Idempotency and ergonomics sweep for the stable core (addresses #40)

[larnsce]

Systemic rule from the 2026-07 review: every function that rewrites user-owned files (DESCRIPTION, README.Rmd, R/*.R docs, _pkgdown.yml, .gitignore, CITATION.cff) follows read-merge-write and is safe to re-run, because the openwashdata workflow explicitly re-runs these functions on packages people have hand-edited.

Tasks:

• setup_website(): idempotent handling of an existing _pkgdown.yml so the guide's "answer No when prompted to preserve styling" instruction becomes unnecessary (a step whose docs warn you to answer a prompt correctly should not prompt)
• setup_website(): do not crash when .gitignore is missing (R/setup_website.R:52-55 readLines errors on a fresh non-git package)
• inst/templates/_pkgdown.yml: adopt pkgreview's corrected copy in full, not just the url field: Pages URL instead of repo URL, the explanatory comments, AND the multi-dataset reference block (skills/pkgreview-core/references/templates/_pkgdown.yml in openwashdata/pkgreview). Goal: pkgreview's fork reduces to a pointer (its CLAUDE.md rule 4 calls the divergence a defect magnet; cross-repo coherence finding C4)
• update_citation(): preserve hand-added CITATION.cff fields across regeneration, keywords in particular. pkgreview's standard adds advisory CITATION.cff keywords (Provenance and FAIR light: @source, README provenance and downloads, CITATION.cff keywords, Zenodo record prompt pkgreview#35) and already warns they must be re-verified after every update_citation() call because cffr regenerates the file wholesale; merge instead of wiping so that manual step disappears (cross-repo coherence finding B1; the canonical keyword mapping is settled in Metadata design decision: canonical sources and field mappings (resolves #47) #67)
• setup_roxygen() re-run: preserve user-written roxygen below the @Format block instead of silently deleting it, and error clearly when an existing file has no @Format line (currently an unbounded while loop crash); closes the actionable half of Add possibility to overwrite existing .R roxygen documentation file using setup_roxygen() #40, whose documentation half is already largely addressed
• setup_dictionary(): correct variable_type for multi-class columns (POSIXct, ordered factors produce garbage like c("POSIXct", "POSIXt") today); groundwork for Feature request for labelled data #20
• load_object(): error informatively on multi-object .rda files and non-rda files in data/ instead of documenting the alphabetically first object or crashing opaquely

Acceptance: running the full core sequence twice in a row on a fixture package produces zero diff and zero data loss; a CITATION.cff with hand-added keywords keeps them through update_citation().

Milestone: v1.1.0 (set manually, see the milestone setup issue).

[larnsce]

Cross-references from the triage in #113 (2026-09-02). #104 folds into this sweep: setup_website() skips the .gitignore edit when .github/workflows/pkgdown.yaml exists, or takes a track_docs argument defaulting to the current behaviour; three lines next to the idempotent _pkgdown.yml task already here. #102 (CITATION.cff in .Rbuildignore) shipped separately in 0dc3ae1 through #114, so the update_citation() task here is only the keyword preservation. Under decision H in #111, this issue is a prerequisite for the washr skill (#115), because the skill re-runs these functions after editing files.

[larnsce]

Done on dev in 0a28d41 (2026-09-02). Not yet on main.

• setup_website(): an existing _pkgdown.yml is kept as it is and only the site is rebuilt (no use_pkgdown() call, so no overwrite prompt); a missing .gitignore is left alone; the example article is created once; and docs stays ignored when .github/workflows/pkgdown.yaml exists or track_docs = FALSE is passed (setup_website() re-tracks docs/ even when deployment is via GitHub Actions #104).
• inst/templates/_pkgdown.yml: the Pages URL as url, the explanatory comments, and a reference block with one entry per data object in data/, rendered by whisker at setup time (omitted when data/ is empty, so the build does not stop on an empty contents list). Content matches the review standard's template with openwashdata where it has ORGNAME and ORGDOMAIN; the pkgreview copy can now be a pointer here plus those two placeholders. use_brand() still rewrites the file through yaml, which drops the comments; that is documented on use_brand().
• update_citation(): a re-run without doi keeps the DOI on file (before, it silently dropped it while the README kept the badge); keywords typed into CITATION.cff by hand move once to X-schema.org-keywords in DESCRIPTION, their canonical home per Metadata design decision: canonical sources and field mappings (resolves #47) #67, and cffr carries them forward from there (cffr pads a single keyword with r-package to satisfy the CFF schema); the README rebuild runs only when the badge is missing.
• setup_roxygen(): only the @format block is regenerated on re-run; the title and description above it and everything below it (@source, @examples, other text) are kept; nothing is written when the output is unchanged; a file without an @format line errors with the file name instead of crashing in the loop.
• setup_dictionary(): variable_type is the first class of a column (POSIXct, ordered, Date), no more deparsed vectors.
• load_object(): a multi-object .rda or a non .rda file in data/ errors with the file name and the reason.

Acceptance: test_idempotency.R runs the re-runnable sequence (setup_rawdata, setup_roxygen, update_description, update_citation, update_metadata) twice on a two dataset fixture and compares the md5 of every file outside docs/; a second test shows setup_dictionary() and setup_readme() refuse to overwrite without loss. setup_dictionary() on re-run still stops rather than merging; the merge is #13. 36 new tests, suite 152 passed, R CMD check clean.