CoManTrack: Conflict-aware and manifold-adaptive RGBT tracking.

RGBT tracking relies on fusing visible and thermal signals, but severe degradation in one modality often pollutes the shared feature space, causing cascading errors and long-term temporal drift. Existing methods struggle with active thermal distractors and suffer from over-smoothed representations and poisoned memory queues during online adaptation. To address these issues, we propose CoManTrack, a spatio-temporally decoupled framework that isolates spatial purification from temporal adaptation. Spatially, we introduce Conflict-Driven Directed Dampening (CD3) to dynamically suppress high-energy thermal noise based on semantic divergence, safeguarding fragile visible structures. The purified features are then enriched by a Tri-Statistic Feature Envelope (TSFE), which captures fine-grained multi-granularity cues through normalized extreme statistics. Temporally, we develop Test-Time Optimal Manifold Adaptation (TOMA) to filter redundant historical states based on distributional novelty. This maintains a diverse and compact target manifold, ensuring a stable geometric foundation for dynamic adaptation. Extensive experiments on the LasHeR, RGBT234, and GTOT benchmarks demonstrate that our CoManTrack achieves competitive and balanced performance against state-of-the-art methods, validating its robustness against severe cross-modal degradation.