feat(orchestrator): wire C2 command dispatch through the WS server

[martinjms]

Completes the wire layer that PR #43 (in-process dispatcher) deliberately deferred. A real driver can now register, receive server-pushed `Command` envelopes, and ack them — and the dispatcher's `submit()` resolves end-to-end against a real WebSocket.

This is the orchestrator-side counterpart to #27 (driver-side `--orchestrator-url` extension); together they form the full C2 wire path.

Protocol additions (backward-compatible)

• `DriverMessage::CommandAck(CommandAck)` — driver→orchestrator ack frame
• `OrchestratorResponse::Command(CommandEnvelope { seq, command })` — server-pushed command frame
• `DriverChannel::send` now takes a `seq` parameter so the dispatcher's monotonic sequence number flows through to the wire envelope and back via the matching ack

Server changes

• `DriverServer::with_dispatcher` constructor wires an `Arc` through to `handle_connection`. Old `new()` (no dispatcher) preserved for call sites that don't need command dispatch.
• `handle_connection` refactored into a reader/writer pair. The writer multiplexes command envelopes + handshake responses onto the WS sink; the reader handles register/heartbeat/deregister/ack messages and routes acks to per-seq `oneshot`s.
• `WsDriverChannel` (new module): production `DriverChannel` impl backed by a per-connection mpsc + oneshot pair; populated on Register, dropped on Deregister or disconnect.

Tests

Two new e2e tests in `tests/ws_command_e2e.rs` drive a real WS round-trip:

• `dispatcher_round_trips_command_to_real_driver_via_ws` — submit returns acks routed back through the wire
• `dispatcher_drops_channel_when_driver_deregisters` — the dispatcher's channel is dropped when a driver deregisters

Test plan

• `cargo test -p arcane-swarm-orchestrator` — 58 passed, 0 failed (2 new e2e + 56 prior)
• `cargo clippy -p arcane-swarm-orchestrator --all-targets -- -D warnings` clean
• `cargo fmt --all -- --check` clean

🤖 Generated with Claude Code