Add reliable detection job dispatch

[fengting124]

What

• persist asynchronous detection commands in a PostgreSQL transactional outbox
• dispatch versioned events to Redis Streams with retry, stale-claim recovery, and stable event-id deduplication
• add bounded inspection and replay operations plus an operator runbook

Why

The previous request path performed a database/Redis dual write. Redis outages or a process crash could lose accepted work. This change makes PostgreSQL the durable source of truth and provides an explicit at-least-once delivery contract without requiring model weights.

Reliability boundary

• no distributed transaction or exactly-once claim
• normal duplicate publication is suppressed and completed tasks are terminal-state idempotent
• atomic ownership for independently injected concurrent task messages is deferred to feature/short-lived-execution-transactions
• real PostgreSQL/Redis restart tests remain deferred until Docker is available

Verification

• Java: 79 tests passed
• frontend: 8 tests, lint, and production build passed
• model service: 6 tests passed without weights
• smoke tests: 3 passed