Why can't we prove decidability of $L= \{ \langle M \rangle : M$ accepts $\epsilon \}$ with a configurations graph?

Since every deterministic Turing Machine can be translated to a graph of configurations such that $$M$$ accepts a word $$w$$ iff there is a path from the initial configuration that matches $$w$$ to an accept configuration, we can define a decider $$D$$ for $$L$$ such that:

$$D$$ verifies $$\langle M \rangle$$ is a (deterministic) TM and then builds the configurations graph and checks if the initial configuration of $$\epsilon$$ is connected to an accept state (there are only finitely many) and returns true if there is and false if there isn't.

Though this seems simple and valid (to me), there is a well-known result showing that $$L$$ is undecidable. I can't figure out what is the flaw in the proof and will be happy for any help to figure it out

• How long does it take to build the graph? What if $M$ runs forever, how long will it take to build the graph? – Andrej Bauer Nov 27 '19 at 16:58

$$D$$ verifies $$\langle M \rangle$$ is a (deterministic) TM and then builds the configurations graph and checks if the initial configuration of $$\epsilon$$ is connected to an accept state (there are only finitely many) and returns true if there is and false if there isn't.
The above is just another way of saying: run $$M$$ on $$\epsilon$$ and wait to see if the machine halts or not. This, hovewer, is not deciding the language $$L$$.