Evaluator /verify transient retry backoff (2, 5, 10 seconds) adds up to 17s before ERROR

[u2135]

Summary

On unhappy paths, POST .../queries/{id}/verify may return transient 429/5xx. The SDK retries with fixed sleeps 2s + 5s + 10s = 17s total before surfacing server failure as ERROR.

Current behavior (SDK)

• src/provably/handoff/evaluator.py:
  • _VERIFY_RETRY_BACKOFF = (2.0, 5.0, 10.0)
  • Transient statuses include 429, 500, 502, 503, 504

Impact

• Correct for resilience, but poor UX when the service is degraded (long tail to ERROR).
• Developer feedback: consider optimizing unhappy/healing paths.

Acceptance criteria

• Product decision: keep conservative backoff vs exponential backoff with jitter vs shorter caps for eval-only path.
• Document worst-case latency for evaluation in README or runbooks.
• Optional: env-tunable backoff for operators.

References

• src/provably/handoff/evaluator.py — _verify_query_record, _VERIFY_RETRY_BACKOFF