In line with what is actually asked in the bold part of the question, I have taken the liberty of changing the title from constructive mathematics to computation. As it stood it was essentially a duplicate of another one, which already contains a lot about constructive but non-computational mathematics and so that would be the appropriate forum for further such discussion.
In fact Valery tagged the answer to the end of his first paragraph, namely @AndrejBauer's Efficient Computation with Dedekind Reals, which describes the ideas behind his Marshall calculator, whilst being based on our Dedekind Reals in Abstract Stone Duality and my Lambda Calculus for Real Analysis.
The setting for this answer is my ASD programme. I refuse to carry the baggage of obsolete programming languages (Turing Machines and Gödel Codings), any form of set theory (CZF, which I think is the setting for Lubarsky and Rathjen's paper, or even a topos) or the two bolted together (computable analysis in Weihrauch's TTE).
Valery's definition of a Dedekind cut is correct, but (to underline my rejection of any form of set theory) I write $\delta d$ and $\upsilon u$ for the down and up predicates instead of his $q\in L$ and $q\in U$ for lower and upper sets.
The predicate $\delta d$ says that (we so far know that) $d\lt x$ where $x$ is the real number being defined and similarly $\upsilon u$ means that (we so far know that) $x\lt u$.
Where this knowledge is incomplete, the pair $(\delta,\upsilon)$ satisfies all but the last axiom (locatedness) and is equivalent to an interval.
In particular, Dedekind cuts are two-sided. Taking the complement would amount to deducing knowledge or termination from ignorance or divergence.
There are plenty of examples of limiting processes that define "uncomputable" numbers that are best thought of as one-sided cuts. For example, the domain of definition of the solution of a differential equation is something that is only known from the inside; it is remarkable if we can obtain some knowledge from the outside, not that the boundary is "uncomputable" in this sense.
In the case where the cut defines a rational number, it should occur on neither side (Dedekind's choice was wrong here).
Defining the usual stuff in elementary real analysis is much easier and more natural using Dedekind cuts than with Cauchy sequences. Consider Riemann integration: $\delta d$ holds if there is some polygon (piecewise linear function) that lies below the curve and has area more than $d$; deducing the various properties from this is easy. This is gratuitously difficult if you insist in advance that the polygon have a particular shape of resolution $2^{-n}$ to make a Cauchy sequence.
The essential difference between a Dedekind cut and a Cauchy sequence as the representation of the solution of a mathematical problem is that Cauchy assumes that the problem has already been solved (to each degree of precision), whereas Dedekind merely combines all the formulae ("compiles" them into an "intermediate code", if you like), so that the problem remains to be solved.
This means that deriving a Cauchy sequence or decimal expansion from a Dedekind cut in general is just as difficult as solving a general real-valued mathematical problem. So I am under no illusion about the difficulty in implementing this proposal.
But it is a novel and so valuable way of looking at things.
The predicates define open subspaces of $\mathbb R$, so I would think that, in terms of recursion, they have to be $\Sigma^0_1$ or recursively enumerable. (I have thought a little bit about "descriptive set theory" (a name that needs to be changed) but haven't got very far with it.)
These predicates $\delta$ and $\upsilon$ are in the language of ASD and you will need to read the papers that I have linked to find out the details.
In the first instance, such predicates are built from $f(\vec x)\lt g(\vec x)$, where $f$ and $g$ are polynomials, using $\land$, $\lor$, $\exists$, recusion over $\mathbb N$ and universal quantification over closed bounded intervals.
We are given that $\delta d$ and $\upsilon u$ hold for certain rationals $d$ and $u$ and want to find some $d\lt m\lt u$ that is nearer to the result and whether $\delta m$ or $\upsilon m$.
In the case of $f(\vec x)\lt g(\vec x)$ we can use the (formal, interval) derivatives of $f$ and $g$ and the Newton--Raphson algorithm to do this. As is well known, this algorithm doubles the number of bits of precision at each step. This would justify the claim that the computation is efficient if we can extend it to more general formulae.
