I realise there are many different pure states that can combine into the same mixed state, so "disassembling" may produce infinitely many solutions.
This is indeed true. The decomposition of a density matrix into a classical distribution is highly ambiguous (as I explained at depth in this previous thread).
However, the loose idea of the "most probable" pure state that helps constitute a given density matrix can indeed be given a unique, rigorous understanding. In essence, a density matrix $\rho$ is a self-adjoint, positive-semidefinite, trace-class operator such that $\mathrm{Tr}(\rho)=1$. The first of those qualifiers is the key, because as a self-adjoint operator you are guaranteed* an eigenvector decomposition of the form
$$ \rho = \sum_n p_n |\phi_n \rangle\langle \phi_n|.$$
There the $|\phi_n\rangle$ are guaranteed to be orthonormal, and they are unique (up to degeneracies). Moreover, the probabilities $p_n$ satisfy $p_n\geq 0$ and $\sum_n p_n =1$.
This structure then lets you take the $|\phi_n\rangle$ with maximal $p_n$ as the "most probable" pure state within the probability distribution.
Moreover, because that maximal $p_n$ coincides with the operator norm of $\rho$, that "most probable" state can also be defined as the state $|\phi\rangle$ such that
$$
\langle\phi|\rho|\phi\rangle
$$
is maximal over the entire Hilbert space (as can be shown by expanding this $|\phi\rangle$ over the eigenvector basis and then calculating this expectation value).
*modulo mathematical complications in infinite dimensions which require more work, but the assumptions are very strong (the operator is bound, and likely compact) so you get a strong spectral theorem in return.