[handrews]

"Collect the package logger's records for a block" was written five times: the loader's error counter, two handler classes in the scripts, one in the test fixtures and a stream handler in the test support. One collecting() in load.py yields a Collector with the records, the errors and the warning set; load counts from it, the claims script holds and replays through it (the one place that still swaps handlers), the expected-warnings script and the fixtures read it.

The per-node integrity checks were listed by hand in three places, so a check added to the loader could be forgotten by the draft checker. phylohist/loader/checks.py holds the file, material, nomenclature and draft tables over one per-node scope; load.py and check_draft.py iterate them, draft_nulls moves beside null_material, the two source_years become one rule, and a contract test holds the draft's tables to the loader's but for the three checks only a draft needs. The loader's output is unchanged; the drafts' is too.

Author keys were resolved by three loops that differed on et al. and on a capitalised surname; resolve_authors in research.py serves the authorities, the sources and the aliases with each one's own wording, and the warning for a capitalised author whose key form is registered now names the record it is on. implied_rank and
Authority.qualified_keys replace the three spellings of "species if lower-case, else genus" and the two key-expectation cores.

load.py and claims.py read the records through accessors (declared_rank, pages, illustrations, cited_as, designation, homonym, original_parent, lang, status on Taxon; author_keys and citation_fields on Source) in place of _data, and an alias shares its target's record outright. Gone, with no reader: Source.count and title, _editors, the month and day of a publication date, Tree.is_new, bracket_end and _type_index, Section.notes, Authority._ex and _disambiguator.

The claims, the names index and the expected warnings are untouched.