fix(ai): keep Codex execution content advisory

[Pigbibi]

Summary

• Successful Codex execution does not independently verify its financial content. Preserve it as advisory/display-only.
• Reuse the existing advisory boundary to prevent shadow feedback writes without changing deterministic routes or order/allocation restrictions.
• Keep API analysis status handling unchanged; no SDK/model extensions, pins, switches, dependencies, or deployment changes.

Verification

• Tests-first public crisis/TACO regression: original code returned ok (2 failed), then passed after the minimal correction.
• Tests exercise the real public consumers and Codex completion path with only the SDK boundary stubbed. Model self-claims cannot change status or permissions, no API fallback occurs, and feedback writes remain zero.
• Focused: 53 passed. Full suite: 244 passed. Full Ruff and git diff checks passed.
• No real model/API/provider calls, financial artifacts, deployment, or runtime operations.