Signals peer unit (atelier.service)

Atelier joins the Signals lattice as peer id atelier. systemd is only the signals.target membership hook. The wire contract is zndx.engine.v1.Engine on :50251.

Gaius-style wrap: ExecStart is devenv up -d. The lattice engine is the devenv process capability-engine, so devenv processes status|restart|stop capability-engine works. The same graph starts on Linux and macOS — devenv already gates linux-vs-darwin (CUDA libs, d2); there is no laptop-vs-server switch.

Lattice accept is Engine/Status at gRPC bind, not vLLM cold-load and not product HTTP. Status also advertises surfaces (kind=primary, devenv Vite :3300, never Metaflow :3000) and answers ServerQuery (REMOTES / PEERS / SURFACES / QUEUES / WORKLOADS). QueueHint is declared leaf shape. WorkloadHint carries model + capabilities + tp/pp (gpu_tokens = tp×pp); never encode those in the queue name. Pick heavy/medium/light from that share. Occupancy over time is zndx.scheduler.v1.Scheduler/RequestQueueShare (RFC 9562 UUIDv7 on every call; UNIMPLEMENTED = Signals-not-yet; REJECTED / other errors #YK.00000007.SHAREFAIL — do not admit). Zero-floor + valid_until when the WRK ends. Atelier never writes queues.yaml.

Fact	Value
Peer id	`atelier`
Unit	`atelier.service` (`After=signals-ready.service`)
Wrappers	`scripts/systemd_{start,stop,unit}.sh`
Engine process	devenv `capability-engine` → `scripts/processes/capability-engine.sh`
Status probe	`scripts/zndx_status_ok.py` (codegen stubs)
gRPC lattice	`:50251` — `zndx.engine.v1.Engine` (+ native `AtelierEngine`)
Product servicer	`:50071` (same devenv graph; not the accept gate)
Postgres	`:5533` — never Signals `:5455` / RustFS `:9010`
Capability	`referee` (and `instruct` when configured)
Status.project	`atelier`

Two gRPC ports (same graph, different accept)

Port	Process	Role
:50251	`capability-engine`	Lattice accept
:50071	`grpc-server`	Workbench product gRPC

Do not wait for vLLM SERVING — Status at bind is enough.

Wrappers

systemd_start.sh is idempotent: if a process-compose-owned listener already answers Engine/Status and a login-shell devenv processes sees that compose, it exits 0. The wrappers pin XDG_RUNTIME_DIR to /run/user/<uid> so systemctl restart atelier and devenv processes restart gateway share one graph. Foreign/old setsid engines and leftover /tmp/devenv-* daemons for this checkout are reaped, then devenv up -d (never foreground just up). Do not pin a dedicated DEVENV_RUNTIME in the unit — that splits systemd from the login-shell devenv.

systemd_stop.sh runs just down / devenv processes down for this checkout and reaps leftover atelier.engine PIDs. It does not:

run just teardown or GPU-deep-cleanup
write, delete, or scan /tmp/zndx-gpu-leases (Gaius :50051 / Ægir :50151 leases stay)

# Same graph as the unit
devenv up -d
devenv processes status          # includes capability-engine
grpcurl -plaintext 127.0.0.1:50251 list
grpcurl -plaintext 127.0.0.1:50251 zndx.engine.v1.Engine/Status
devenv processes restart capability-engine
just down

Operator (from Signals)

just install-systemd --peers atelier --enable
sudo systemctl start signals.target    # not bare "signals"
grpcurl -plaintext 127.0.0.1:50251 list
grpcurl -plaintext 127.0.0.1:50251 zndx.engine.v1.Engine/Status
just lattice-ci --require atelier

Platform Metaflow

export METAFLOW_SERVICE_URL=http://127.0.0.1:30180

Health

Lattice-ci probes Status + reflection only. Product health stays Atelier-local. GPU leases: /tmp/zndx-gpu-leases — stop must stay lease-aware so it does not wipe Gaius / Ægir.