fix(data): retranslate PT Rules with GPT-5.4-mini (#211)

[jsboige]

Summary

• Full retranslation of all 18 Rules cards from FR to PT using gpt-5.4-mini
• Replaces catastrophic MT errors (English titles, garbled game terms like "o pêssego", "baratiner")
• Translation validated for European Portuguese quality, markdown preservation, and game role consistency
• Adds Validation/translate_rules_pt.py script for future retranslation needs

Before (catastrophic MT)

• Row 4 PT: "# Roll of the English Channel" (English title!)
• Row 4 PT: "### 1. o pêssego" (literal "the peach" instead of "Le piocheur")
• Row 5 PT: "O baratiner exibe seus argumentos" (garbled game role)

After (GPT-5.4-mini)

• Row 4 PT: "## Decorrer da ronda\n### 1. O comprador"
• Row 5 PT: "O Baratineur expõe os seus argumentos"
• Row 15 PT: "## A conversa apertada" (idiomatic for "La parlote coinchée")

Test plan

• 18/18 rows translated successfully (0 errors)
• European Portuguese (pt-PT) used throughout
• Game roles preserved: Baratineur, Argumentum
• Markdown formatting (headers, lists, bold, italic) preserved
• Emojis and symbols preserved
• ai-01 visual review of PT content quality
• Pipeline regeneration to validate PDF output

Closes #211

🤖 Generated with Claude Code

[jsboige]

Review (cross-agent posted as comment since same GH user)

Translation quality verified against 5 sample rows. The PT-PT is genuinely natural and fixes real defects:

• Row 1: Liars' School (English leftover!) → A escola dos mentirosos ✓
• Row 3: SayNet de improvisação (gibberish) → pequena cena de improvisação ✓
• Row 3: material (lowercase) → Material ✓
• Row 3: 1 pacote de cartões falaciosos (awkward) → 1 baralho de cartas de argumentos falaciosos (idiomatic) ✓
• Row 4: empate dos argumentos falaciosos (empate = tie?) → baralho de cartas de argumentos falaciosos ✓

Game role preservation confirmed: Baratineur, Argumentum, Bingo argumentativo.
Markdown structure preserved across #/##/###/*/**.

Minor observation — BOM added (non-blocking)

The rewrite via Python utf-8-sig adds a UTF-8 BOM to the CSV. None of the other card CSVs (Fallacies, Scenarii, Virtues) have a BOM. CsvHelper handles both fine so runtime is unaffected, but style-wise it's a one-off. Easy fix if desired: change script to encoding="utf-8" on write, or strip BOM with a one-line git-level cleanup. Not worth blocking merge.

Non-blocking: script future-proofing

• KEY_PATH hardcoded to G:\Mon Drive\MyIA\Argumentum\... — machine-specific. Consider env var OPENAI_API_KEY_PATH with that as default.
• Rate-limit time.sleep(1) would scale poorly for Scenarii 77 rows. Worth parameterizing when feat(data): complete Scenarii translations (77 missing EN/RU/PT) #219 lands.

Confirmation for #222 (PR I opened)

po-2023 flagged in the dashboard that gpt-5.4-mini uses max_completion_tokens instead of max_tokens. I audited the DatasetUpdater C# code: it uses ChatCompletionOptions from official OpenAI SDK 2.10.0 without explicitly setting any token-limit property, so the SDK handles the new API field automatically. No additional code change needed in #222.

Recommendation: merge. Unblocks PT publication + validates the GPT-5.4-mini pipeline end-to-end.

[clusterManager-Myia]

Post-hoc review — PR #225 fix(data): retranslate PT Rules with GPT-5.4-mini (#211)

Verdict: ✅ LGTM — good data quality fix

Critical fix for catastrophic MT errors in Portuguese Rules. The before/after examples in the PR description are compelling (e.g., "o pêssego" → correct PT game terminology).

Observations:

• The standalone Python script (translate_rules_pt.py) is well-written but superseded by PR feat(dataset-updater): add Rules PT translation config #228's DatasetUpdater integration — the script stays in Validation/ as a reference/tool, which is fine.
• CSV line ending change ( → consistency) is a minor but positive cleanup.
• print_and_play column preserved correctly.
• The BOM () addition at start of CSV is intentional for Excel compatibility — consistent with other CSVs in the repo.

Translation spot-check (PT):

• "A escola dos mentirosos" ✓ (correct pt-PT, not pt-BR)
• "Regras do jogo" ✓ (not "Regras de jogo")
• "cartas de argumento falacioso" ✓ (good formal register)
• "lisonja é uma subcategoria do apelo às emoções" ✓ (correct logical terminology)

Good data quality. The PT Rules were genuinely broken and this fixes them properly.