Signal extraction is not diagnosis: a rule-based trait pipeline on 4,042 Chinese-Wikipedia biographies reveals a steep recall-precision cliff and the necessity of biography-grade evidence for historiometric neurodivergence research

Companion to the native-language eminence-calibration study (P12, DOI 10.5281/zenodo.23621482). We parsed the full 3.43GB Chinese-Wikipedia wikitext (2026-10-01), extracted complete articles for 9,603 targets, and built a six-domain trait dictionary in two iterations. A broad dictionary fired for 3,105 people but was dominated by encyclopaedic achievement words; subject-attribution filtering, removal of high-frequency achievement words, fictional exclusion, and institutional vetoes reduced the firing set to 564 and those with >=1 hard idiosyncratic signal to 478. Of 180 real-person cards (zero overlap with the 5,631-card database), 144 were adjudicated at full-article level: only 1 (Sandor Petofi) reached a stable Class-3 attribution (0.7%), 39 (27.1%) were defensible multi-domain leads, 104 (72.2%) collapsed to weak/contextual signals, and 1 (F. Macias Nguema) was a genuine psychopathology exclusion. Three structural false-positive classes were identified and removed: kin pathology (Ginsberg's mother), politically imposed psychiatry (Soviet dissident Bukovsky), and political/career rather than scholastic expulsion (party purges). A prodigy with six precocity hits (Chopin) was retained as grey because prodigy is not autism without social/restricted-behaviour evidence. We conclude that a rule-based native-language pipeline is an excellent triage/ranking instrument but a poor diagnostic instrument, and that biography-grade evidence is mandatory for the final attribution. Dictionary versions, filters, adjudication rules, all evidence sentences, and scripts are released for reproducibility.