[handrews]

The loader kept its records in class-level registries (Author._authors, Source._sources, Taxon._taxa, Section._sections and the rest), an index every Tree root wrote (_new_index), a repository registry in material.py and a rank-hub cache in claims.py that went stale when a taxon was added. load() filled them as a side effect and returned (data, roots); a second load in one process overwrote and leaked; the extractor and the claims script read the private dicts; and the tests snapshotted and restored all of them by hand before every test.

phylohist/loader/corpus.py holds the Corpus: the five record dicts in insertion order, the repositories, the raw data and the roots, with add_* methods that build a record against the corpus and then store it (a constructor that ends the load stores nothing), copy() for a test's own view, and protologue_sources() walked from the roots in place of the index. Every constructor takes the corpus it resolves records in; a root Tree takes it beside its metadata and its nodes share it; no record keeps a reference to its corpus, and Source.publication is resolved where the record is built. load() builds a fresh corpus per call and returns it; the loader's own checks, the extractor (extract, manifest, names_index, rank_variants, holotype_conflicts) and the claims script take it as their first argument, and claims.py imports no record class and keeps nothing between calls.

The tests take records, a per-test copy of the session corpus, in place of the restored registries; support.with_documents gives a loader check its synthetic documents, and an autouse guard fails a test that changes the session corpus. The three autouse registry fixtures and the snapshot machinery are gone.

The claims, the names index, the manifest and the repositories file are byte-identical, as is the loader's log over the data and drafts.