Abstract
Historical music dictionaries become digital heritage objects through a chain of editorial, technical, and interpretive transformations. A scan, an OCR file, a segmented entry, a keyword rule, and a plotted category count each preserve something from the historical source, but each also changes what can be found, compared, and cited. I examine that process through Lexicon C19 v3.2.1, a reproducible corpus of eight long nineteenth-century music, dance, biographical, and general reference traditions. The corpus contains 3,760 rule-selected records and a complete pool of 25,186 parsed candidates, with processed witnesses in English, French, German, Russian, and Portuguese; Russian and Portuguese are represented by fewer than 100 selected records. A provenance-first analysis combines source composition, selection yield, category profiles, OCR-surface regularity, exact headword overlap, a 200-record single-annotator audit, and a candidate-level rule-collision diagnostic. Three English-language processed witnesses supply 81.9% of selected records. Only 203 selected records, or 5.4%, match terms from more than one top-level category, which bounds the direct effect of rule precedence on top-level labels and shifts attention to larger asymmetries of sampling frame and ontology coverage. The audit estimates selected-candidate precision at 97.0%. Among real and relevant sampled rows, broad-category self-consistency is 89.7%. These figures are not historical prevalence. They show how sampling frame, translation, typography, OCR, segmentation, and classification rules construct the musical-heritage object available to computation. I argue for a critical computational musicology in which recoverability is modelled as evidence and distant reading organises source-specific close reading rather than replacing it.