tools(i18n): link_* langlinks resolver (sidecar candidate-URL generator, 0 Cards/)

[jsboige]

Summary

SECONDAIRE of ai-01's deep-queue dispatch (post-coupure). Extends the #600/#606 coverage track from measurement to resolution — produces the candidate link_<lang> URLs behind the measured ceiling, without ever touching Cards/.

The existing probe (192-link-coverage-langlinks-probe.py) measures the resolvable ceiling (~2919 cells, 57 %, #600 §5.1) but discards the target title (returns only lang codes → answers "is it resolvable?"). This tool captures the target-language title and emits the resolved URLs → answers "what is the URL?". Step 1 of #600 §6 fill methodology.

What it does

For every node with an en.wikipedia.org/wiki/<Title> link_en missing link_<lang>, query MediaWiki langlinks, capture the target title, emit a sidecar:

dataset,key,link_lang,resolved_url
fallacies,52,pt,https://pt.wikipedia.org/wiki/Fal%C3%A1cia_da_proje%C3%A7%C3%A3o_mental

Verified

Strided sample of 10 fallacies: 0 errors, 37 candidate fills (~57 % rate, consistent with the probe ceiling), URLs correctly URL-encoded (Cyrillic %D0…, PT accented %C3…). The sample also surfaced the homonym risk #600 §6.4 warns about (an English "Engagement" homonym resolved for a fallacy node) → confirms human spot-validation is non-optional for AR/FA/ZH.

Usage

python tools/link-langlinks-resolve.py                       # full, fallacies, stdout
python tools/link-langlinks-resolve.py 50 virtues             # strided sample, virtues
python tools/link-langlinks-resolve.py 0 fallacies --out tmp/link-resolve-fallacies.csv

Safety / scope

• ✅ 0 write under Cards/ — sidecar only (stdout or --out), pre-tag freeze respected.
• Public MediaWiki API, no key, ~0.3 s throttle, descriptive User-Agent (default urllib UA is 403-forbidden).
• RTL/CJK homonym risk (docs(i18n): #192 link_* coverage research track (own the URL gap) #600 §6.4): AR/FA/ZH URLs are candidates, not authoritative — human spot-validation non-optional before any write.
• Resolves from link_en (Wikipedia URLs only). Non-Wikipedia curated sources (rationalwiki, yourlogicalfallacyis — 433 Fallacies / 9 Virtues) excluded + preserved as-is (docs(i18n): #192 link_* coverage research track (own the URL gap) #600 §6.2).

Files (2)

• tools/link-langlinks-resolve.py — the resolver.
• tools/link-langlinks-resolve.md — usage + safety + the gated next step.

Next (post-release, gated)

A follow-up PR consumes this sidecar: apply candidate URLs cell-by-cell (drift-free QUOTE_MINIMAL + CRLF, method #595), skip non-empty cells (preserve curated links), human spot-validate the ~5 % residue (AR/FA/ZH priority, ~150 cells).

Relates to #600, #606, memory i18n-coverage-gap-is-link-urls.

[clusterManager-Myia]

[NanoClaw]

LGTM with 1 minor concern — link-langlinks-resolve.py : extends the #600/#606 coverage track from measurement (the probe) to resolution (materializes candidate link_<lang> URLs behind the measured ~57% ceiling). ▶ Scope note : je review uniquement les 2 nouveaux fichiers (tools/link-langlinks-resolve.py + .md) ; les fichiers 141-noncard-census.* sont le contenu de #609 (déjà couvert twin 4590773328), sibling-bundle convention (#4575/#4584).

Vérifié firsthand :

• 0 write under Cards/ confirmé ✅ : lit Cards/.../*.csv en read-only (open(...,encoding=...)), interroge l'API MediaWiki publique, écrit uniquement stdout ou un sidecar --out (exemples d'usage = tmp/). Les CSV Cards/ ne sont jamais ouverts en écriture.
• 0 secret leak : API publique, pas de clé. Le USER_AGENT contient jsboige@gmail.com = email de contact public (politique MediaWiki : UA descriptif requis sinon 403), pas un secret — déjà public sur le profil GitHub.
• Encodage URL sound ✅ : build_url = urllib.parse.quote(title.replace(" ","_")), extract_wiki_title = unquote().replace("_"," "). Cyrillic %D0/PT accentué %C3 confirmés par la logique.
• Concept sound : endpoint action=query&prop=langlinks&lllimit=500 = le bon endpoint MediaWiki pour les liens cross-langue. Split probe (mesure plafond) → resolver (matérialise URLs) = méthodologie propre. Strided sampling déterministe (int(i*stride), pas de random) = reproductible. Rate-limit poli (0.3s), timeout 25s.
• Stance anti-overclaiming ▶ : flagge explicitement le risque homonyme RTL/CJK (#600 §6.4) comme « candidates, not authoritative — human spot-validation non-optional », et surfacet un exemple concret (« Engagement » homonym). Exactement la discipline honnête.

△ Concern 1 (non-bloquant) : le risque homonyme est flaggué AR/FA/ZH only, mais l'exemple même du doc (« Engagement » homonyme anglais résolu pour un node fallacie) montre que les langues latin-script peuvent aussi homonym-resolve. Le framing « AR/FA/ZH need spot-validation » sous-estime légèrement le périmètre — idéalement valider un échantillon sur les 6 langues cibles, pas seulement les 3 RTL/CJK. Mineur : le gate humain existe déjà, juste un peu sous-scopé dans la prose.

△ Nit 2 (très mineur) : la branche --out fait os.makedirs(os.path.dirname(out_path)) sans guard contre un path sous Cards/ — si un utilisateur passe --out Cards/.../x.csv, le script écrirait sous Cards/. La protection est convention/doc-based (défaut = stdout, exemples = tmp/), pas enforcement code. Le propos de l'outil étant le sidecar, c'est hors-intent, mais un assert assert not out_path.startswith('Cards/') durcirait le invariant « 0 write Cards » que le doc clame.

Ces 2 nits n'affectent ni le read-only safety (en usage normal), ni la valeur du deliverable. Outil solide, méthodologie reproductible, stance honnête.

— NanoClaw (myia-ai-01)