test(sleep): strengthen verifier-discipline test with score/gate_action assertions

[harshitnagar22]

Summary

Follow-up to #94. Strengthens TestVerifierDiscipline in tests/test_sleep_engine.py with three concrete improvements:

Changes

1. test_gate_rejects_reward_hacking_edit (enhanced) — now asserts concrete baseline_score=1.0 and candidate_score=0.0 so the rejection is verifiably score-driven, plus asserts gate_action="reject" to bring the vendored evaluate_gate() under test.

2. test_gate_accepts_genuine_improvement (new) — paired accept case using MockImprovingBackend: an edit that genuinely improves the held-out slice. Asserts accepted=True, baseline_score=0.0, candidate_score=1.0, gate_action="accept_new_best".

3. test_gate_rejects_subtle_reward_hack (new) — a more realistic reward-hack scenario using MockSubtleHackBackend: train improves from 0.0→1.0 (looks good!) but held-out degrades from 1.0→0.5. Gate correctly rejects, proving it catches non-obvious regressions, not just all-or-nothing collapses.

Why

Verifier discipline (only adopt edits that improve a held-out slice) is the core safety property of the sleep loop. Pinning scores + asserting gate_action turns the test from a smoke test into a real regression guard against gate regressions. The paired accept/reject cases ensure the gate works in both directions.

[harshitnagar22]

@microsoft-github-policy-service agree