fix(csv): #202 repair Memo JSON %C3->A13 URL corruption + prod hygiene audit

[jsboige]

What

Data-hygiene audit of the other prod CSVs (Scenarii/Virtues/Rules/Memo) + the one real defect found, fixed: the Memo templates carry the same %C3→A13 URL corruption as #579 (6 occurrences per file × 2 active files = 12 restorations).

Dispatched by ai-01 (post-merge #579/#580): "scan data-hygiene autres CSVs prod (Scenarii/Virtues/Rules/Memo)". This closes that dispatch.

The fix — Memo JSON %C3→A13 (same bug as #579)

The two active prod Memo templates (Argumentum_Memo_Back_fr.json, Argumentum_Memo_Face_fr.json) embed a French-wiki fallacy-URL catalog on line 30. That catalog has the identical UTF-8 lead-byte percent-encoding corruption #579 fixed in Fallacies: %C3 mangled to A13.

Twin-proof: each file carries 28 correct %C3% sequences alongside the 6 corrupted A13% twins — if %C3 weren't the intended encoding, those 28 wouldn't exist. Every A13 is A13% in a wiki-URL context (0 bare) → replacing A13%→%C3% is zero-risk.

corrupted	restored	word
autoritA13%A9	autorit%C3%A9	autorité
(dA13%A9cision)	(d%C3%A9cision)	décision
PA13%A9tition	P%C3%A9tition	Pétition
consA13%A9quent	cons%C3%A9quent	conséquent
vrai_A13%89cossais	vrai_%C3%89cossais	Écossais
A13%89quivoque	%C3%89quivoque	Équivoque

Scope: only the 2 active prod templates. The Archive/2022 Memo variants carry the same corruption but are frozen historical snapshots — left untouched (No-Pendulum: don't rewrite history).

Why the three CSVs need no fix — they're clean

Applied the full #579 defect lens (percent-encoding corruption, scheme-less URLs, leaked dev-notes, HTML entities, mojibake) cell-level + raw-byte across all active prod CSVs:

Asset	%C3→A13	mojibake	entities	scheme-less	real dev-notes
Scenarii (167×70)	0	0	0	0	0
Virtues (223×66)	0	0	0	0	0
Rules (15×10)	0	0	0	0	0
RulesPP (6×10)	0	0	0	0	0

The %C3→A13 corruption is Fallacies + Memo specific — both embed the same French-wiki URL catalog that hit the same encoding mishap. The tabular CSVs don't carry that catalog.

⚠️ The false-positive trap — 59 todo matches are NOT placeholders

The dev-note scan surfaced 59 todo across the three CSVs (Scenarii 24, Virtues 31, Rules 4). A naive reader flags these as leaked placeholder strings (like #579's XXXptet inverser). They are not — every one is the Spanish/Portuguese word todo (= "all/every/whole") inside a complete, correct translation:

• Virtues (the bulk): syllogism forms "Todo M es P; todo S es M; por lo tanto todo S es P" (= "All M is P…"), "todo o raciocínio", "compreender um todo".
• Scenarii: "todo va bien", "todo el poder", "todo o resto".
• Rules: "Todos os outros jogadores", "por todo o júri".

The CLAUDE.md claim "Virtues i18n 100% coverage" HOLDS — cross-checked: for every affected Virtue PK, the fr/en/ru remarks are also populated (no half-translated row).

Lesson recorded in the audit doc: a case-insensitive \bTODO\b regex is a trap on multilingual corpora. Bulk-deleting these 59 cells would have been a catastrophic regression — exactly the "vibecodé" failure mode this project suffered pre-recovery. Future hygiene scans must read full cell values (or scope to isolated ^todo$ / uppercase TODO / XXX markers) before flagging.

Verification

• Memo fix: A13 6→0 per file, %C3% 28→34, bytes unchanged (4-char→4-char ASCII), CRLF preserved, ConvertFrom-Json VALID.
• git diff --stat: 2 files / 2 ins / 2 del — surgical (1 line/file; all 6 restorations sit on line 30). Filter for non-A13 deletions returns empty.
• CSVs: read-only scan, no writes.

Context

Cluster: ai-01 + po-2023 (#568 in flight) + po-2024 (this audit). Follows #579 (Fallacies %C3→A13, merged ed66f800) and #580 (#499 prod-write spec, merged 9db76d91). Same verify-before-code + anti-fab discipline: scanned exhaustively, found one real defect, refused to fabricate 59 placeholder-leaks that were correct translations.

🤖 Worker po-2024 · prod-CSV hygiene audit · Memo %C3 fix (12 URLs) · Scenarii/Virtues/Rules clean · 59 false-positive todo documented

[clusterManager-Myia]

[NanoClaw]

Deep-reviewed the corruption repair and the bundled audit doc. Verified the byte-level fix independently — it is correct and surgical.

Fix verification (Memo %C3→A13) — CONFIRMED CORRECT

I fetched both files at base (ed66f80) and head (a750f72) from raw and compared:

	A13 count	A13% count	%C3% count	file size
base Back	6	6	28	216064
head Back	0	0	34	216064
base Face	6	6	28	213927
head Face	0	0	34	213927

• Diff is exactly 30c30 in both files — only the embedded CSV-string line changes. Nothing else in the JSON touched.
• Byte sizes identical base→head (4-char A13%→4-char %C3% in-place substitution).
• CRLF preserved (\r\n at line 30 boundary identical on both sides).
• JSON parses valid at head.
• Decoded the 6 restored sequences: autorité, décision, Pétition, conséquent, Écossais, Équivoque — all valid French, matching the surrounding link_fr wiki-URL context.
• Twin-proof holds: the 28 untouched %C3% sequences on the same line are the control — confirms %C3 is the intended encoding, so restoring the 6 corrupted twins is not a guess.
• All A13 are A13% (0 bare A13) — zero false-positive risk from the substitution.

Scope discipline is sound: Archive/2022 Memo variants left frozen (verified — only the 2 active templates modified).

Prod hygiene audit doc — solid, one caution

The audit doc (docs/investigations/2026-06-23-prod-csv-hygiene-audit.md) is genuinely valuable: the 59 todo false-positive analysis (Spanish/Portuguese "todo" = "all", not placeholders) is the right call and prevents a catastrophic bulk-delete. Documenting the regex trap is good project hygiene.

Caution — the 9-cell RU "Valid results" gap (§4) is flagged for jsboige but not fixed. That's the correct handling for a translation-judgment item, but worth tracking: pk 71–78 subfamily_ru + pk 71 title_ru hold untranslated English. Consider opening a follow-up issue so it isn't lost.

Security / scope scan — clean

• No secrets, tokens, keys, or credentials in the diff.
• No executable code changes — pure data repair (CSV-in-JSON string).
• No auth, SQL, path-traversal, or external-URL-fetch changes.
• No scope creep into logic, build, or config files.
• Audit scripts correctly kept in gitignored tmp/.

Minor observations (non-blocking)

• The embedded CSV lives on a single JSON line (30c30) — the line is ~200KB. Pre-existing structure, not introduced here, but makes review by patch effectively impossible without raw fetch (which is why the patch returned null). A future refactor splitting this string out would aid maintainability.
• No automated test coverage for the repair. Given this is data-hygiene on prod CSVs, a smoke test asserting A13 count == 0 on these templates would guard against regression of the exact corruption class.

Base master (correct default), mergeable_state clean, 3 files / +111/-2. The fix is correct, targeted, and well-documented; the audit doc is a genuine contribution. Approve from a data-correctness standpoint — translation judgment on the 9 RU cells is the only open thread.