scripts/data-inventory.py: a dataset manifest without a single value

.claude/skills/replication-package/SKILL.md

@@ -38,7 +38,7 @@ Write `replication_package/README.md` (the AEA template, fields below). Leave a
 
 - **Overview / paper citation** — title, authors, abstract one-liner.
 - **Data Availability Statement** — for each dataset: public / restricted / proprietary, and whether it is redistributed in the package. This is the single most-rejected DCAS field; be explicit.
-- **Dataset manifest** — a table, one row per file: `filename | description | source (URL/citation) | access (public / DUA / purchase) | license | provided in package? (Y/N)`.
+- **Dataset manifest** — a table, one row per file: `filename | description | source (URL/citation) | access (public / DUA / purchase) | license | provided in package? (Y/N)`. Start it from `python3 scripts/data-inventory.py data/raw --format md` (add `--header` once you have checked that each CSV's first row is a header), which lists each file's size, SHA-256, rows, column names, and the columns whose *names* suggest personal data, and never a value. Everything in it is still derived from the data. On restricted data the author runs it on the permitted machine, and the **whole** manifest (file names, sizes, row counts, column names) goes through the agreement's disclosure review (`/disclosure-check`) before it leaves that machine or reaches a model. `--no-counts` drops the row counts.
 - **Computational requirements** — OS, software + versions (R / Stata / Python), key packages, approximate runtime, RAM, any HPC/cluster need.
 - **AI-generated analytical outputs** (only if the analysis used a model to *produce data*, such as variables coded by an LLM, classified text, or extracted fields) — for each output: the exact model ID or snapshot and the date it ran, the prompts verbatim, the parameters, where the raw model output is kept, the post-processing applied, and who checked a sample and how. AI help with *writing code* is not an analytical output; it is disclosed under the journal's AI policy (`/submission-disclosures`). The distinction follows the World Bank's reproducibility guidance (`worldbank/wb-reproducible-research-repository`).
 - **Step-by-step run instructions** — the single master-script invocation, then the expected outputs.

CHANGELOG.md

@@ -16,12 +16,20 @@ An **adoption release** carrying what the 2026-11-06 audit found worth building
 - **`/verify`** (62nd skill). Since 2.1.286, Claude Code tells Claude to run a project skill named `verify` right before committing; docs-only and tests-only commits are exempt. This one runs the backtest, and the quality score on staged `.qmd`/`.tex`/`.R` files, and stops the commit on a failure. It skips itself when the git pre-commit hook will run the same gates. It is a reminder, not enforcement: the hook and `/commit` remain the gates. It is bridged for Codex and Antigravity.
 - **`templates/triangulation-ledger.md`.** For each robustness check, write down and commit what result would support the claim, weaken it or leave it undecided, before the check runs; the result is filled in afterwards in a new commit. It is method-neutral and joins `spec-ledger.md` on date and `Script:line`.
 - **Template, not fork, for unpublished work** (README and guide, step 1, placed before the clone commands). The repository is now a GitHub template repository, so a copy can be made private; a fork of a public repository cannot.
+- **`scripts/data-inventory.py`: a dataset manifest that never carries a value.** For each file it lists the path, size and SHA-256 and, for tables, the row count, the column names, and the columns whose *names* suggest personal data (name, email, address, birth date, ...).
+  - **CSV and TSV** use the standard library. A first row is listed as column names only with `--header`, your check that every file given has a header, because no rule can tell `Alice,London` from a header. Even with `--header`, a numeric first row is withheld.
+  - **`.dta`, `.sav` and `.sas7bdat`** use pyreadstat's metadata-only read, and **Parquet** uses its footer, when those packages are installed.
+  - **`--no-counts`** drops the row counts.
+  - **On restricted data**, the author runs it, and the whole manifest goes through the agreement's disclosure review before it leaves the machine or reaches a model.
+  - `/replication-package` starts its dataset manifest from it.
+  - Six portability cases pin it, under cp1252 and with non-ASCII file names. Three mutants each fail their own case: one leaks a value, one drops the numeric-row check, and one lists names without `--header`. Codex's review on PR #184 found the all-text-first-row leak and the clearance gap.
+  - The idea follows the World Bank's reproducibility skill; the code is our own.
 - **`/replication-package`: AI-generated analytical outputs.** When a model produced data, the README records the model ID and date, the prompts, the parameters, the raw outputs, the post-processing and who checked them. AI coding help stays a disclosure matter. The idea follows the World Bank's reproducibility guidance; the wording is our own.
 - **Links:** Brodeur et al. 2026 (*PNAS*: AI-assisted teams reproduce about as often as human-only teams but catch fewer major coding errors; AI-led teams reproduce far less) in `verification-ladder.md`; `paulgp/econ-ai-detector` in `writing-with-ai.md` and the ecosystem section; `posit-dev/skills` in README and the guide.
 - **TROUBLESHOOTING: "Every Bash, Edit and Write is refused, with a hook error"**, covering a missing hook script (recovery from a terminal) and a missing `python3` on Windows.
 
