The product speaks eight languages: English, Hebrew, Arabic, Russian, Spanish, Portuguese, German and Japanese. Two of them read right to left. Supporting that is mostly discipline — logical CSS properties, no hardcoded strings, key parity across every locale file, enforced by a test.
What discipline does not catch is the handful of things that are correct in code and wrong on the page.
The emphasised word
Our headline puts its key phrase in italic. In Hebrew and Arabic there is no italic tradition and the fonts ship no italic face, so the browser fakes one by shearing the glyphs. To somebody who reads the script that does not look like emphasis, it looks like a rendering fault. The colour was doing the work in every language anyway; the slant was only ever the Latin half of the idea.
Digits inside a mirrored sentence
A clock time is left-to-right content even when the sentence around it is not. Without an explicit override the bidirectional algorithm renders 07:12 as 12:07 in Hebrew — in a section whose whole structure is a sequence of times. Invisible until somebody reads a real one.
The name that is not the brand
"Memorise" is identical in all eight locales. The companion’s name is not: she is אלי, إيلي, Элли and エリー. A single hardcoded "Ellie" in a phone mockup put Latin script inside a right-to-left interface, and it passed every test, because every test ran in English.
And one that is not a bug
Grammatical gender. Hebrew, Arabic, Russian and Spanish all inflect for the gender of the person being addressed, and a companion that gets it wrong is not making a typo — it is calling a woman "he" every sentence, forever. For a long time an unrecorded gender was silently treated as male across seventeen places in the code.
Unknown is now a third case rather than a default, and when she does not know she is instructed to prefer wordings that do not inflect at all. English cannot show you whether any of that works, because English barely inflects — which is why the test fixture for it had to be written in Hebrew.