[BUG] Codex WebSocket session metrics lose model attribution after model switch #3027

Description

When a Codex conversation switches models while keeping the same WebSocket connection open, Headroom can record the response metrics under unknown instead of the model that handled each turn. This makes per-model usage, savings, and cost reports inaccurate.

Reproduce Steps

1. Run the OpenAI Responses WebSocket proxy.
2. Keep one WebSocket session open.
3. Send a turn using model-a and complete it.
4. Send another turn on the same session using model-b and complete it.
5. Inspect the recorded request metrics.

Minimal client frames:

{"type":"response.create","response":{"model":"model-a","input":"first turn"}}
{"type":"response.create","response":{"model":"model-b","input":"second turn"}}

Expected Behavior

The recorded models should be model-a, then model-b.

Actual Behavior

Before the fix, the same session recorded unknown, unknown. The failure was reproduced against an archived copy of main.

Code Sample

response.completed.response.model = model-a
response.completed.response.model = model-b
recorded metrics = [model-a, model-b]

Error Output

No exception is raised. The failure is incorrect telemetry attribution:

assert ['unknown', 'unknown'] == ['model-a', 'model-b']

Environment

• Headroom version: source checkout at the pre-fix main revision
• Python version: 3.13.3
• OS: Windows (win32)
• LLM Provider: OpenAI Codex Responses WebSocket

Additional Context

Codex can keep one connection open for a multi-turn conversation while the user changes models. Each completed turn should be attributed to the model that handled it so the dashboard and reports remain accurate across model changes.