 **Inventory at release: 62 skills, 18 agents, 37 rules, 11 hooks, 11 gates**
-(standing hook battery: 322 cases, seconds to run; portability suite: 233 cases).
+(standing hook battery: 322 cases, seconds to run; portability suite: 239 cases).
 
 ---
 

quality_reports/qualification/LEDGER.md

@@ -62,6 +62,7 @@ meaningless:
 | 2026-11-06 | `scripts/file-issue.py` (PR #177 review, #174: `--repo` pinned to `origin` as HOST/OWNER/NAME in any spelling — `GitHub.com`, `host:/owner/name`, an ssh-config alias resolved with `ssh -G`, a GitHub Enterprise host; an origin that is neither local nor identifiable stops the script, exit 2, without repeating the URL; the printed duplicate follow-ups carry the same pin) | `scripts/hook-battery.sh` run with `FILE_ISSUE` pointing at the first version (0bd456d), at the first repair (the copy an external GPT-6 Pro review read — `quality_reports/oracle_audits/2026-11-06_pr177-fork-safety/`), and at a mutant that passes `--repo X` as one argument | case-varied host · ssh alias to github.com · ssh alias to an Enterprise host · Enterprise host · the template's owner/name on another host · scp absolute path · unreadable origin carrying a token · fused `--repo` argument · unpinned follow-ups | 9 (i11–i14, i16–i18, i20, plus the fused mutant) | 0bd456d red on i11–i14, i16–i18, i20; the first repair red on i11, i12, i14, i16–i18, i20 (and on i19, which it mis-pinned); the fused mutant red on i11–i13, i16–i18 | 0/2 controls on the final version (i15: the template on github.com; i19: a `file://` origin) | Codex (P1) found the first class; the GPT-6 Pro review found the rest in the first repair | **PASS** |
 | 2026-11-06 | `check-skill-integrity.py` check 6 — **the `.agents/skills/` bridge is current** (every skill classified in `scripts/sync-agent-skills.py`; every generated wrapper for Codex and Antigravity matches its `.claude/skills/` source) | four copies of the working tree in the session scratchpad, outside the checkout, one seeded defect each; the real script | wrapper description edited by hand · a wrapper deleted · a stray file in the bridge | 3 | 3/3 (exit 1, the file and the fix named) | 0/2 (an unmodified copy: exit 0; a fork's new skill in neither list: a P2 advisory naming the fix, exit 0, by design) | new check; the bridge did not exist before | **PASS** |
 | 2026-11-06 | `check-ledger-coverage.py` — **reads the inline hook launcher** `sh -c '<code>' <name> <script>` (#179, #180) as running `<script>` | the pre-change and changed `script_of()` loaded side by side in the session scratchpad, given the launcher drafted for #180 | launcher naming a real hook · launcher naming a misspelt hook · launcher with no script | 3 | the changed parser registers the real hook, flags the misspelt one, refuses the scriptless one; the old parser flagged all three as missing (a false alarm on the valid form) | 0/1 (the plain `python3 <script>` form: registered by both) | needed before the owner applies the launcher patch on #180, or the gate would fail the repository | **PASS** |
+| 2026-11-06 | `scripts/data-inventory.py` (new: a dataset manifest that never carries a value; CSV/TSV names only on `--header`; `--no-counts`) | `tests/portability/test_scripts.py::DataInventory` run against the script and three mutants loaded from the session scratchpad | a data value reaches the output · a numeric first row listed as names · any first row listed as names without `--header` (the all-text-row leak Codex found on PR #184) | 3 | 3/3 (each mutant fails only its own case) | 0/6 on the real script, on Python 3.9 and 3.14, with cp1252 output and non-ASCII names | new script | **PASS** |
