tools(i18n): unified gated apply-spec Fallacies i18n-hierarchy gap (consolidates #712 + #716, dispatch h2utyb)

[jsboige]

Scope

Dispatch h2utyb PRIMARY (ai-01 2026-07-06 07:09) — consolidate the two i18n-propagation gap inventories (#712 family-gap + #716 subsubfamily-gap) into ONE post-tag apply spec + gated byte-targeted script. Base master 7ebeda18. GATED: 0 write to prod CSV (release freeze on Cards/ until v0.9.0 tag) — script + tests ship and run, --apply is post-tag.

Why consolidate

The 48 rows from #712 (empty Family/Subfamily) are the exact same 48 rows #716 found with empty Subsubfamily (overlap 48/48). Each is a pure i18n-propagation gap: FR Soussousfamille filled, all three EN hierarchy columns empty. The #712 apply would have filled 2 columns and missed the third — this does all three in one pass.

Files

• tools/fallacies_i18n_hierarchy_apply.py — Python 3 stdlib, same byte-targeted discipline as mnemonics_to_latin.py (chore(i18n): mnémoniques Virtues — normalisation (DEFER post-tag, scope à arbitrer) #654): BOM preserved, CRLF, QUOTE_MINIMAL, only empty target cells change. Re-derives the canonical FR→EN map at apply time (never hardcoded). Modes: --dry-run (default), --apply (gated), --report.
• tools/test_fallacies_i18n_hierarchy_apply.py — 11 unit tests, all PASS: canonical-map derivation, conflict detection, plan building (3-col/partial/novel/no-FR-key/full-row), apply round-trip (BOM+CRLF preserved, never overwrites filled cells), + real-CSV grounding (asserts 144 cells = 48×3, 0 conflict, 0 novel on master).
• docs/taxonomy/i18n-hierarchy-apply-spec.md — the unified spec: gap, 6-tuple canonical map, script usage, 5 validation gates, post-tag apply procedure, impact, relationship to the link_fr→link_en gap (24 URLs, human-research, out of scope).

Dry-run on 7ebeda18 (the report)

144 cells (48 rows × 3 EN columns), 0 conflict, 0 novel key, canonical map 63 FR keys (consensus). The 6 gap FR keys → tuples:

• Jeu de pouvoir (33) → Influence / Psychological manipulation / Power games
• Conditionnement (6) → Influence / Psychological manipulation / Conditioning
• Mensonge (4) → Cheating / Spin doctoring / Lying
• Langage persuasif (2) → Influence / Rhetorical device / Loaded language
• Ambiguïté narrative (2) → Misleading language / Ambiguity / Narrative ambiguity
• Repoussoir (1) → Influence / Appeal to emotion / Foil

Validation gates (enforced in code, abort exit 2)

1. Re-derive mapping at apply time (no hardcoded tuple) 2. Conflict (>1 tuple per FR key → abort) 3. Novel key (FR value absent from map → abort) 4. Row count preserved 5. BOM + CRLF + QUOTE_MINIMAL byte-preserved.

Post-tag apply = one command

python tools/fallacies_i18n_hierarchy_apply.py --csv <PATH> --apply (after dry-run confirms 144/0/0 + tests pass). Supersedes #712 + #716 as the single authoritative apply spec.

---

DoD

Item	Status
One unified spec (supersedes #712 + #716)	✅ docs/taxonomy/i18n-hierarchy-apply-spec.md
Gated byte-targeted script (#654 discipline)	✅ tools/fallacies_i18n_hierarchy_apply.py
Validation gates documented + enforced in code	✅ 5 gates, exit 2 on any failure
Unit tests (synthetic + real-CSV grounding)	✅ 11/11 pass (144 cells, 0 conflict, 0 novel)
Release freeze respected	✅ 0 write Cards/, 0 --apply execution

---

Gate boundaries (HARD — spec + gated script, NOT executed)

• ❌ No prod CSV write, no build, no prod test, no OWL regen.
• ❌ No --apply execution during release freeze.
• ✅ Script + unit tests ship; dry-run on real CSV produces the 144-cell / 0-conflict report.

Relates: dispatch h2utyb (PRIMARY), #712, #716, #654 (byte-targeted discipline), #587 (CSV hygiene), #707 §1. Base 7ebeda18.

🤖 Generated with Claude Code