Let $L_1$ be the language of all $(M, p, y)$ such that if we start $M$ (a DPDA with $\epsilon$ transitions) in state $p$ with a stack consisting of only $y$ (which is either a single symbol or nothing), then reading an empty string as input would cause $M$ to eventually return to state $p$, with a $y$ on top of its stack if $y$ was a symbol rather than nothing (if $y$ was originally nothing, we don't care about what's on the top of stack after $M$ returns to $p$; anything is fine).
claim: $L_1$ is decidable.
proof: modify $M$ as follows: make a copy $p_0$ of $p$ with the same outgoing transitions but no incoming transitions, and start the DPDA in $p_0$ with a symbol $y$ (or empty) in its stack. If $y$ is empty, then make $p$ the only accept state. Otherwise if $y$ is a symbol, then after creating $p_0$, modify the outgoing transition of $p$ with $y$ on the top of its stack to instantly go to a state $w$ that is now the only accept state (see note 1). Now use the standard CFL emptiness decider on this modified DPDA and we are done.
Note 1: If there is no such transition to modify due to determinism of $M$, i.e. $p$ only has one outgoing $\epsilon$ transition that either does nothing or pushes some symbol onto the stack, then do the following instead: delete the transition, add one that pops $y$ and goes to $w$ and also add trivial no-op transition that pops then immediately pushes back all other symbols that are not $y$.
Next, let $L_2$ be the language of all $(N, q, x)$ such that if we start the DPDA $N$ in state $q$ with a stack having only a single symbol $x$ (or empty), then with an empty input string, it will run forever without ever attempting to pop an empty stack. Note Sipser calls $(q, x)$ a "looping situation".
claim: $L_2$ is decidable (Sipser third edition problem 4.32 asks to prove this)
proof: take $N$ and delete all its input-reading transitions (leaving only $\epsilon$ transitions). Then, start $N$ in state $q$ with an $x$ on its stack, and for every state $p$ it enters, take its stack top $y$ and run the decider for $L_1$ on $(N, p, y)$. This procedure is guaranteed to terminate in either an accept for $L_1$ for some $p$ and $y$, or $N$ itself will terminate (either by attempting to pop an empty stack or by running out of $\epsilon$ transitions to follow). (Why? Suppose $N$ runs forever. Consider all of the pairs $(p,y)$ that it encounters at each step during execution. Since there are only finitely many states and symbols, by the pigeonhole principle, there must be some pair $(p,y)$ that is encountered infinitely often during execution; consider all the time points where we encounter $(p,y)$. Since the size of the stack cannot decrease forever, there must be some adjacent pair of such time points where the stack is larger at the second than at the first. Now consider the execution between between those two times, and find the time in that range when the stack is smallest. Let $p'$ be the state of the DPDA and $y'$ the value on the top of its stack. Then starting the DPDA at state $p'$ and just $y'$ on the stack and running forward will eventually return to state $p'$ with $y'$ on the top of the stack. Why? Well, we start $(p',y')$, then reach a point with state $p$ and $y$ on the top of the stack, then continues on to a point with state $p'$ and $y'$ on the top of the stack, and we never pop off the initial $y'$. Basically, we've decomposed the trajectory $(p,y) \leadsto (p',y') \leadsto (p,y)$ into its second part $(p',y') \leadsto (p,y)$ followed by its first part $(p,y) \leadsto (p',y')$, but with things chosen so that we can start with just the symbol $y'$ on the stack and nothing more, and it will never get popped off. It follows that $(N,p',y')$ is accepted by $L_1$, i.e., if $N$ runs forever, it will visit a pair that is accepted by $L_1$.)
Finally, we get to the problem in the main question: define $L_3$ as the language of $(M, x)$ where $x$ is some input string such that the DPDA $M$ halts on input $x$.
claim: $L_3$ is decidable.
proof: run $M$ on input $x$, and after every input symbol read, run the decider for $L_2$ on the current state and stack top: either $M$ itself will halt or the $L_2$ decider will return an acceptance. (Why? Suppose $M$ never halts normally. Then consider the configuration $(q,s)$ where $s$ is the shortest stack among all configurations, resolving ties arbitrarily. Since $s$ is among the shortest stacks, it will never pop below the element on top of $s$. Consequently, $(q,y)$ is accepted by the decider for $L_2$, where $y$ is the top of $s$.)
Note 2: it's probably not necessary to run $L_2$ decider to prove decidability of $L_3$: we can just run the $L_1$ decider instead.