This paper describes the participation of team SMMTech in the SMM4H-HeaRD 2026 Shared Task 2: Detection of Insomnia in Clinical Notes. We present a comparative architectural study exploring the friction between extractive token-classification models and generative Large Language Models (LLMs) in clinical span extraction, on the MIMIC-III Clinical Database. During the validation phase we established baselines using encoder-only transformers such as BERT, ClinicalBERT, BigBird and Clinical BigBird. For the official test phase, we deployed a 4-bit quantized generative hybrid pipeline using Llama3-Med42-8B to evaluate its multi-hop reasoning capabilities. While the generative pipeline achieved an F1-score of 0.4783 on Subtask 1 (Classification), it struggled with exact span matching on Subtask 2.In this paper we present the mechanical limitations of zero-shot JSON extraction and the necessity of decoupling clinical reasoning from character-level span extraction.