Run the fixture review: acceptance gate for PR #14

[larnsce]

Goal

Run the full review workflow against the defective fixture package and confirm every planted defect is caught. This is the acceptance gate for the checklist consolidation (#5) and the skill refactor (#6), and the last unticked box on PR #14. It doubles as the observed "first skill-driven review" for the rollout gate in #8.

Procedure

1. Stand up a throwaway repo (the fixture cannot be reviewed in place; gh would create issues on this repo):

   cp -R fixtures/pkgreviewtest ~/tmp/pkgreviewtest
   cd ~/tmp/pkgreviewtest
   git init -b main && git add -A && git commit -m "Fixture package for review workflow test"
   gh repo create larnsce/pkgreviewtest --private --source . --push

2. In a new Claude Code session in that directory, run the full sequence: /review-package, then per area /review-issue [number], merge the PR to dev, /create-next-issue, and finally /review-complete.

3. Score against fixtures/SCORECARD.md. A defect counts as caught when the analysis or PR for its area names it:

   • Metadata: D1 (MIT license), D2 (wrong CITATION author)
   • Data: D3 (latin1 encoding), D4 (-99 sentinels), D8 (character dates), D9 (inconsistent status values), D10 (duplicate id), D11 (camelCase column), D12 (dead code in data_processing.R)
   • Docs: D6 (vignette outside articles/), D7 (missing Plausible header)
   • Tests: D5 (no R-CMD-check workflow or badge)

4. Watch the guardrails throughout: every check-in honored, hard STOP after each PR, no PR against main before /review-complete. One violation trips the rollout gate (Keep behavioral guardrails always-in-context and add mechanical enforcement #8).

5. Afterward: gh repo delete larnsce/pkgreviewtest --yes and remove ~/tmp/pkgreviewtest.

Expected checks during the run

• Standards file written as CLAUDE.md with Standard version: 1.0.0-dev filled in
• Version stamp line present in the first review issue body
• All 12 scorecard defects caught in their respective areas
• No guardrail violations (check-ins, STOPs, PR bases)
• Result recorded here and the last box ticked on PR Refactor review workflow into Claude Code skills (issues #4-#11) #14

On failure

A missed defect means the consolidated checklist or a skill weakened the standard; report which defect and hold PR #14 until fixed. A guardrail violation halts rollout until enforcement is mechanical (#8).

Related: #5, #6, #8, #9, PR #14.

[larnsce]

Status: the v1.0.0 release (PR #26, including the amendments from the fslogisticskampala review) was shipped with this gate waived by the maintainer. The scorecard mapping was re-checked statically (every planted defect still maps to a surviving checklist item after the wording changes), but the full run against fixtures/pkgreviewtest is still outstanding and now validates v1.0.0 as released.

[larnsce]

Procedure refresh after v1.5.0 (2026-09-05). The body above predates the org profile layer, the 17-defect scorecard, and the check script; these are the differences that matter for the run.

• The throwaway repository must live in a registered organization. /review-package Step 0 resolves the org from the pasted URL and stops on an unregistered one, so larnsce/pkgreviewtest cannot be reviewed (observed today, recorded on Release v1.5.0: reconcile the standard with washr 1.1.0 and the owdata catalog #68). Use openwashdata/pkgreviewtest (private; the fixture DESCRIPTION already points there).
• The standard is 1.5.0: the standards file reads Standard version: 1.5.0 and Organization profile: openwashdata, and the first issue carries both stamps.
• Scoring by area, current mapping: Metadata D1, D2; Data D3, D4, D8, D9, D10, D11, D12, D13, D14, D15, D16, D17; Docs D6, D7; Tests D5. D18 is exercised by the history fixture, not this run. D13 and D14 stop the run at the intake screen; the maintainer acknowledges at the check-in and the review proceeds (scorecard, "How to use").
• Already observed (Release v1.5.0: reconcile the standard with washr 1.1.0 and the owdata catalog #68): Steps 0 to 2 of /review-package stop on D14 and D13 and create nothing; the mechanical report on the throwaway reproduces 19 FAIL + 2 FLAG. What this issue still needs is the interactive remainder: issue creation after the acknowledgement, four /review-issue rounds with their plan approvals and PRs into dev, the merges, /create-next-issue three times, /review-complete, and the guardrail observation throughout.
• New since v1.5.0 to watch: the docs issue creates .github/workflows/pkgdown.yaml and untracks docs/ (the fixture has neither); /review-complete Step 3 verifies it. The throwaway is private, so GitHub Pages cannot be enabled there; the final PR body names the Pages setting as the maintainer action, and that line counts as met for the gate.
• The check script report is posted as the first comment on issue 1 and again after each /review-issue round; the per-area FAIL lines are the verified facts the plans start from.

The throwaway is standing: local copy at ~/tmp/pkgreviewtest, remote openwashdata/pkgreviewtest (private). Open a Claude Code session in that directory and type /review-package; afterwards gh repo delete openwashdata/pkgreviewtest --yes and remove the directory.

Expected checks, refreshed:

• Standards file written as CLAUDE.md with Standard version: 1.5.0 and Organization profile: openwashdata
• Version and organization stamp lines in the first review issue body; intake screen section at its top
• All 17 scorecard defects caught in their areas, and no finding that maps to no defect
• Docs PR adds the pkgdown workflow and leaves docs/ untracked; final PR body names the Pages action
• No guardrail violations (check-ins honored, STOP after each PR, no PR against main before /review-complete)
• Result recorded here