fix(codex): keep commentary out of visible content

[callexcapital]

Summary

• do not promote Codex Responses phase=commentary/analysis message text into assistant.content
• preserve commentary/analysis message items only as Codex replay/debug metadata while still executing structured tool calls
• sanitize interim assistant commentary before gateway/UI callbacks
• add regression coverage for the Discord-visible Need inspect... commentary leak with a real tool call

Related

• complements fix: strip leaked memory context from commentary #12237, which strips leaked memory-context fences from interim/stream paths
• covers the separate failure mode where plain Codex commentary text is copied into visible assistant content during tool-call turns

Test plan

• pytest tests/run_agent/test_run_agent_codex_responses.py -q -o 'addopts=' → 60 passed
• pytest tests/run_agent/test_deepseek_reasoning_content_echo.py tests/gateway/test_session_boundary_security_state.py -q -o 'addopts=' → 41 passed

[teknium1]

The Codex commentary-phase leak this PR targeted is now fixed on main via PR #58868, which salvaged #58696 (broadest coverage: adapter + streaming + fallback) combined with #41343's reasoning-channel routing. Commentary/analysis-phase text now displays like thinking instead of leaking as visible content, in both normalized responses and live stream deltas.

You were the earliest submitter in this cluster — thank you for identifying and fixing the leak first. Closing since the fix that landed takes a different mechanism (phase-based routing at the adapter/runtime layer rather than callback sanitization).