[thystra]

Summary

• enable five delayed retries for every relay fan-out target;
• retain the shared activity body for fifteen minutes;
• reduce remain_count only on success or final retry exhaustion;
• add bounded structured delivery-attempt correlation and remote-error details;
• add focused regression tests for retry retention, successful recovery, terminal
  cleanup, retry horizon, metadata parsing, and bounded response capture.

RC1 finding corrected

RC1 production diagnostics showed that some receivers synchronously fetch
actors, keys, objects, parents, contexts, or collections before returning from
/inbox. Transient five-second delivery timeouts were terminal because
relay-v2 tasks had RetryCount: 0. The shared payload also expired after two
minutes and was decremented after each failed attempt.

RC2 keeps the existing five-second per-attempt HTTP circuit breaker but retries
the target at bounded Fibonacci delays. The delivery model remains at-least-once.

Validation

• full serialized Go tests;
• full serialized race tests;
• go vet ./...;
• web tests;
• ops tests when present;
• disposable Redis only;
• clean build and version verification;
• git diff --check;
• private-host and private-key screening.

No release tag, package publication, or deployment is performed by this PR.