feat(sandcastle): enrich planner with acceptance criteria and root cause hypothesis

feat(sandcastle): enrich planner with acceptance criteria and root cause hypothesis

The planner now produces structured analysis per issue: issueType,
confidence, rootCauseHypothesis, and acceptanceCriteria. These flow
into the actor prompt (confidence-gated hypothesis + criteria) and the
critic prompt (criteria as verification checklist).

- Confidence controls plan specificity: high → full context, medium/low → criteria only
- All planner-generated fields are sanitized and length-bounded
- Critic evaluates observable outcomes, never plan adherence
- Backward-compatible: missing fields result in empty template variables