Split prompt-visible validation from target graders

[Jordak]

Summary

Previously, task.test served two roles at once: the runner used it as the deterministic target-grader command list, and build_agent_prompt showed those same commands to the agent harness as validation guidance. That made target graders prompt-visible by default.

This changes the task contract so test: remains grader-facing and prompt-visible validation must be configured explicitly with visible_validation:. Codex/Claude prompts and the manual adapter now show only visible_validation commands. Task cards can render visible validation separately for maintainer review, while existing target grader reporting and runner behavior remain unchanged.

Tracked issue: GitHub issue #118 remains open for the landing workflow.
Related blocker: GitHub issue #91 remains blocked until this prompt-policy split lands.

Readiness evidence: GitHub issue #118
Readiness verdict: Ready to Implement

Validation

• python3 -m unittest tests.test_agent_prompts tests.test_tasks tests.test_task_cards
• python3 -m unittest tests.test_task_bundle_integrity tests.test_runner tests.test_results tests.test_task_execution
• python3 -m unittest discover
• python3 -m agentlab task validate --check-task-cards tasks
• git diff --check
• pre-commit task bundle integrity check during commit