Understanding and Mitigating Structural Forgetting in Fine-Tuned Time Series Foundation Models

While fine-tuning is the standard recipe for adapting Time Series Foundation Models (TSFMs) to downstream tasks, we reveal that it triggers a critical yet overlooked problem: structural forgetting. Distinct from widely studied point-wise error degradation, structural forgetting represents a fundamentally more destructive phenomenon: the erasure of universal structural patterns acquired during pre-training. Unlike point-wise errors, this structural collapse renders models incapable of supporting basic structure-dependent decisions (e.g., trend-based trading), a critical vulnerability that existing forgetting mitigation strategies fail to address. To bridge this gap, we first introduce a diagnostic framework equipped with a novel Temporal-Frequency Attribution analysis, mechanistically revealing that fine-tuning unnecessarily overwrites a sparse set of pattern-critical parameters. We then propose NeST, a post-hoc framework that precisely restores these components while compensating surrounding weights to preserve task adaptation. Extensive experiments across 3 TSFMs and 9 datasets demonstrate that NeST recovers an average of 85\% of structural awareness without sacrificing task performance, offering a new paradigm for structurally-preserved TSFM adaptation.