Model editing updates factual knowledge in large language models (LLMs) without retraining, but simultaneous large-batch edits can degrade generalization and locality as edit interactions accumulate. We study this problem through a multi-scale analysis of feed-forward network (FFN) and multi-head self-attention (MHSA) geometry. Weak FFN–MHSA token-level alignment can coexist with non-trivial distributional coupling, while Jacobian-inferred FFN local-response geometry provides a distinct proxy for locally supported response directions. Based on this distinction, we propose PRISM, which replaces unconstrained residual-target fitting with controlled residual correction within edit-conditioned, Jacobian-inferred local-response subspaces. On the standard CounterFact large-batch benchmark, improves the baseline’s Score by 1.3–1.7 points and Generalization by 1.7–1.8 points at 10k edits, with larger Generalization gains in beyond-10k scaling tests at 20k edits while maintaining competitive Specificity.