As also discussed in the answers to What do noncontextual scenarios with no quantum model represent? and Can the Peres-Mermin square be reframed as a statement on the associated conditional outcome probabilities?, a contextuality scenario, as defined in (Leifer and Duarte 2020), is a triple $\mathfrak C=(X,\mathcal M,\mathcal N)$ with $X$ a set of outcomes, $\mathcal M$ a set of measurement contexts, and $\mathcal N$ a set of maximal partial measurement contexts.
A typical example of this is the Specker triangle as the contextuality scenario with $X=\{a,b,c\}$, $\mathcal M=\{\{a,b\},\{a,c\},\{b,c\}\}$ and $\mathcal N=\varnothing$. The idea is to model a situation where there are three possible outcomes, with labels $a,b,c$, and three possible measurement settings (or measurement contexts, if you will), corresponding to the elements of $\mathcal M$. Given any measurement context $M\in\mathcal M$, one and only one outcome is possible.
One can then consider value functions and states on a contextuality scenario. A value function is some $v:X\to\{0,1\}$ which represents a deterministic outcome assignment. It prescribes for any possible outcome whether that outcome will be found or not. A state is the probabilistic counterpart of a value function: a map $\omega:X\to[0,1]$ assigning a probability to each outcome (both $v$ and $\omega$ are required to satisfy a few constraints to make sure the above interpretations are sensible). For example, the Specker triangle above has no value function, but it admits the state $\omega(a)=\omega(b)=\omega(c)=\frac12$.
Another function one can consider is a noncontextual state, which is defined (c.f. Def II.11 in Leifer and Duarte 2020) as a state $\omega:X\to[0,1]$ such that $$\omega(a) = \sum_{v\in V_{\mathfrak C} } p_v v(a),$$ with $(p_v)_v$ a probability distribution on $V_{\mathfrak C}$, and where $V_{\frak C}$ was defined previously as the set of all value functions. States that are not noncontextual are said to be contextual.
What confuses me about this is the fact that, looking at the fact that all states are functions exclusively of the outcomes, how can any state be contextual? The paper explicitly mentions that value function are always noncontextual, being defined on the set of outcomes regardless of the context. Why doesn't the same argument apply to states? Where is the information about the context entering the definition?
To be clear, I'm aware of the fact that one can have contextual quantum scenarios, in the sense of not being possible to determine probability outcomes without specifying the measurement setting. My doubt is specifically about the fact that, using this definition of a "state", I'm not seeing where information about the context is to be fed.