refactor: adopt Hard Signal Framework for create-expert verification

[FL4TLiN3]

Summary

• Replace soft signal verification (LLM judgment-based) with Hard Signal Framework: Ground truth × Context separation × Determinism
• Consolidate 3 test queries into 1 comprehensive query — coverage comes from signal design depth, not repetition
• Add reproducibility check to verify-test: re-run same commands, diff results to confirm signals are truly deterministic
• Simplify verdict system from PASS/SUFFICIENT/CONTINUE to PASS/CONTINUE

Test plan

• Run create-expert against a sample domain and verify plan.md produces hard signals (commands + expected outputs)
• Verify the build loop executes write → test → verify in single-query mode
• Confirm verify-test's reproducibility step catches non-deterministic signals
• Validate structural checks still detect missing delegates arrays, pick lists, etc.

🤖 Generated with Claude Code