Published skill `objectstack-i18n` teaches that validation messages are not a translation group — false once #14253 lands

[os-warren]

Blocked-by: #14315

Found while implementing #14253. Not fixed in that PR on purpose: skills/** is a governed surface (Prime Directive #14), a mixed diff is refused whatever the proportion, and the governed files have to be split into their own human-merged PR.

The claim, and why it goes false

skills/objectstack-i18n/SKILL.md:270 carries a blockquote stating that validation messages are not a translation group. That was true and correct when written: validationMessages had been retired in 17.0.0 (#4667, ADR-0049) precisely because no resolver read it, and the only honest advice was "author the message on the rule".

#14253 adds objects.OBJECT._validations.RULE.message, object-scoped, resolved on the write path by the ObjectQL rule evaluator through the i18n service that has localized built-in messages since #3957. #14253 has now landed, so the sentence is false on main today — it is a signpost pointing at a dead end, which is the same defect the retired key had, one layer up, and this catalog ships to codebases this repo cannot see.

Scope

• Correct the objectstack-i18n claim, and check its neighbours: the same file teaches the bundle's group inventory in several places, and the datasets group and _views.VIEW.bulkActions are new too.
• ⚠️ Budget. This is a correction, not an expansion — maintainer ruling 2026-08-21: 「对外发布的 skills 是整个平台的最大价值，尤其要整体考虑和评估。」「……不能为了一个小功能扩写很多。」 Three new key families do not license three new sections; if the net line/token increase cannot be kept near zero, stop and report the gap rather than self-authorizing a bigger budget.
• ⚠️ Collision. Skills optimization flight — skills/objectstack-i18n: DIET (≈ −1,685 tok, −27%) — "Core Concepts" restated in full four sections later, a per-namespace layout with zero real usages, os i18n check taught in six places, and the #1 real pattern (a package shipping its own generated bundle) dead-ends #14315 is an open DIET card on this same file (≈ −1,685 tok). Whoever picks either up should read the other first — a correction landing under a rewrite, or a rewrite landing over a correction, loses one of them silently.

Landing

Governed surface: push the branch, open the PR, and stop there. ⛔ Never merge, queue, arm auto-merge on, or flip out of draft. A finished task here is a PR visibly awaiting the maintainer's own merge.

Generated by Claude Code

[huangyiirene]

Triage: documentation · priority:p2 · pm:blocked · domain:skills · type Task.

The collision is now encoded, not just described. The card named #14315 as a hazard in prose; prose is invisible to the unlock sweep and to whoever picks a card off the queue. A second body line — Blocked-by: #14315 — is added beside the existing Blocked-by: #14253, so this card is released only when both are done. That is the honest state: the sentence is not false until #14253 lands, and correcting a file that a −1,685-token rewrite is about to replace loses one edit or the other silently. pm:blocking derived onto both upstreams.

⇒ Ordering falls out of it: #14315 first (it is pm:queue and unblocked today), then this correction against the rewritten file. ⛔ Whoever takes #14315 must not pre-emptively fix line 270 — the claim is still true on main, and "fixing" it early would put a falsehood in the published catalog for however long #14253 takes.

p2, not p1. The distinction from the skills-flight cards (#14315 and its siblings, p1) is that those are a maintainer-directed programme with a standing budget; this is one factual correction that is not yet false. It joins the same lane and the same file, at the priority its own consequence earns.

The budget fence is the load-bearing instruction and I am restating it as a dispatch condition, not advice. The maintainer's ruling is quoted verbatim in the body — 「对外发布的 skills 是整个平台的最大价值，尤其要整体考虑和评估。」「……不能为了一个小功能扩写很多。」 ⇒ Three new key families do not license three new sections. Target a net token change near zero; the group inventory already exists in the file, so the new families belong inside it rather than beside it. ⛔ If the net increase cannot be kept near zero, stop and report the gap — a dev agent does not widen a published-skill budget on its own reading of "but the reader needs to know".

Governed-surface landing, verbatim from the card and not negotiable: push the branch, open the PR, stop. ⛔ Never merge, queue, arm auto-merge, or flip out of draft. A finished task here is a PR visibly awaiting the maintainer's own merge.

Scope check while in the file. The card is right to widen the read without widening the edit: the same file teaches the group inventory in several places, and datasets plus _views.VIEW.bulkActions are new in the same landing. Correct every place the inventory is stated, so the file does not end up internally inconsistent — that is still a correction, not an expansion.

Size/model suggestion: S, at CONTRACT_REVIEW_TIER — skills/** is a published contract surface and the budget judgement is the hard part, not the diff.

---

Generated by Claude Code

[huangyiirene]

Unlock scan (R+92): one of two dependencies discharged — card stays pm:blocked.

#14253 has landed (translateDataset is now in the generated spec surface on origin/main @ f7d92d3), so the exhausted Blocked-by: #14253 line is deleted and the body now states the claim is false today rather than prospectively. ⚠️ Blocked-by: #14315 remains — that DIET card on the same file is still pm:queue and open — so this stays blocked, which is the whole point of encoding the collision as a second line last round rather than leaving it in prose.

⇒ Ordering is unchanged: #14315 first, this correction second. The published skill now teaches something false, which is unpleasant, but landing a one-line correction into a file that is about to lose ~1,685 tokens loses one edit or the other silently — and the rewrite is the larger, harder-to-redo half. ⛔ Whoever takes #14315 should not fold this correction in on the way past: the budget ruling applies to both, and mixing them makes the net token change unreadable.

Everything in the R+91 grading note stands, in particular the budget fence as a dispatch condition (net token change near zero; ⛔ three new key families do not license three new sections; stop and report rather than self-authorising a bigger budget) and the governed-surface landing (push, open the PR, stop — never merge, queue, arm auto-merge or flip out of draft).

---

Generated by Claude Code