feat(agent-device): simlock runs the agent-device daemon

[V3RON]

Closes #470

Status

Implement: done (83/83 green) Review: round 2 + 2 fix checks passed Hardware: n/a Gate: running
Next: gate

Done when

• The tests above pass, src/lint-rules.test.ts included.

Assumptions

• Assumption: the AgentDeviceSupervisor constructor takes the ports and facts as named options (baseEnvironment, clock, config, driverEnvironment, filesystem, healthProbe, home, leaseMaxTtlMs (the value of lease.maxTtlMs), logger, nodePath, nodeVersion, platforms, processRunner, processSupervisor, tcpProbe) — the spec names the inputs but not the shape; the tests fix this one.

• Assumption: the supervisor may signal the child through the runner's handle or through ProcessSupervisor; the tests accept either — the spec says only "sends SIGTERM".

• Assumption: StartDaemonOptions.agentDeviceHealth is the test hook the spec names; the wiring tests pass it through a cast until the green step adds the field.

• Assumption: MemoryFilesystem creates (writeFileAtomic, writeFileExclusive, mkdir) follow a symlinked ancestor, as reads already did — the spec test with SIMLOCK_HOME reached through a symlink writes daemon.json under the unresolved spelling and reads it back resolved, which no implementation could satisfy otherwise.

• Assumption: the supervisor keeps its state dir and pid record under the real path of SIMLOCK_HOME, the spelling the marker argument uses — one spelling, so a symlinked home cannot split them.

Review

Spec review: 3 blocking, 1 fixed, 2 notes. Code review: 11 blocking, 10 fixed, 4 notes. Claims review: 0 blocking, 0 fixed, 0 notes.
Rounds: 2. Tokens: 10.3M.

Rejected:

• spec: supervisor.test.ts:1099 "removes the record" unproven — duplicate of the code finding, fixed there
• spec: supervisor.ts:248 start-failed remedy points to an empty log — duplicate of the code finding at :244, fixed there
• code: supervisor.test.ts:1099 deleting removeRecord stays green — duplicate of the spec finding, fixed
• code: supervisor.ts:745 stop-waiter removal untested — the leak is not observable without asserting private state

Fixed outside review: the daemon protocol bump 29 → 30 the spec asks for (fix check passed).

Written by an agent.

[V3RON]

Review notes

Not blocking, not verified. Each is one reviewer's claim.

• spec: src/agent-device/host/agent-device-supervisor.ts:114 a child killed by a signal reports "exited after a signal" without the signal name, though the spec asks for the code or signal.
• spec: src/agent-device/host/agent-device-supervisor.ts:525 with iOS on but no SIMLOCK_IOS_DEVICE_SET (the e2e fake driver returns {}), the daemon gets no device set and sees the host's default simulators.
• code: src/daemon/main.ts:485 agentDevice.stop() is awaited unguarded; if it rejects, every driver disposal is skipped, leaving simlock's adb server running.
• code: src/agent-device/host/agent-device-supervisor.ts:598 the warn line naming the surviving pid can be deleted with a green suite, because #fail already logs a message the test matches.
• code: src/gateway/service.test.ts:1205 test title says a worker one protocol behind lacks device.add/driverDeviceId; with max now 30 that worker has both. False since feat(leasing): physical devices that go missing become absent, end their lease device-lost, and come back reclaimed #458.

Written by an agent.