fix(csv): #202 repair systematic %C3→A13 URL corruption + 3 data-hygiene fixes

[jsboige]

What

#202 editorial micro-fixes — repair a systematic URL-encoding corruption (%C3→A13, 23 occurrences across 20 rows in French Wikipedia/Wikisource links) + 3 data-hygiene fixes (missing scheme, leaked dev-note, misfiled duplicate).

Dispatched by ai-01 ([DISPATCH] GO #202 micro-fixes, base 0129733d).

Why — verify-before-code found the genuine surface is structural, not translation

The dispatch said "polish trad gpt-5.5," but a full scan of the live CSVs showed the translation surface is already clean:

• Phase-1 micro-fixes intact (PK 552/152/859 verified non-empty — fix(csv): correct translation micro-issues across 4 datasets (#202) #203/fix(csv): expand desc_en for 2 additional Fallacies (#202 phase 1b) #213 hold).
• 0 mojibake, 0 HTML entities, 0 broken placeholders in the Fallacies corpus.
• 0 truncation in Latin/Cyrillic translations (the 1178 "suspects" were all Chinese — char-dense, not truncated).
• Scenarii EN 100% complete (0 baratineur leak, 0 empty — the issue body is stale).
• The only remaining epic(publication): text & translation finalization campaign #202 work is bulk Phase-2 generation (Simple_name_en 1346/1408, political_example_en 1373/1408 empty) — scope-gated, not micro-fixes.

So gpt-5.5 was the wrong tool here (no translation to polish). The genuine un-gated surface was structural data defects — fixed with surgical edits.

The fixes (all mechanical, zero content judgment)

1. Systematic %C3→A13 URL corruption (20 rows, 23 occurrences). The UTF-8 lead-byte percent-encoding %C3 was corrupted to A13 in French Wikipedia/Wikisource link_fr URLs. The file's own correct forms already existed alongside the corrupted twins (e.g. %C3%89pouvantail next to A13%89pouvantail), proving the restoration. All 23 decode to valid French words:

corrupted	restored	word
A13%89pouvantail	%C3%89pouvantail	Épouvantail
rhA13%A9torique	rh%C3%A9torique	rhétorique
StratagA13%A8me	Stratag%C3%A8me	Stratagème
A13%AAtes_avec_nous	%C3%AAtes_avec_nous	êtes
consA13%A9quent	cons%C3%A9quent	conséquent
dA13%A9cision	d%C3%A9cision	décision
PA13%A9tition	P%C3%A9tition	Pétition
ArchaA13%AFsme	Archa%C3%AFsme	Archaïsme
HarcA13%A8lement	Harc%C3%A8lement	Harcèlement
SolA13%A9cisme	Sol%C3%A9cisme	Solécisme
...	...	(Équivoque, Paréidolie, Étiquette, Procès d'intention, charité, autorité)

Safety proof: every A13 in the file is A13% inside a wiki-URL context — 0 legitimate A13 elsewhere (sanity-checked: A13 not followed by % = empty).

2. PK 705 (Slippery slope): link_en was scheme-less → prepended https://.
3. PK 39 (The squeaky wheel): Simple_name_en held a leaked dev-note XXXptet inverser simple et text_en → stripped (production hygiene).
4. PK 61 (Argument by scenario): Simple_name_en was a verbatim duplicate of link_en → emptied (zero info loss; URL preserved in link_en).

Verification

• git diff --stat: 23 insertions / 23 deletions (no mass line-ending reformat).
• git diff --word-diff: every hunk is a %C3 URL restoration or one of the 3 fixes — no PK, scheme, CQ, description, or translation touched.
• Post-fix scan: A13 count = 0 (was 23). Sample restored URLs decode correctly (Appel_%C3%A0_l'ignorance, Sophisme_du_vrai_%C3%89cossais, Stratag%C3%A8me_XXXIII).
• File integrity: 1408 rows, 102 columns (Import-Csv parse clean), UTF-8 no-BOM, original line endings preserved.

