- Shipped
- September 16, 2026 at 2:24 AM UTC
- Author
- Kamo
- Commit
- 6895ea8
The last pass read the English and pronounced it clean. It was: the double hyphens were all in the TRANSLATED files, where an em dash comes back as `--`. The Chinese contact page said "星期一 -- -- 星期五" for "Monday – Friday". Fourteen of those were live, and nothing in the build would ever have caught them. **************** now fails the build on a double hyphen, on a glued em dash (word—word), and on any message that does not parse as ICU — across all 22 locale files, not just English. It runs in prebuild and test, so it reads the real dictionary in CI, which is the only place a bad sync can be caught. Five ICU plurals came out of the source rather than being repaired, because repairing them fixes one sync and no other: the service translates `plural`, `one` and `other` as prose and breaks them again every time. `andMoreFeatures` is a flat "+{count} more"; the KB result count, the hunt duration, and the two signal strings split their singular into a sibling key that the component picks. extract-collections stops sending `keywords` to the translator, for the same reason it already withholds slugs and vendor names: 856 of 903 keyword arrays had come back as mistranslated product names. And the numbers that contradicted each other. The demo page advertised "50% Cost Reduction" with no source anywhere, "10min Avg Setup Time" against "about a minute" everywhere else, and "30+ Integrated Tools" against the homepage's 100+. Six competitor pages claimed SOC 2 Type II while every first-party surface says SOC 2 aligned, and one asserted ISO 27001 that another put on the roadmap. Dictionary side is 000895df.