Terminology consistency wave: corpus-splitting terms, the 边际/边缘 pattern residue, and mistransliterated researcher names

The 2026-08-20 full-corpus quality review (project-translation#50) found within- and cross-lecture terminology inconsistency is the edition's dominant editorial defect — present in roughly a third of lectures. This issue collects the wave; each ruling should also feed the engine glossary so syncs stop reintroducing the losing variant.

Corpus-splitting terms (rule one rendering each)

Worst within-lecture offenders (for the editor queue): mix_model (agent five ways: 类型1智能体/代理/学习者/第1类人/第1类主体), likelihood_ratio_process_2 (agent four ways across 116 occurrences), two_auctions (FPSB and SPSB each three ways, including the title), ge_arrow (Arrow securities as 阿罗证券/箭头证券/untranslated; 各州 for state-by-state).

Extend the #201 ruling from the term to the pattern

#201 eliminated 边际分布, but the same character confusion survives in wordings its sweep never matched — statistical marginal-X should be 边缘X: exchangeable (边际密度, systematic), likelihood_ratio_process_2 (×3), linear_models, mle (one each, both files also use the correct form nearby), phillips_drifts_volatilities (边际先验/边际后验). Economics marginal (成本/效用/倾向) stays 边际.

Four name mistransliterations (one-line fixes)

• Kiyotaki → 北原 (bounded_rationality; standard is 清泷)
• Chentsov → 陈氏 (blackwell_kihlstrom)
• Tsallis → 查理斯 (blackwell_kihlstrom, two docstrings)
• Zha → 扎 (phillips_drifts_volatilities; convention 查)

Plus unstable transliterations to normalise within single lectures: Wald (three forms in wald_friedman_2), Kolmogorov, Bewley, Raphson, Phelps 费/菲.

Also worth one policy ruling

Wikipedia links are silently swapped for Baidu Baike / Zhihu / CSDN targets in ≥8 lectures — plausibly deliberate accessibility localisation, but undeclared and inconsistent, and rand_resp's differential-privacy link points at an unrelated page. Rule it once, then sweep.

Every finding with quotes and locations: data/quality.json in the audit directory; the per-lecture table in the report is the priority queue.

Start with data/quality.json and the per-lecture table in the audit report, then inspect the named lectures such as mix_model, likelihood_ratio_process_2, kalman_2, and phillips_drifts_volatilities. Review ruling #201 before sweeping the listed terminology and names. Done means consistent renderings across the corpus, glossary updates that prevent regression, and a documented Wikipedia-link policy.