fix: gate background loops behind enable_background_workers

[nishant-uxs]

Problem

lifespan starts the scheduler, dispatcher, and watchdog unconditionally. Uvicorn runs lifespan once per worker process, so multi-worker / multi-replica deployments run replicas × --workers copies of each loop against the same DB. Watchdog copies fan out sandbox-API verifies (contributing to HTTP 429 → false FAILED, see #285) and put background CPU on webhook-serving processes.

Root cause

Ownership of background work was tied to process startup, not to a deployment-level singleton. There was no way for a request-serving worker to skip the loops.

Solution

Implement the preferred fix from #286:

• New setting AUTOMATION_ENABLE_BACKGROUND_WORKERS (default true)
• When false, this process starts zero of scheduler / dispatcher / watchdog
• When true, starts exactly one of each (prior single-process behavior)

Extracted start_background_worker_tasks() so lifespan shutdown only awaits tasks that were actually started.

Operator model

Process role	Setting
Local / single-process	default true (no change)
Dedicated background owner	true
Request-serving replicas	false

This is an explicit single-owner gate, not leader election — enabling the flag on every worker still fans out (covered by a regression test so the invariant stays documented).

Why this approach

• Matches the issue author's preferred fix (vs. claim-before-verify, which only softens watchdog fan-out)
• No public API change; default preserves local/dev behavior
• Minimal diff — no Temporal rewrite, no distributed lock redesign

Risks

• Multi-worker production deploys must set the env var correctly on request servers; leaving the default on every worker retains today's duplication
• Helm/compose charts are out of scope for this PR (config knob only)

Testing

• uv run pytest tests/test_background_workers.py tests/test_config.py tests/test_health.py — 57 passed
• Confirmed fail-before-fix: with the gate forced off, test_disabled_* and test_multi_worker_simulation_single_owner fail (3 tasks started when disabled)
• ruff check / ruff format clean on touched files
• pyright — 0 errors on app.py / config.py

Regression coverage

• Disabled → no scheduler/dispatcher/watchdog
• Enabled → exactly one of each
• Multi-worker simulation (2 request + 1 owner) → one of each across the deployment
• All workers enabled → still 3×3 (documents operator invariant)
• Default true; env false respected

Fixes #286