[codex] Fix Chinese name order and splitting

[sergeyf]

Summary

• Preserve explicit all-Han group boundaries for spaced Chinese input such as 增友 叶.
• Add a guarded given-first order bonus for two-token romanized names where surname-frequency evidence strongly supports preserving input order.
• Add a given-position-only gold splitter for surname-like fused tokens, fixing Junjie Fang-style outputs without splitting ambiguous one-anchor tokens such as Yuxuan or protected compound-surname variants.
• Add regression coverage for the new order, splitting, and all-Han grouped-input behavior.

Root Cause

The parser was collapsing whitespace-separated all-Han groups into character-level tokens before deciding order, so 增友 叶 lost the original given/surname boundary. Separately, two-token romanized names could be flipped even when the second token was overwhelmingly more likely to be the surname. Finally, the fused-token splitter had a broad surname early-exit, so valid given-name splits like Jun + Jie were skipped when the full token also appeared as a surname form.

Validation

• uv run --no-sync --python 3.12 pytest -q tests/test_regression_proposals.py::test_given_context_gold_split_bypasses_surname_guard tests/test_regression_proposals.py::test_given_context_gold_split_keeps_ambiguous_non_gold_tokens_unsplit tests/test_regression_proposals.py::test_guarded_low_frequency_surname_ratio_preserves_given_first_order tests/test_regression_proposals.py::test_guarded_low_frequency_surname_ratio_keeps_compound_and_common_surname_boundaries tests/test_regression_proposals.py::test_spaced_all_chinese_given_first_preserves_group_boundary
• uv run --no-sync --python 3.12 python scripts/check_test_status.py

Special status script result: 44 individual test-case failures, performance tests passed. This improves from the previous expected baseline of 52 failures.