 | 2026-09-26 | `slide-qa.py` (new: measures a rendered Reveal.js deck in headless Chrome) | two regression decks in a scratchpad (30 slides MathJax 2 incl. knitr output, 4 slides KaTeX) + the earlier 8-slide seeded deck + `Quarto/HelloWorld.qmd`; Playwright 1.63.0 driving Google Chrome 152 | too-tall list · too-wide box · sideways-scrolling code · too-wide, too-tall and inline-fraction formulas (MathJax and KaTeX) · kable table and plot past the bottom · overflow:hidden wide div · fade-out fragment · scrollable slide · slide with its own transition · auto-animate target · second tab of a tabset · lazily loaded image · uncounted slide | 22 | 22/22 (18 in the regression decks + 4 in the seeded deck, each on its own slide) | 0/23 (16 clean regression slides — math, fragments, columns, small table, small and stretched images, tabset, transition, filler — 4 clean seeded slides, 3 HelloWorld slides) | first version: 4/4 on its own deck, but an adversarial review then found 13 blind spots (chunk output never measured, inline MathJax box, transitions mid-animation, uncounted slides skipped, and others); every one is a slide in the regression decks | **PASS** |
 | 2026-09-26 | `validate-findings.py --check-quotes` (every double-quoted span of 12+ characters in a finding's evidence must occur in the cited file, or in another file the evidence names by path) | 24 seeded findings against LaTeX, Quarto, Beamer and PDF fixtures in a scratchpad, plus a binary file | fabricated quote · changed number · text from a file not named · misquoted number after a `%` · swapped notation (`\hat\alpha` for `\hat\beta`) · elided fragments stitched from far apart · parity quote in neither file · fabricated quote from a PDF | 8 | 8/8 (exit 1, each miss listed with the closest line) | 0/16 (exact LaTeX; curly quotes, minus sign, spacing; `...` and `[…]` elisions; hyphenation both ways; backticked command; a short quoted term before a real quote; `\%`; LaTeX quote marks; forced break and `---`; inline code inside a quote; a Beamer quote in a parity finding; a genuine PDF quote) — and an unreadable binary file left unchecked, exit 0 | first version: an adversarial review found PDFs always failing, short quoted terms shifting the pairing, text after `%` never checked, and notation swaps passing; all fixed before this row | **PASS** |
 | 2026-09-26 | `check-repo-hygiene.py` (append-only check for the replication log and the specification ledger: every committed line kept, in order; `APPEND_ONLY_BASE` in CI; layout and R/Python environment files allowed) | a throwaway git repository | committed line edited · log deleted · path turned into a directory · edit staged while the working tree is restored · `git rm --cached` · a rewrite committed without the hook, caught with `APPEND_ONLY_BASE` | 6 | 6/6 (exit 1, naming the reference, where, and the first lost line) | 0/5 (an appended entry; a CRLF working tree; a CRLF tree with an appended entry; a co-author's rows inserted by a merge; the same rewrite locally with no base set, which HEAD already contains) | before: append-only was only a request, and `output/` — the layout's own results folder — failed the gate; the first version of this check also failed CRLF checkouts, missed `git rm --cached` and could not fire in CI (all fixed) | **PASS** |