feat(http): expose the control plane over HTTP

[V3RON]

What is this?

This PR gives simlock a network-facing HTTP API (closes #65). Until now the daemon was reachable only over its unix socket, so every agent had to live on the same machine as the simulators. With http.enabled turned on, a remote agent can lease a device, watch the acquisition progress, renew, and release — over plain HTTP behind a tunnel — while the daemon keeps sole authority over queueing, provisioning, and cleanup.

It ships off by default and changes nothing for existing CLI/MCP users: the gateway is a third sibling frontend over the same role interfaces, and the core is untouched except for one addition (per-request cancel).

How does it work?

Everything lives under /v1, documented in the new docs/HTTP-API.md:

• Auth: every route (except GET /v1/healthz) requires Authorization: Bearer slk_…. Tokens are minted with the new simlock token create --role agent|operator command and stored as SHA-256 hashes in tokens.json; the token is the requester identity, so one-lease-per-agent is enforced by credential rather than by an honor-system env var.
• Leasing is an async resource: POST /v1/lease-requests returns 201 immediately; the client then polls GET /v1/lease-requests/{id} (with ?wait= long-poll) or streams SSE from …/events through the same queued → provisioning → booting → granted stages the CLI prints today. DELETE cancels a still-queued request via a new core cancelPending, which reuses the queue timeout's safety envelope — a request with device work already in flight reports 409 REQUEST_NOT_CANCELLABLE instead.
• Leases are detached-only: TTL plus POST /v1/leases/{id}/renew — the held-mode "connection = lease" model deliberately doesn't cross the network. Renew responses carry notices (device_unhealthy / device_recovered) drained since the last renew, and GET /v1/leases/{id}/events streams the same facts live, ending with lease_lost.
• Operators get GET /v1/leases, /v1/devices, /v1/events(?since=|/stream) behind the operator role. nuke is deliberately not exposed over HTTP.
• Lifecycle: the gateway binds only after startup convergence completes (HTTP handlers call the role interfaces directly, so serving earlier could observe half-converged state) and is the first thing shut down on daemon stop, via a new stopAuxiliary hook. Lease payloads carry a reserved dataPlane: null field for the agent-device integration (agent-device integration: bind the HTTP control plane into agent-device as a remote lease provider #66) so its arrival is additive.

The app layer is a pure Request → Response Hono app — the only file touching a real listener is the @hono/node-server adapter — so the entire route surface unit-tests through app.request() with fakes and a manually-advanced clock, in keeping with the repo's ports discipline. New e2e coverage drives the full loop (token → lease → renew → release, plus 401/403/409 paths) against the fake driver.

Why is this useful?

• Remote agents — CI jobs, coding agents on other machines — can consume simlock's fair queueing and warm-pool machinery without shell access to the host; this is the control-plane half of the self-hosted-device-farm direction, with agent-device integration: bind the HTTP control plane into agent-device as a remote lease provider #66 tracking the data plane.
• Real authentication replaces the local-trust model at the network boundary: hashed bearer tokens, roles, and ownership checks (an agent cannot see or release another requester's lease).
• The async lease-request resource survives flaky networks: idempotency keys make retried POSTs safe, daemon-side timeoutMs means a vanished client cannot hold a queue slot, and a restarted client recovers its lease from GET /v1/leases/{id} instead of leaking it.
• Nothing moves for existing setups: the feature is opt-in config, the socket protocol is unchanged, and the one core addition (cancelPending) reuses existing rejection machinery and emits through the existing lease.rejected event family.