i18n: Opus round 26 (134 files) + 88 fixes + fix unreproducible Tier-4 rubric

[JanLahmann]

Round 26

Tier-4 Opus deep-review, seed 20260711 (--max-leaks 2 --drift-focus --per-locale 8 --exclude-reviewed), 1237-file unreviewed pool. 134/134 reviewed, 0 errors. Coverage 3155 → 3289 / 7276 (45.2%).

Verdicts: PASS 46 / MINOR 63 / FAIL 25 — far off rounds 23–25 (~105/27/2). That anomaly is what this PR ultimately explains.

88 fixes (17 locales)

All non-PASS findings remediated by Sonnet: 88/88 FIXED, 0 could_not_fix. A 5-way adversarial gauge (each skeptic told to refute, defaulting to refuted) returned 5/5 real, 0 refuted, so the findings were trustworthy even though their severity labels were not.

Representative:

• tl sinasakal (strangling the next generation) → sinasanay (training) — fluent, grammatical, and a total inversion
• ar الرمز الغليظ ("thick code") → transliterated رمز الغروس, restoring the gross = 144 wordplay that motivates [[144,12,12]]
• ar hedging restored on "should explain"; duplicate untranslated EN bullet removed

Gates: lint 0/88 fail · 0 source-hash edits · 17/17 locales 0-stale/0-critical · known-mistranslations 0 · wrong-language 0 · smart-quotes-in-JSX 0 · docusaurus build PASS on all 17 locales.

The rubric fix (the important part)

The Tier-4 verdict was not reproducible, because the prompt contradicted itself. Under --drift-focus:

• preamble: an isolated awkward phrase in a meaning-faithful file "is a PASS"; "do NOT spend your read hunting leaks/polish/consistency"
• DEEP CHECKS: "flag translationese, calqued English syntax, stiff/robotic phrasing — even when grammatically correct"
• rubric: MINOR_ISSUES = exactly those non-misleading slips; FAIL = "any ... wrong domain terminology"

Both were live at once, so the verdict depended on which half the agent followed.

Measured on 20 byte-identical files (0 changed since round 25, 0 overlap with the fixes above):

Instruction set	Result
preamble live (round 25, Jul 5)	20/20 PASS
preamble absent (same files, today)	1 PASS / 12 MINOR / 7 FAIL

~95 points of swing on identical bytes.

Fix: one discriminator — would this MISLEAD a learner?

• FAIL = yes: drift/inversion/hallucination, a term that teaches a false concept, raw MT throughout
• MINOR = no, but a native editor would still change it: calque, imperfect-but-recognizable term, within-file split, dropped qualifier, register slip
• PASS = no, and nothing an editor would change

Plus binding calibration rules that stop silent re-scaling (a single imperfect term is MINOR never FAIL; within-file inconsistency is MINOR however often it recurs unless the concept becomes unidentifiable; stiffness never FAILs alone; defensible word-choice is not a defect). The focus note now defers to the rubric instead of overriding it.

Validation: two identical replicas under the fixed rubric

Metric	Result
A vs B agreement	12/14 (86%)
FAIL reproduced in both	1/1 (100%)
Disagreements	2, both at the PASS↔MINOR boundary

The signal we act on (FAIL) is now stable; residual noise is confined to the marginal PASS/MINOR call.

Consequence for the metric

PASS now means "screened, no misleading defect found" — not "certified clean." The 3,201 historical PASS verdicts were graded under the ambiguous rubric and are not comparable to post-merge verdicts. Per maintainer decision we are not re-reviewing them; fresh rounds target the ~4,000 unreviewed files instead.

⚠️ One open flag: under the fixed rubric, 1/14 previously-PASS files (ja/measuring-computational-cost.mdx) is FAIL-grade. Small sample, but if it generalizes it implies ~200 misleading defects sitting in the PASS set. Noted for a future decision, not actioned here.

🤖 Generated with Claude Code

[JanLahmann]

Reproducibility test completed (was 14/20, now 20/20)

The 6 files lost to a session limit have been re-run in both replicas. Final numbers, two independent replicas over the same 20 byte-identical files under the fixed rubric:

Metric	Result
Verdict agreement	18/20 = 90%
FAIL sets	identical (1/1 — ja/measuring-computational-cost.mdx)
Tally, replica A	2 PASS / 17 MINOR / 1 FAIL
Tally, replica B	2 PASS / 17 MINOR / 1 FAIL
Disagreements	2, both PASS↔MINOR (cs/serverless-run-first-workload, es/noise-learning)

The signal we act on is now 100% stable. Residual noise (~10%) is confined to the marginal "would an editor bother changing this?" call, which is an inherent judgement boundary rather than a prompt defect. So a single PASS/MINOR label still shouldn't be treated as authoritative — but FAIL is trustworthy, and that's what drives remediation.

Sharpens the open flag

Same 20 files, three graders:

Grader	Result
Round 25 (preamble-dominant)	20 PASS / 0 / 0
Round 26-era rubric (preamble absent)	1 PASS / 12 MINOR / 7 FAIL
Fixed rubric (both replicas)	2 PASS / 17 MINOR / 1 FAIL

The fixed rubric lands exactly where it should — between two artifacts. Real issues surface as MINOR; FAIL is reserved for genuinely misleading defects.

But note what it says about the PASS label: of 20 files round 25 certified PASS, only 2 are genuinely clean. Extrapolated over ~3,201 PASS verdicts that implies roughly 160 FAIL-grade (misleading) defects and ~2,700 MINOR-grade sitting behind the label. n=20, so the FAIL rate is 5% ±5.

This is the strongest evidence yet that PASS never meant "clean" — consistent with the "screened, not certified" decision, and it makes a future ~100-file audit of the PASS set worth considering if the label ever needs to carry weight.