Re-measured on 1,000 labels across the whole 20,000-row split: 3 to 50 forms per label, 0 of 5,423 multi-form words begin with a FINAL form, and reshape(NFKC(stored)) reproduces the stored string for 1,000 of 1,000. Ships verify_yousefmd.py so every figure is re-derivable from one sample.
Correct the visual-order claim about Yousefmd/arabic_ocr_dataset; add a corrections policy
Only the shaping step ran there. Words are in logical order and plain NFKC recovers them. The repair advice previously given was wrong in the harder direction. Also: forms-per-label range 4-54 (was '23 to 29'), and a stated way to dispute any finding.
Measure the tokenizer cost: one stray glyph makes an Arabic word [UNK] on AraBERT and mBERT
