers must key reassembly by `(session_epoch, track_id, frame_sequence)`,
drop incomplete frames at their playback deadline, and reject fragments whose
offset plus payload exceeds `frame_size`. A new session epoch invalidates all
fragments retained from a previous connection. `frame_sequence` is generated
by the QUIC edge service and remains strictly increasing per track throughout
the media epoch, including after a device source restarts.