Pole-Zero cancellation is one of those phantoms in control which are often either sidestepped or only treated informally. Personally, I've never deeply understood the reasoning behind some of the rules--in fact, there isn't even really a consensus on what that reasoning is.
For the purpose of this question, suppose $G$ is a SISO transfer function. Let $p$ be a pole which is exactly cancelled by a zero at the same location. We treat the system $G$ formally in the sense that we assume 100% fidelity to whatever we are trying to model. If you prefer to think of the system in terms of differential equations, this is equivalent to saying the differential equations perfectly model the system. The question is whether or not the system $\tilde{G}$ obtained from cancelling $p$ is distinguishable from $G$.
My answer is no by the following reasoning.
If $G$ has 100% fidelity then every system parameter is uniquely determined by $G$. It follows that if $H$ is another transfer function and $H(s) = G(s)$ for all $s$, then $H$ is mathematically indistinguishable from $G$ and by extension so too are their system parameters.
As regard $G$ and $\tilde{G}$ we have $$G(s) = \frac{s-p}{s-p}\tilde{G}(s)$$ by their definition. Clearly $G(s) = \tilde{G}(s)$ whenever $s\neq p$, so we need only check that $G(p) = \tilde{G}(p).$ To do so we may evaluate $$ \lim_{s\rightarrow p}G(s) = \lim_{s\rightarrow p}\frac{(s-p)\tilde{G}(s)}{s-p}, $$ which by L'Hopital's rule is equivalent to $$ \lim_{s\rightarrow p}\ (s-p)\tilde{G}'(s) + \lim_{s\rightarrow p}\ \tilde{G}(s) = \tilde{G}(p). $$ This calculation finds $G = \tilde{G}$, showing the system before and after pole-zero cancellation are equivalent.
I therefore conclude that exact SISO pole-zero cancellation under the assumption of 100% model fidelity has no affect on the system.
What are the thoughts of others?
Edit
Stelios has made some thoughtful points regarding the nature of $p$ as a removable singularity. Unfortunately some of the conventions I've used also appear to have caused confusion, and I will clear them up presently.
(1) When I write $G(p)$ I mean $\lim_{s\rightarrow p}G(s)$. I concede that it is unclear whether showing $\lim_{s\rightarrow p}G(s) = \lim_{s\rightarrow p}\tilde{G}(s)$ is equivalent to saying $G(s) = \tilde{G}(s)$ for all $s$, since, as Stelios has argued, it is unclear whether or not $G(p)$ can be assigned any meaning in this way.
(2) Supposing that two transfer functions $G$ and $H$ can be assigned meaning for every $s$ in a domain which they share, it is no more than a basic set theoretic statement to define $G =H$ iff $G(s) = H(s)$ for all $s$. In fact, if $A$ and $B$ are sets and $f$ maps $A$ to $B$, then WLOG we may regard $f$ as a subset of $A \times B$ consisting of the points $(x,f(x))$ for $x \in D(f)$. If $g$ is another function and $D(g) = D(f)$, then if $g(x) = f(x)$ for each $x$ in their domain it follows that $f$ and $g$ describe the same subset of $A \times B$. This is surely enough to say $f = g$.
I presume you consider rational transfer functions as typically employed in signal processing and control theory.
I think your reasoning is invalid due to the following conceptual error. You "prove" that $\tilde{G}(s)$ and $G(s)$ are "mathematically equivalent" (which, according to your definition, means that $\tilde{G}(s)=G(s)$ for all $s$), yet $G(s)$ is not formally defined at $s=p$. Therefore, a comparison between $G(s)$ and $\tilde{G}(s)$ at $s=p$ makes no sense.
What you are essentially asking is whether a function $G$, that is analytic in a domain $\Omega$ except a single (isolated) point $s_0$, can be "replaced" by another function $F$ that is exactly the same over $\Omega\setminus \{s_0\}$, yet appropriately defined at $s_0$ such that it is analytic over all $\Omega$ (including $s_0$). Why we want to do so? Because it makes life easier and, for the majority of applications, this replacement has not effect. For a trivial example, why would one use the function $f(s)=s/s$, defined everywhere expect $s=0$, instead of $f(s)=1$, defined for all $s$?
The above question is a standard one in (complex) analysis and is related to the concepts of singular points (and Laurent series). The theory says that in the case of $s_0$ being a removable singularity for $G$, we can replace $G$ by $F$, defined as $F(s)=G(s)$ for all $s\neq s_0$ and $F(s_0)\triangleq\lim_{s\rightarrow s_0}G(s)$.
In your example, $p$ is indeed a removable singularity with $\lim_{s\rightarrow s_0}G(s)=\tilde{G}(s_0)$, therefore we can "replace" $G(s)$ by $\tilde{G}(s)$ which is equal to $G(s)$ for all $s\neq s_0$, yet also analytic at $s=s_0$.