Document conversion and import are a large but under-studied source of real-world noisy text, distinct from the social-media noise usually studied in normalization. We study missing-space artifacts in a low-resource, Arabic-script setting: an Urdu Qur’anic tafseer exported from InPage .inp to .txt and then to JSON/markdown, a pipeline that fuses adjacent words into single tokens (e.g., اورقرآن for اور قرآن, "and the Qur’an"). Restoring spaces is hard: Urdu’s cursive Arabic script does not require spaces at word boundaries, and the text interleaves Urdu prose with Qur’anic Arabic, proper names, and markdown, so a careless split can corrupt preservation-sensitive content. We cast this as precision-first normalization and present a confidence-gated pipeline that proposes only space-insertion edits, verifies a character-preservation invariant, and holds protected regions out of automatic correction. On a 1,000-row silver-labeled sample, the high-confidence auto-accept operating point reaches 96.9% strict precision on decisive cases, against 30.5% for an ungated function-word splitter and 39.7% for a frequency-only baseline, with 0 character-altering edits. This suggests that confidence gating is necessary before applying whitespace restoration to preservation-sensitive Urdu–Arabic text. As our labels are model-assisted silver rather than expert-gold, we treat the precision figures as provisional and release derived artifacts for reproducibility.