Why do different attributes carry unequal explanatory weights in causal explanation? Although both momentum and colour comply with conserved-quantity transmission in physical interactions, momentum is universally recognized as a core causal factor, whereas colour is dismissed as a trivial epiphenomenon. This long-standing puzzle, formalized as Salmon’s blue chalk counterexample, has perplexed the philosophy of science for nearly three decades. This paper argues that the persistent paradox arises not from inherent defects in existing causal mechanism theories, but from cross-level cognitive misalignment. Treating human cognition as a resource-bounded information compression process, we identify four naturally emergent discrete cognitive stable states: Primordial Percept (L1), Symbolic Percept (L2), Formalised Percept (L3), and Algorithmic Percept (L4). We demonstrate that Salmon attempted to resolve an L1-originated ontogenetic puzzle through L3–L4 formal and algorithmic tools. The primitive bias for explanatory relevance stems from evolutionary embodied filtering, rather than physical conservation rules or statistical correlation. We further formulate six fundamental relations between cognitive stable states: Compression (P1), Generation (P2), Realisation (P3), Dependence (P4), Emergence (P5), and Rollback (P6). Compression and rollback form the complementary dual wings of cognitive dynamics. As an intrinsic compensatory mechanism for unavoidable information loss in compression, cognitive rollback dismantles rigid stable configurations and restores cognitive freedom, rather than merely reversing compression. Together, the six propositions construct a unified ontogenetic framework of cognition, which fundamentally grounds the theory of causal explanation. Furthermore, this framework demarcates the structural boundaries and upper capacity limits of artificial causal reasoning in AI systems.