What's wrong with my reasoning?
The box is of course supposed to be "closed" such that any observations that may be collected "at the outside" allow the conclusion that "the cat is inside, not outside", but under no circumstance any characterization of the cat in terms of "live or dead" or "having survived or having died".
A "meow from the box", however loud and shrill and indeed unavoidably audible "at the outside", is in this sense not supposed to be in any way indicative of whether or not the cat is about to expire;
a "thud on the ground" is meant to be attributable to a "cat having fallen dead" just as well as to a "cat having fallen asleep, but living on".
The suitably idealized measurement operators "fate of the cat" (with value range "having survided" or "having died") and "confinement of the cat" (with value range "inside the box" or "outside the box") are meant to be strictly incompatible operators. Observations which are sufficient for determining the "fate of the cat" should therefore in turn not in any way indicate whether the cat "has been entirely inside the box, or outside the box". Just as the state of "being in the box" can formally be expressed as superposition of "having survived" and "having died" (and correspondingly the state of "being outside the box", too), the state of "having survived" should be formally expressible as superposition of "being in the box" and "being outside" (and likewise the state of "having died".
Specificly, after lifting the lid of the box (and vigorously shaking it, just in case), and again closing the lid, the cat might in any case be found "outside the box";
and it might then count as "having survided" if, after again lifting and then closing the lid even repeatedly it would at least sometimes be found again "inside the box";
or correspondingly as "having died in the box" if, after lifting and closing the lid repeatedly, it would always be found "outside the box".
What counts as “observation” in Schrödinger's Cat
When considering any particular measurement operator one can distinguish the (sets of) observations which allow some particular value to be derived as "result of the measurement", by applying the operator under consideration on one hand, and on the other hand any other (set of) observations. Both cases may include observations which were collected in actual trials as well as hypothetical descriptions of observations.
As sketched above, the (actual, or hypothetically imaginable) sets of observations which count towards determining whether "the cat is inside" or "the cat is outside" are different from the (actual, or hypothetically imaginable) sets of observations which count towards determining whether "the cat had survided" or "the cat had died".
And of course there are many other (actual, or hypothetically imaginable) observations which have no relevance to either determination.
and why are superpositions possible?
The salient description of the state of the cat before the lid had been lifted is of course "the cat is inside, not outside", which is an eigenstate of the "confinement operator". Expressing this state instead as "half surviving and half dead, with a suitable phase between these parts" makes the relation to eigenstates of the "fate operator" explicit, and depends on the relations between the correspondingly relevant sets of observations.
So -- writing some particular state as superposition is certainly possible as an abstract expression. However, and here I believe to side with Schrödinger, it seems unpalatable to consider any formal superposition if it couldn't be primarily understood as eigenstate of a suitable operator.