I am studying the rough path theory of Lyons and am struggling to understand a heuristic claim made in several books and articles.
Let $(\Omega,\mathcal{F},\mathcal{F}_t, P, B_t)$ be a filtered probability space on which a standard Brownian motion $B_t$ is defined. Let $f:\mathbb{R}\to\mathbb{R}$ be a function (say, Lipschitz). Suppose $X_t$ is an $\mathcal{F}_t$-adapted process which solves the SDE $dX_t = f(X_t)dB_t$, that is, such that $X_t$ satisfies $$ X_t - X_0 = \int_0^tf(X_s)dB_s\quad\text{a.s.}, $$ where the the integral is in the sense of Itô. The Itô integral can then be obtained as the limit in $L^2$ of the Riemann-Stieltjes sum: $$ \sum_{0=t_0<t_1<\cdots<t_n=t}f(X_{t_i})(B_{t_{i+1}}-B_{t_i})\xrightarrow{\sup_i|t_i-t_{i+1}|\to 0}_{L^2(\Omega,\mathcal{F},P)}\int_0^tf(X_s)dB_s. $$ Many authors claim that rough path theory constitutes a "pathwise SDE solution theory": given a realization $B_t(\omega)$ of the driving Brownian motion, one can solve the corresponding "rough differential equation" driven by the corresponding realization of the Brownian rough path. In particular, the analogous statement is claimed to not hold for the Itô integral, because the above limit is only in $L^2$ (indeed, the above limit cannot be almost sure because Brownian motion has unbounded variation on any compact interval). On the other hand, a routine application of the Borel-Cantelli lemma shows the existence of a sequence of partitions along which the above limit is almost sure. My question is: in what sense does the realization of the Itô integrals $\int_0^tf(X_s)dB_s$ as an a.s. limit of Riemann-Stieljtes sums fail to constitute a "pathwise solution theory?" What defect of this "pathwise" construction does rough path theory mitigate?