[codex] Calibrate identity self-ID patterns with ZenMux data

[toby-bridges]

Summary

• Calibrates Step 5 natural-language identity matching against ZenMux Arena's 2026-06-01 "Who Are You?" dataset.
• Adds current frontier-vendor aliases while keeping common names anchor-gated so product comparisons do not become identity findings.
• Fixes CJK no-whitespace matching for distinctive ASCII model names such as 我是DeepSeek.
• Regenerates the standalone audit.py artifact and documents why self-ID mismatches remain consistency signals, not attribution proof.

Validation

• python -m pytest (783 passed)
• Local ZenMux calibration rerun over 29,700 joined records: extractor non-Anthropic identity coverage improved from 49.2% to 62.1%; Anthropic-claim false flags stayed low at 0.2%.