<p>Qualitative failure cases across preprocessors.</p>

Stemming reduces vocabulary by up to 83% relative to the normalization-only baseline, proportionally reducing encrypted-index size and transmission cost. We find that per-query search latency is driven primarily by result-set size rather than vocabulary size: the normalization-only baseline is fastest per query because it matches the fewest documents, while preprocessors that broaden retrieval incur higher latency. At scale, all five configurations — including the normalization-only baseline — operate successfully to 45,500 documents, and the result-set-size effect on latency persists.

In terms of search quality, light stemming and morphological segmentation achieve the best trade-off between retrieval precision and recall, while root-based stemmers sacrifice precision without commensurate recall gains. Based on these findings, the framework provides actionable, evidence-based guidelines mapping deployment requirements — memory scalability, search latency, search quality, and retrieval breadth — to concrete preprocessor choices for Arabic DSSE system designers.</p></div>