Make the daily soak exercise real workflows and report useful results

[rmcdaniel]

Related #137. Fixes #151.

Synthetic workers now maintain their registrations, reject stale HTTP-200 poll responses, and must sustain cache activity across the measured window. A low-rate PHP canary reuses the standard-workflow binding and validates results, completed status, and ordered activity history. Health/readiness/cluster-info probes always run. CPU/memory, latency percentiles, backpressure, cache drain, and explicit coverage limits appear in the Action summary; artifacts expire after 14 days.

Docker supervises the existing hourly queue-worker recycle. Real activity replay caches retain evidence through the five-minute lease plus grace, so the drain is 310 seconds without flushing cache or reducing Server TTLs. Load remains two hours on the same disposable host, with bounded drain/setup/cleanup time additional. Idle drain does not dilute load memory-slope or CPU measurements.

Security: provider credentials remain in the protected main-only environment, never on the VM. PR checks receive no provider credentials. The SDK uses only isolated disposable data and credentials. No live deployment or Cloud/customer data is involved.

Validation:

• 23 Python tests and 28 PHP contract tests (446 assertions) pass; PHP formatting passes.
• Native isolated Compose run at 0997b9c: 90 seconds, 24 synthetic workers, eight namespaces, 16 queues; 18/18 validated standard workflows, zero execution/result/history errors, p95 3.837s.
• Polling activity in all 18 periodic samples, zero polling errors, healthy health/readiness/cluster-info probes, final Server cache keys zero after the real TTL drain.
• Native Laravel queue:restart during load caused Docker restart count to advance from zero to one; polling and workflow completions continued. The ordinary max-time exit was separately observed to return zero.
• The final measurement-only change at 882628f has focused regression coverage excluding idle drain from load CPU/slope; it does not change Server semantics.
• All task containers and volumes were removed. No additional paid VM or paid rerun was created.

This short run proves the changed harness and recycle mechanism, not a completed two-hour soak, capacity claim, cross-SDK qualification, backend failover, or completion of the broader runtime comparison. The next scheduled paid run must supply the long-duration evidence.