[codex] fix(agent): preserve Codex recovered stream usage

[xyang199]

Summary

• preserve terminal response.completed.response.usage when recovering Codex Responses streams after response.output = null
• carry recovered usage through both the main Codex stream path and auxiliary Codex adapter
• add regressions that reproduce iterator-time SDK TypeError recovery while asserting exact usage is retained

Root Cause

#32963 recovers already-streamed output when the OpenAI SDK raises TypeError: 'NoneType' object is not iterable while parsing a terminal Codex Responses event with output = null.

That terminal event can still contain accurate usage, but the synthesized recovered response used usage=None, so token accounting state such as context_compressor.last_prompt_tokens, session counters, and footer/context display could remain at 0.

Validation

python -m pytest tests/run_agent/test_run_agent_codex_responses.py::test_run_codex_stream_falls_back_when_stream_iteration_parses_null_output tests/agent/test_auxiliary_client.py::TestCodexAuxiliaryAdapterNullOutputRecovery::test_recovers_output_item_when_sdk_raises_during_iteration -q
# 2 passed

python -m pytest tests/run_agent/test_run_agent_codex_responses.py tests/agent/test_auxiliary_client.py -q
# 243 passed

Fixes #33019

[alt-glitch]

Competing with #33017 — both preserve Codex recovered stream usage from null-output recovery (#32963). Both fix #33019. Recommend consolidating into a single PR.

[teknium1]

Closing as obsolete. PR #33042 (merged commit cb38ce28c) removes the OpenAI SDK's client.responses.stream(...) helper from both Codex call sites entirely — we now use client.responses.create(stream=True) raw event iteration that assembles the final response from response.output_item.done events as they arrive. The terminal event's output field is never read for content reconstruction, so the SDK's TypeError: 'NoneType' object is not iterable parser crash on output=null is structurally impossible — no code path can produce it. Same defensive strategy OpenClaw uses for the same backend.

The symptom-patch + recovery logic this PR added on top of the SDK helper is now redundant. Thanks for the work during the outage.