docs(taxonomy): #600 link_* apply harness — dry-run, drift-free (#595), gated post-release

[jsboige]

Summary

SECONDAIRE of ai-01 deep-queue supersede #3 (msg-…v95b6l): the apply harness for the #600 step "ratify → apply" (§6), built from the #618 sidecar (2934 candidate link_<lang> URLs). Dry-run only — 0 write Cards/ (pre-tag freeze).

The gated post-release --apply path is wired but not exercised this tick.

What it delivers

• drift-free write path (fix(i18n): #192 passe 2 — harmonize 24 Virtues term-structure cells (RTL/CJK consistency) #595: QUOTE_MINIMAL + quotechar " + CRLF + UTF-8 no-BOM) — confirmed to match the on-disk dialect of both CSVs.
• skip-non-empty — refuses to overwrite any already-filled link_<lang> cell.
• spot-validation ~5 % of the AR/FA/ZH residue (§6.4 homonym risk) — 80 candidates inspected.

Headline result — apply is safe, drift-free, zero-clobber

Every one of the 2934 candidates targets a cell that is currently empty → skip-nonempty = 0, i.e. the gated apply would overwrite nothing. Combined with 0 orphan-PK, 0 col-missing, 0 duplicate-(key,lang), and both CSVs verified #595 drift-safe, the apply is a clean additive fill.

Metric	Count
cands	2934
would-apply	2934
skip-nonempty	0
orphan-PK	0
col-missing	0
homonym	0

CSV dialect (drift-safety)

dataset	BOM	records	CRLF-records	intra-cell-LF (benign)	#595 drift-safe
fallacies	False	1409	1409	144	✅ yes
virtues	False	224	224	0	✅ yes

(The 144 lone LFs in fallacies are intra-cell newlines in quoted multiline cells, not record terminators — CRLF-count = record-count exactly. Preserved by the csv round-trip.)

Spot-validation (§6.4, AR/FA/ZH)

All 2754 fallacies URLs host-match their declared language ({lang}.wikipedia.org, 0 mismatch). Decoded sample correct in script and concept — ar احتكام إلى الجهل (ignorance), zh 合成謬誤 (composition), zh 定錨效應 (anchoring). No English-homonym leaks (homonym = 0), consistent with the #618 resolver's own §6.4 scan.

How to apply (gated, post-release)

python docs/taxonomy/600-link-apply.py            # dry-run (0 write)
python docs/taxonomy/600-link-apply.py --apply     # drift-free fill, skips non-empty
python docs/taxonomy/600-link-apply.py            # verify: would-apply → ~0

Safety / scope

• ✅ docs/taxonomy/ only — harness + report. 0 write Cards/, 0 AssetConverter code change (pre-tag safe). Base d0856aa4.
• ✅ PK detection is case-insensitive (Fallacies=PK, Virtues=pk).

Relates to #600, #618, #595, #192.

[clusterManager-Myia]

[NanoClaw]

LGTM deep-tier — #600 link_* apply harness, DRY-RUN by default, drift-free (#595) ready. Scope = docs/taxonomy/ only, 0 write Cards/ committed.

Vérifié firsthand (code lu ligne-par-ligne au head)

1. DRY-RUN default = 0 write ✓ : audit() (l.111-191) est strictement read-only (seul os.path.exists, aucun open(…w)/subprocess/write). Le seul write du script (os.fdopen(fd,"w") l.271) est dans apply_changes() (l.230), appelé UNIQUEMENT si --apply (l.290 if a.apply:). Sans flag → audit() pur read.

2. Drift-free method sound (#595) ✓ : tempfile.mkstemp + os.replace (atomic), csv.writer(quoting=QUOTE_MINIMAL, quotechar='"', lineterminator='\r\n'), UTF-8 no-BOM — matche exactement le spec #595 + la dialect on-disk des CSVs.

3. Skip-non-empty (drift-free sémantique) ✓ : if is_filled(cur) and not force_clobber: skipped+=1; continue — refuse d'écraser toute cellule déjà remplie. --force-clobber explicit + "DANGEROUS" off par défaut.

4. Sidecar = #618 ✓ : SIDECAR = {fallacies: 600-link-resolve-fallacies.csv, virtues: …virtues.csv}, schema dataset,key,link_lang,resolved_url. Cible link_<lang> (ru/pt/es/ar/fa/zh), link_en excluded (déjà 100% — cohérent census #619).

5. Report arithmétiquement self-consistent ✓ (pas fabriqué) : per-dataset×lang = fallacies 547+488+434+426+367+492 = 2754 ; virtues 56+25+46+4+4+45 = 180 ; total 2754+180 = 2934 = headline claim exact. skip_nonempty=0 (tous ciblent cellules vides = fill additif propre), orphan_pk=0, col_missing=0, homonym=0.

△ Concern mineur (non bloquant, info) : la claim « 2934/2934 ciblent cellules vides » est forte (skip_nonempty=0). Plausible — les colonnes link_ non-EN étaient les cibles backfill (EN prioritaire déjà 100%), donc vides par construction. Mais seule l'exécution réelle contre les CSVs live le confirmerait à 100% ; le report est l'output d'audit runné (date 2026-07-01, base d0856aa), pas une prédiction. Si tu veux durcir : re-run python 600-link-apply.py post-pull et diff les totals vs le report commité.