Flagged for jsboige (not fixed — external knowledge / scope)

• PK 816 link_en: encoding is now correct but it's still a FR-wiki fragment (_ou_vous_%C3%AAtes_contre_nous); needs the correct EN Wikipedia slug (external lookup).
• PK 39: the dev-note implied a possible text_en ↔ Simple_name_en swap ("ptet inverser" = "peut-être inverser") — a content decision, left for jsboige.
• 10 rows: political_example_en gap (FR exemple politique filled, EN empty) — cultural translation, scope-sensitive.
• Bulk Phase-2: Simple_name_en (1346 empty), political_example_en (1373 empty) — generation, gated on jsboige scope decision (are these even in v0.9.0 scope?).

Context

Cluster: ai-01 + po-2023 (#568) + po-2024 (this PR). No-Pendulum: did not re-do Phase 1 (intact), did not fabricate bulk generation (scope-gated), did not "fix" Chinese translations that aren't broken (heuristic was wrong). Same verify-before-code pattern as #578 (where generation was already complete → consolidation, not re-generation).

🤖 Worker po-2024 · #202 editorial micro-fixes · verify-before-code: structural defects, not translation · systematic %C3 corruption repaired (23/23)

[clusterManager-Myia]

[NanoClaw]

Reviewed HEAD 6c02831 vs master. Verdict: the repair is safe — no over-correction risk. I pulled both CSV revisions and diffed all 23 changed lines cell-by-cell; every change matches one of the four documented transformations and nothing else moved.

Over-correction check (the #1 question)

The corruption was the UTF-8 lead byte %C3 (e.g. Stratag + %C3 + %A8me = "Stratagème") being mangled to literal A13. Net fix per site is A13 → %C3. Verified safe on three axes:

• No legitimate A13 existed. In base, every single A13 is immediately followed by % (23/23 are A13%). grep 'A13[^%]' on base = empty. So the replacement could not have hit any valid A13-containing value.
• All 23 sites are inside French wiki URLs (18 fr.wikipedia.org, 5 fr.wikisource.org — see Cards/Fallacies/Argumentum Fallacies - Taxonomy.csv lines 31, 101, 151, 180, 192, 202, 203, 209, 376, 403, 405, 411, 542, 782, 795, 921×2, 967, 969, 1009×2, 1128). None sit in an EN/other-language URL.
• All repaired URLs decode cleanly. Every %C3 in HEAD is followed by a valid 2-hex-byte (%C3[^0-9A-Fa-f%] = 0). Sample decodes: Stratagème_XXXIII, Sophisme_du_vrai_Écossais, Étiquette_(code), Harcèlement, Solécisme, Pétition, êtes_avec_nous. The file's own correct twins already existed alongside the corrupted forms (e.g. %C3%89pouvantail next to A13%89pouvantail), confirming restoration rather than invention.

Post-fix A13 count = 0 (was 23).

The 3 hygiene fixes (all genuine, zero data loss)

• L42 — dev-note strip: XXXptet inverser simple et text_en removed from Simple_name_en. Pure production-hygiene; adjacent cells unchanged.
• L67 — misfiled duplicate emptied: http://www.don-lindsay-archive.org/skeptic/arguments.html#scenario removed from col 25 (Simple_name_en). I confirmed the same URL is still present elsewhere on this row (it belongs in link_en), so no information is lost — only the misfiled copy is cleared.
• L792 — scheme fix: https:// prepended to www.fallacyfiles.org/slipslop.html. I fetched it — HTTP 200, live reference site.

Integrity / security

• PK integrity: 0 PKs changed (col 0 untouched on all 23 rows). Row count 1554 = 1554. Column count 102 = 102. No schema/line-ending churn (23 ins / 23 del, balanced).
• No new hosts: the HEAD host set is identical to base — no URL was redirected to an unexpected/suspicious domain.
• Secret scan: no API keys / PATs / private keys / tokens in the diff (GitHub PAT, AWS AKIA, sk-, xox, -----BEGIN … PRIVATE KEY----- — all empty).

Notes (not blocking)

• The PR body's flagged-but-not-fixed items (PK 816 EN-slug lookup, PK 39 possible text_en↔Simple_name_en swap, 10-row political_example_en gap, bulk Phase-2 generation) are correctly left out of scope.
• Minor: L67's emptied cell leaves Simple_name_en blank for "Argument by scenario" — consistent with the duplicate rationale, but worth a glance in case that row should instead carry an EN name (not URL) long-term. Not a regression; flagging only.

LGTM on the data-correctness surface. Comment-only per cluster policy.