Job Monitor: extract a pure retry reconciliation planner

[mmitche]

Parent: #17171

Make retry decisions independently understandable and testable.

Scope

• Given an immutable Helix observation and invocation context, produce explicit actions such as leave running, resubmit selected items, record an actionable failure, or take no action.
• Keep API calls, logging, and mutable counters in an executor/orchestrator layer.
• Preserve one-shot-on-entry behavior and stage-attempt reconciliation invariants.
• Represent resubmission failure explicitly rather than with ambiguous nulls where practical.

Acceptance criteria

• Retry policy is expressed by a deterministic pure component.
• Existing retry, crash-recovery, and multi-attempt scenarios remain unchanged.
• A reviewer can understand retry behavior without following service calls.