Build hardened metadata parsers with fixtures

[larnsce]

Goal

Hardened parsers for the per-package metadata files, with a fixture suite covering every known pathology across the org. They live in R/parsers.R, take text already fetched by a backend, never touch the network, and are unit tested against tests/testthat/fixtures/.

Rewritten 2026-09-04 per dev/decisions-2026-09.md. Start from R/parsers.R and the fixtures on the closed PR #15 branch.

Inputs to parse

1. DESCRIPTION via desc::desc(text = ): Package, Title, Description, Version, Authors@R (maintainer, ORCID, all aut/cre), License, Date, URL, and the washr 1.1.0 fields X-schema.org-keywords, X-schema.org-spatialCoverage, X-schema.org-temporalCoverage
2. data-raw/dictionary.csv: the five standard columns; washr's own reader requires only file_name, variable_name and description, so keep that minimum
3. CITATION.cff via cffr::cff_read() on a temp file: doi and version
4. Roxygen data docs in R/: dataset title, description, and dimensions from the @format line
5. The file listing of inst/extdata, for the csv and xlsx URLs in Define and ship the catalog data model #5

Known dictionary.csv pathologies

Fully quoted headers, space-padded names, trailing empty columns, extra columns (unit_type, error), a leading unnamed column, double-double-quoted headers, the deprecated two-column schema with a UTF-8 BOM, a git-lfs pointer file, latin1 encoding, and multi-class variable_type written as a deparsed vector such as c("POSIXct", "POSIXt") by washr releases before 1.1.0.

Parser behaviour: read as text, detect git-lfs pointers and refetch through the media endpoint in the GitHub backend, convert latin1 to UTF-8, strip the BOM, normalise header names, drop unnamed and empty columns, keep only the five standard columns, map the two-column schema to the package's single dataset when it has exactly one, normalise variable_type to its first class. Unrecoverable failure lands in the problems table.

Known @Format variants

"A tibble with N rows and M variables", a trailing colon after the tag, #'@format without a space, reversed order with a thousands separator. Anything unrecognised returns NA dimensions.

Acceptance criteria

• Every fixture parses to the normalised schema or lands in the problems table with a reason
• No parser call can abort a harvest run
• A DESCRIPTION written by washr::update_description() and a CITATION.cff written by washr::update_citation() round-trip without a problems row
• Non-ASCII text survives: a fixture with ö and é parses to the same characters, never to <U+00F6>