refactor: per-agent start lock — complete the concurrent-start race hardening from #1809

[webmixgamer]

Summary

#1809 shipped minimal recreate-race hardening: container_remove tolerates NotFound and a 409 name-conflict on the run adopts the winner's container. That converts the loser's 500 into a graceful success, but two concurrent starts can still both enter the recreate sequence (double stop/remove work, interleaved injections).

The full fix is a per-agent start lock serializing start_agent_internal, following the existing Redis SETNX+TTL precedent (agent:data_op:{name} in routers/agent_data.py).

Why deferred from #1809

New lock infrastructure on a hot path has its own failure modes (stuck lock blocks starts; TTL sizing; Redis-down behavior; wait-vs-409 contention semantics) — deliberately kept out of a P2 bug fix. See the #1809 plan's decision trail.

Design notes

• Key: agent:start:{name} (register in services/agent_runtime_state.py per the bug(lifecycle): stale transport circuit-breaker Redis key survives agent delete/recreate — a fresh healthy agent fast-fails as "unhealthy" #1560 keyspace-parity rule — tests/unit/test_1560_agent_redis_key_parity.py fails CI on unregistered agent:* keyspaces)
• Fail-open on Redis down (start must never be blocked by lock infrastructure)
• TTL ≥ worst-case recreate duration (stop timeout + run + readiness poll)
• Second caller: prefer short wait-then-adopt over an immediate 409 ("start" is idempotent from the caller's view)

Acceptance criteria

• Concurrent starts of the same agent serialize; exactly one performs a needed recreate
• Lock failure modes fail open (never blocks a start when Redis is down; no stuck-lock lockout beyond TTL)
• Keyspace registered in agent_runtime_state.py with parity-test coverage

Refs #1809

[webmixgamer]

Heads-up from #1860 (PR #1912): lifecycle.restart_agent_internal(name, *, stop_timeout=30) now exists as the canonical stop→cold-start helper, consumed by restart_fleet. Its docstring names this issue as the future home of the per-agent start lock — when the lock lands there, the remaining inline copies of the stop→start_agent_internal shape can consolidate onto it in the same pass: routers/system_agent.py (restart/reinitialize), routers/systems.py (restart_system loop), and services/subscription_auto_switch.py (which additionally holds its own #799 process-local lock and uses stop timeout 10 — parameterized via stop_timeout).