ai-bookkeeper reads bank exports in many layouts. Probing parseCents with odd inputs showed one clear gap: an amount like ١٢٣٫٤٥ (Arabic-Indic digits, Arabic decimal mark) came back NaN, and a date like ٢٠٢٦-٠٣-١٢ came back null, so those rows were dropped or flagged.
The fix is one small function, asciiDigits, that maps Arabic-Indic (U+0660), Persian (U+06F0) and fullwidth (U+FF10) digits plus the Arabic decimal and thousands marks to ASCII before parsing. ASCII input passes through unchanged. arabic-digits.test.mjs covers amounts (negatives, parentheses, thousands) and dates (ISO, day-first).
What is not measured: this is checked on hand-written strings, not on a real export from a bank that prints native digits. A real one is still wanted.