Phase G: verifier-derived training rewards with kernel-observed divergence penalty

[txmed82]

What

• Wires kernel-observed episode divergence (diverged_observed) into export_verifiers.py:rollout_vector context during training rollouts.
• Adopts context.get("diverged_observed") in LumenVerifier and wrap.py verifier generator: when engine info lacks a divergence channel, non-finite tags observed across full episode steps correctly mark diverged = True.
• Evaluates diverged episodes as failed gates and sets the diverged metric to True, driving declarative projection to 0.0 without manual reward shaping.
• Adds regression unit test in tests/test_eval_run.py covering verifier divergence gate failure under diverged_observed.
• Updates docs/NEXT_STATUS.md ledger reflecting D5 cloud parity (feat: Phase D5 explicit cohort subgroup validation, head-covered trial bindings, and clustered continuous bootstrap #64, vector-cloud Phase 5: decision rule and contestation #5) and Phase G training reward adoption.

Why

Phase G requires that verifier-derived training rewards reflect authoritative vectors rather than fabricated physics or unobserved convergence. Wiring kernel-observed divergence ensures diverged solver rollouts penalize training rewards directly.

Verify

• uv run ruff check --fix && uv run ruff format --check && uv run mypy clean.
• Evaluation test suites green (65 passed).

[txmed82]

@droid review

[factory-droid]

Droid finished @txmed82's task —— View job

---

No inline comments. The divergence-observed wiring looks consistent across export_verifiers, wrap verifiers, and the Lumen example, and the added regression test covers the intended behavior.

[factory-droid]

Droid finished @txmed82's task —— View job

---

No inline review comments to add. The changes look consistent with wiring kernel-observed divergence into verifier scoring so diverged rollouts fail gating as intended.