The definition of Evaluation of Filter in Sec.18.5.2 (Evaluation Semantics) currently says:

eval(D(G), Filter(F, P)) = Filter(F, eval(D(G),P), D(G))

Notice that the function on the right-hand side of this equation is invoked with three arguments:
i) F (the filter expression),
ii) eval(D(G),P) (the multiset of solution mappings that results from evaluating the child pattern P), and
iii) D(G) (the dataset and active graph).

Yet, when looking at the definition of this function,

Let Ω be a multiset of solution mappings and expr be an expression. We define:
Filter(expr, Ω) = ...

it turns out that the function is defined for only two arguments:
i) a filter expression and
ii) a multiset of solution mappings.

This inconsistency needs to be fixed.

By the way, it was not like that in SPARQL 1.0. Back then, the Evaluation of Filter used the Filter function with its two arguments.

I assume that the reason for passing the active graph as a third argument is related to the idea that the evaluation of Exists requires access to the active graph.

Fixing this issue properly is not entirely trivial I am afraid. All related definitions need to be extended such that D(G) can be passed as an additional argument, including the definition of how expressions are evaluated.