As Generative AI (GenAI) expands in language teacher education, a considerable gap persists between pre-service teachers’ declarative AI literacy and real-time execution under cognitive load (TPACK-in-Action). To address initial self-assessment inflation and uncritical algorithmic reliance, this study implemented an AI-self-assessment coordination loop (ASACL) framework. Utilizing an embedded single-case design (N = 7), Korean pre-service teachers were tracked across four lesson plan iterations (L1 to L4) featuring instructor macro-scaffolding and GenAI Socratic micro-scaffolding. Quantitative findings via a 33-point rubric revealed meaningful expert-evaluated quality gains (ML1 = 24.6 -> ML4 = 29.6, Gain = +5.0) alongside metacognitive calibration, narrowing the perceptual discrepancy gap from 1.58 to 0.25. Qualitative analysis of the discourse data showed three developmental themes: (1) Socratic counter-questions triggered metacognitive dissonance; (2) participants exercised epistemic agency through dialectic negotiation and philosophical rejections of context-blind machine outputs; and (3) external scaffolding faded, consolidating pedagogical sovereignty. These findings demonstrate that when GenAI is algorithmically constrained as a Socratic thinking mirror rather than an automated script generator, pre-service teachers evolve from passive algorithmic consumers into reflective, pedagogically capable designers. This study provides empirically verified design principles for operationalizing GenAI in language teacher education.