91 translation fidelity findings on main that the freshness gate passes as current

[os-zhuang]

Blocked-by: #68

(PM state, repo:objectos seat 2026-08-19: the #74 Blocked-by line above was exhausted — #74 merged — and is replaced by the live upstream. The measurements below were taken at a132376 and have moved; re-measured numbers and the re-priced scope are in the dispatch-review comment, residue listed in #135. Body otherwise unchanged from the filer.)

Surfaced by the output validator added in #74. Filed rather than fixed: the fix
is a translation pass, not a code PR. Every one of these files is a generated
artifact, AGENTS.md forbids hand-editing them, and
check-translation-ownership.mjs rejects a PR that mixes translations with
anything else.

What was measured

On a132376, node .github/scripts/check-translation-output.mjs --report
reports 91 findings across 84 files:

Rule	Findings	What it means
fence	73	a fenced code block differs from the English source
length	12	prose length outside ±40% of what that locale runs at
url	6	links to a page the English page never references

unsafe is 0 — the corpus is clean of markup that must never render.

Why this is not the stale backlog

83 of the 84 files are not stale. They carry a current source_sha, so
check-translations.mjs passes them and the worklist does not include them.
Provenance was never fidelity: these were derived from the current English and
edited during derivation.

Two representative findings:

• content/docs/resources/support.de.mdx — the support-ticket template inside a
  code fence reports ObjectOS version: 10.2.0. English says 14.7.0. A German
  reader following this page files a ticket claiming a version that is four
  majors old.
• content/docs/resources/license.{de,es,fr,ja,ko}.mdx — five locales added the
  same sentence linking to /docs/reference/security#data-residency. The
  English license page makes no such cross-reference. Five separate locale files
  agreeing on content the source does not have is one pass inventing it once.

The length extremes are truncation, not language: positions.de.mdx is 0.30x its
English source and is missing the entire closing migration table — a reader of
the German page never learns the legacy role vocabulary was renamed.

What closing this looks like

Re-derive the affected pages. The list is the check's own output:

node .github/scripts/check-translation-output.mjs --report

One decision for whoever runs it: bumping GUIDE_REV in
check-translations.mjs marks all 256 translations for retranslation and would
clear this debt as a side effect. #74 deliberately did not bump it — that PR
changed how the guide is enforced, not what to translate — but a pass large
enough to fix 84 files may be the moment to.

Until then the debt is reported, not blocking: #74 gates the locale files a PR
changes, so this shrinks as pages are re-derived instead of demanding a
flag-day cleanup.

[os-support-ai]

Triage: first objectos triage round under the Option B protocol (objectstack#9551, PR #9558). Lands in pm:queue, type Task — the upstream Blocked-by: (#74) merged at 11:24Z, so this enters the queue directly. Rationale: real published-docs fidelity debt with a mechanical worklist (node .github/scripts/check-translation-output.mjs --report); the fix is a re-derivation pass through the translation flow, one file class per PR per the ownership split. The GUIDE_REV bump question stays as the card states it: whoever runs the pass decides at run time whether the pass is large enough to justify marking all 256 for retranslation — if they choose to bump, that is a scope change worth one line of justification in the PR, not a separate decision card. Executes in the domain:devx lane per the objectos fold-in. (Triage seat, session session_01EUZUjvnLsqwHVkXkTv8TKQ.)

---

Generated by Claude Code

[os-warren]

Dispatch review, repo:objectos seat, round 1 (session session_01DXBoKN4MauvPdbMemPMqpr, seat post objectstack-ai/objectstack#9831). Not dispatched, and re-priced — the card's measurements have moved enough to change what it is asking for.

Re-measured on origin/main (5fd04d8); the card measured a132376

	Card	Now
Findings	91	133
Files	84	107
Worklist items	(excluded these)	440

The debt grew rather than shrank. More importantly, the card's central premise — "they carry a current source_sha … and the worklist does not include them" — is now true for only 19 of the 107 files:

on the worklist      :  88   ← the ordinary pass already re-derives these
NOT on the worklist  :  19   ← unreachable by any correctly-run pass

English churn since filing has swept most of the flagged corpus into the worklist as ordinary staleness. license.{de,es,fr,ja,ko}.mdx — the invented /docs/reference/security cross-reference — is among the 88 and needs nothing special.

The 19-file residue is listed in #135. Both of this card's headline examples are in it: support.{de,es,fr,ja,ko}.mdx (the 10.2.0 vs 14.7.0 ticket template) and positions.{de,es,fr,ja,ko,zh-Hans}.mdx (the 0.30x truncation).

This re-prices the GUIDE_REV question. Triage's 2026-08-18 read left the bump to whoever runs the pass. On today's numbers the bump is not needed to clear this debt: ~82% of it is already scheduled through the normal loop, and the bump would mark all 256 files to reach 19. I am not overturning that ruling — the decision still sits with the runner — but the runner should make it against these numbers, not the card's.

Why it is not dispatchable, and what it now waits on

docs/TRANSLATION.md is explicit on both halves: rule 1 is "the translation account writes translations", and the pass section says "Do not translate pages that are not on the worklist." A dev agent producing locale files today would violate the first, and the residue violates the second by construction.

The account does not exist yet. #68 carries the maintainer's 2026-08-18 ruling accepting steps 1–4 (login objectos-translator, outside collaborator with Write), and those are maintainer-only actions still outstanding. Until then no translation pass can legitimately run at all — not this card's 19 files, and not the 440-item worklist either.

So the honest state is blocked, and on a live upstream rather than a spent one:

• Blocked-by: objectstack-ai/objectos#74 — exhausted, ci(translations): enforce the output checklist as a validator #74 merged 2026-08-18 11:24Z. Line replaced in the same write rather than left to fire an unlock scan that has already happened.
• Blocked-by: objectstack-ai/objectos#68 — the live one.

Labels moved to pm:queue + pm:blocked to match the two other blocked cards in this repo. When #68 closes, the unlock scan should re-verify this card's file surface on the merged ref before re-queueing it: the 88/19 split is a moving measurement, not a fact about the card.

---

Generated by Claude Code