Align df.metrics() instance status counts with persisted df.instances (exclude rollback orphans)

[Copilot]

df.metrics() could report higher failed_instances than df.instances/df.list_instances('failed') after rolled-back df.start() calls, because failures from orphan duroxide.executions rows were being counted as workflow failures. This changes metrics semantics to reflect persisted workflow instances for status counts.

• Metrics source of truth

  • total_instances, running_instances, completed_instances, failed_instances now come from df.instances (persisted workflows).
  • total_executions and total_events continue to come from duroxide.executions / duroxide.history.

• Query shape

  • Replaced Client::get_system_metrics() usage in df.metrics() with explicit SQL aggregation.
  • Uses a single scan over df.instances with FILTER (...) counts and cross-joins execution/event totals.

• Regression coverage

  • Added an E2E regression in tests/e2e/sql/05_monitoring_and_explain.sql:
    • triggers a rolled-back df.start(),
    • waits for worker-side orphan failure effects to settle,
    • asserts df.metrics() instance counters match df.instances counters.

SELECT
  COUNT(*)::BIGINT AS total_instances,
  COUNT(*) FILTER (WHERE status = 'running')::BIGINT AS running_instances,
  COUNT(*) FILTER (WHERE status = 'completed')::BIGINT AS completed_instances,
  COUNT(*) FILTER (WHERE status = 'failed')::BIGINT AS failed_instances
FROM df.instances;