For more complicated logical expressions, Andrej developed a method of approximating these expressions with simpler ones. See his paper for details of this. This is quite closely analogous to methods of approximating real-valued functions by low(er) degree polynomials that have a very long history but have been developed in interval computation by Michal Konecny. Whilst Michal himself has used these computationally, they have not yet been incorporated into Andrej's calculator, but I hope that this will be done in future.
Computations like this with so-called "exact real arithmetic" are necessarily non-deterministic with regard to the Cauchy sequences tat they produce, which means that the existence of the completed sequence necessarily relies on Dependent Choice. This is entirely natural from a computational point of view and I don't see what use there is in discussing it as an issue of constructivity.
Best Answer
It's provable in ZF that every Baire-2 function is effectively Baire-2. It suffices to prove the following:
(ZF) There is an explicit function which maps each Baire-1 function $f: \mathbb{R} \rightarrow \mathbb{R}$ to a sequence of rational polynomials that pointwise converge to it.
The first step is to construct a sequence of reals $\langle r_n \rangle$ such that each $r_n$ codes a rational-valued function $f_n$ such that $|f_n(x)-f(x)| \le \frac{1}{n}$ for all $x \in \mathbb{R}.$ We'll construct $r_1,$ which immediately generalizes to other $n.$
Let $U_k$ enumerate the basic open sets. We use transfinite recursion to define a descending sequence of closed sets $\langle C_{\alpha}: \alpha<\omega_1 \rangle$:
If $C_{\alpha} \neq \emptyset,$ then $f \restriction_{C_{\alpha}}$ is continuous at some $x \in C_{\alpha}$ by Baire's Characterization Theorem. Then $x \not \in C_{\alpha+1},$ so $C_{\alpha+1} \subsetneq C_{\alpha}.$ Define a surjective partial map $g: \omega \rightharpoonup\beta = \{\alpha: C_{\alpha} \neq \emptyset\}$ by sending $k$ to the greatest $\alpha$ such that $U_k \cap C_{\alpha} \neq \emptyset,$ and a well-founded partial ordering $(\omega, \prec)$ by $i \prec j$ if $g(i) < g(j).$ Define $h: \text{dom}(g) \rightarrow \mathbb{Z}$ by $h(k)=\lfloor \sup_{U_k \cap C_{g(k)}} f \rfloor.$
Define $f_1: \mathbb{R} \rightarrow \mathbb{Z}$ as follows: for given $x,$ let $j=\min\{k<\omega: x \in U_k \wedge \forall \alpha (x \in C_{\alpha} \leftrightarrow U_k \cap C_{\alpha} \neq \emptyset)\}.$ Set $f_1(x) = h(j).$
Let $r_1$ be a real which encodes $(\omega, \prec)$ and $h.$ From $(\omega, \prec),$ one can determine the sequence $\langle C_{\alpha}\rangle$ and hence $g.$ Thus, $f_1$ can be determined from $r_1.$ This completes the construction of $r_1$ (and hence all $r_n$).
Let $r$ encode $\langle r_n \rangle$ (and thus $f$). By Shoenfield absoluteness relativized to $r,$ we have in $L[r]$ that $r$ encodes a Baire-1 function $\hat{f}.$ Enumerate the rational polynomials by $\langle q_n \rangle.$ A Stone-Weierstrass argument shows that there is $\langle n_k \rangle$ such that $\langle q_{n_k} \rangle$ converges pointwise to $\hat{f}$ in $L[r].$ Let $\langle n_k \rangle$ be the $L[r]$-least such sequence. Applying Shoenfield absoluteness upwards relative to a real encoding $(r, \langle n_k \rangle),$ we see in $V$ that $\langle q_{n_k} \rangle$ converges pointwise to the Baire-1 function coded by $r,$ namely $f.$ This completes the construction.
Some additional remarks: The above construction can be carried out in $Z_2.$ I don't know enough about subsystems of analysis to say more with confidence. By analyzing how much of $L[r]$ is really needed in the above, you can probably get this down to $\Pi^1_1-CA_0.$
As I showed in the comments, it is not provable in ZF that every Baire-3 function is effectively Baire-3, since if $\mathbb{R}$ is a countable union of countable sets, then every indicator function is Baire-3.