Large language models recognize explicitly marked discourse relations far better than implicit ones; on an engineering-tutorial benchmark, current LLMs score below 9% F1 on implicit relations despite reaching 41% on explicit ones. We ask what it is, in the form of texts themselves, that carries implicit coherence through this bottleneck. Using the open Georgetown University Multilayer corpus (GUM), we triangulate three annotation layers—PDTB-style relation labels from the GUM Discourse Treebank, human-annotated signaling inventory from enhanced Rhetorical Structure Theory, and automatically computed distributional indices—at 4,178 causal, concessive, and contrastive relation sites (1,260 in 84 expository documents). Implicit sites differ from explicit ones in a cue-family-selective way: they carry 79% more syntactic construction signals and 5.7 times more morphological signals, and their arguments are more syntactically parallel (d = 0.44) and more often relaunched by demonstrative phrases (OR = 2.1), whereas semantic redundancy signals (meronymy, lexical chains) instead cluster at explicit sites. A signaling gradient across explicit, alternatively lexicalized, and implicit relations, and a sensitivity analysis within cross-sentence relations, support a selective-compensation account. Sense information, by contrast, is only weakly recoverable from the same indices. We interpret the pattern as markedness without specification: unmarked sites announce that a relation holds without revealing which one, leaving the content of the relation to inference—a division of labor that explains both how human readers bootstrap implicit coherence and why distributional machines do not.