Preprint. Completed manuscript.The Appearance of Meaning (AoM) showed that contextualized token-in-context states causally controlmeaning-like preference margins in transformer models. This paper asks whether that already-localized con-textual control is better recovered in the raw residual stream or in a sparse learned feature basis. In Gemma2 2B, we compare matched raw residual and sparse autoencoder (SAE) interventions at shared resid_postsites under hard invariants and endpoint-native accounting. On lexical disambiguation (DISAMB), layer-4SAE patching yields a larger mean donor-directed effect than raw patching and higher RMS-based effect-efficiency under scored-position disturbance accounting, although the paired SAE-over-raw difference is onlydirectionally positive and its 95% confidence interval includes zero. FP64 decomposition shows that thepaired difference is captured by expected-set and other-set candidate log-probability terms; diagnostic Δ log𝑍is zero in the primary single-token regime. Matched-rank PCA recovers an intermediate effect, random or-thogonal projections recover little, and an SAE reconstruction/residual split shows that the early result isnot well explained by generic compression or by a simple monotonic reconstruction-fidelity account. Thisearly, task-conditional pattern weakens or reverses at later layers and does not generalize uniformly tocounterfactual preference (CF) and discourse coherence (COH) tasks, where raw residual patching is nearparity or stronger on effect magnitude. We therefore conclude that AoM-relevant control is basis-sensitive,layer-dependent, and task-heterogeneous rather than uniformly sparse. The result extends AoM from causallocalization to representational basis without implying meaning proper or sparse semantic atoms.