fix(registry): evict stale lease on heartbeat so re-registration can fire

[atsyplikhin]

Summary

Fixes a server-side bug where a device that TTL-expired in etcd would stay missing from the registry until some unrelated event forced a fresh register().

Stacked on arm#41 (which is stacked on arm#38). This PR's own diff is just the registry fix below.

DeviceRegistry.refresh() returned early whenever the process-wide in-memory leases dict still held a handle for the device -- even after etcd had already TTL-expired that lease. The handle is never evicted on its own, so on a "stranded" heartbeat:

• etcd3gw's Lease.refresh() does not raise against an expired lease. Per the etcd keepalive spec the response omits the TTL field, and etcd3gw returns -1 in that case (see etcd3gw/lease.py).
• So the stale handle survived, has_lease() kept reporting True, and the heartbeat path in registry/service/main.py never sent requestRegistration.
• The device stayed absent from the registry until an unrelated fresh register() call (e.g. the edge's NATS-disconnect path) happened to fire.

Fix

refresh() now inspects the return value of lease.refresh() (and still catches transport/server errors): on TTL == -1 or an exception, it drops the stale handle and falls through to the existing recovery code.

if lease:
    try:
        new_ttl = lease.refresh()
    except Exception:
        new_ttl = -1
    if new_ttl >= 0:
        return
    self.leases.pop(lk, None)   # stale -> drop, fall through to recovery

After eviction has_lease() correctly reports False, the server asks the device to re-register, and the edge replies with its registration payload (the requestRegistration wire path already exists on both sides). Recovery completes on the next heartbeat -- no client-side change required.

The architectural choice ("registry asks the device to re-register on a stranded heartbeat") was already made; it was just gated behind a stale-state check that lied.

Test plan

• test_refresh_evicts_expired_lease_no_raise -- refresh() -> -1, no raise: handle evicted, has_lease() False, no phantom lease created when etcd has no data
• test_refresh_evicts_stale_lease_on_error -- refresh() raises: handle still evicted
• test_refresh_recovers_when_data_survives_lease -- stale lease but doc still in etcd: re-issues lease + re-stores data
• Existing TestRefreshHeartbeat tests updated to return a realistic int TTL from the mocked lease (a bare MagicMock return would not survive the new >= 0 comparison)
• Full registry suite: 48 passed

🤖 Generated with Claude Code

[atsyplikhin]

Reopening against arm/device-connect instead of the fork.