tools(dnn-i18n): #457 content-type extractor + reimport dry-run (fixture, Option C tooling)

[jsboige]

What

#457 DNN i18n tooling — bricks 1 + 3 (content-type extractor + re-import dry-run verifier) + round-trip DoD test. Dispatched by ai-01 (msg-…141408, non-gated #457 tooling while #131 exec waits on jsboige's target decision).

Key finding that shaped the work

Brick 2 (DatasetUpdater Option C config) was ALREADY merged in #487 (ca9a8640) — entity DnnUiString, KnownDataSets.DnnUiStrings, prompts, task config (DatasetUpdaterRootConfig.cs:2636-2696, Enabled=false, gpt-5.5). The investigation (SDDD triple-grounding) confirmed the config-only gpt-5.5 path is canonical. So this PR adds the missing tooling around it rather than rebuilding the config:

Brick	What	Status
(2) config Option C	gpt-5.5 task ui./res.	✅ already done (#487)
(1) extractor	content-type → CSV (codifies PHASE1 manual audit)	🆕 net-new
(3) re-import verifier	key-set diff + payload render (dry-run)	🆕 net-new

Brick 1 — extract_dnn_ui_strings.py

Parses the Argumentum 2sxc .cshtml templates, emits the dnn-ui-strings.csv dialect (key,context,source_file,fr,en,ru,pt,es,ar,fa,zh,notes):

• ui.* = anchor-based (NOT free-text): each entry declares a verbatim anchor that must exist in the named source. Fails loud (exit 2) if a refactor removes a string — never silently drops a row, never fabricates.
• res.* = @?Resources.<Key> regex capturing both @Resources.X (directive) and @Html.Raw(Resources.X) (nested expression). The nested form recovers RuleMemoInstructions (the one a naïve @Resources. regex misses). fr left empty + flagged DB-only (canonical values live in SQL/2sxc App Resources — pattern appris).

Brick 3 — reimport_dnn_ui_strings.py

• verify — key-set diff vs reference. KEY-SET match is HARD (exit 1 on divergence). res.* fr empty-in-extract is by-design WARNING.
• reimport — renders the canonical DNN/2sxc payload to stdout only (never writes — live apply is DB/RDP-gated, jsboige).

DoD proof — test_roundtrip.py

[1/4] extract OK: 10 rows (2 ui.*, 8 res.*)
[2/4] key set OK (10 keys)
[3/4] verify OK (key sets match, res.* fr empty by design)
[4/4] reimport dry-run OK (payload rendered, nothing written)
ROUND-TRIP DoD PASS — exit 0, zero prod mutation

Anti-fabrication (proven, not asserted)

• Negative test: breaking a ui.* anchor → exit 2 (verified, not piped).
• Cross-validated vs REAL prod templates: extractor run on DNNPlatform/Portals/1/2sxc/Argumentum/ yields 10/10 keys matching docs/dnn-localization/dnn-ui-strings.csv. Only delta = res.* fr (empty-in-extract by design — DB-only).

Gate boundaries (HARD)

• ❌ No live DNN DB / portal / 2sxc App Resources mutation (DB/RDP-gated jsboige).
• ❌ Does NOT touch dnn-ui-strings.csv (po-2024 lane, feat(dnn-i18n): #457 translate DNN UI strings FR -> 7 langs via gpt-5.5 #490).
• ❌ Does NOT enable/modify the DatasetUpdater config (feat(dataset-updater): #457 DNN UI strings localization infra (Enabled=false) #487 stays Enabled=false).
• ❌ No translation run (gpt-5.5 translation is the config's job, gated on FR-complete which needs the portal export).
• ❌ No QA verdict declared (ai-01 only). Stdlib Python, 0 deps.

🤖 Generated with Claude Code

[clusterManager-Myia]

[NanoClaw]

Clean i18n tooling (bricks 1+3) for DNN localization: anchor-based ui.* extractor (anti-fab: fails loud if anchor disappears, never invents rows), res.* reference scanner (DB-only values honestly flagged), CSV reimport dry-run verifier, and a round-trip test with fixture. Stlib-only Python, cross-validated against prod templates (10/10 key match). Gate boundaries explicit: no DB touch, no live reimport, no CSV modification, no DatasetUpdater config change. Subprocess only used to run its own test script. Good README with anti-fabrication guarantees. LGTM.