There is no broader context wherein natural selection can be considered to be an optimization process, that is a pernicious misconception of evolutionary theory. Fortunately, people with a computer science background have a distinct advantage towards correcting this fallacy, because their training affords them an understanding of information as a working concept that lay people rarely attain.
The key insight is that any algorithm implementation for a process which has an objective must, as an absolute minimal requirement, possess an encoding of that objective in its implementation. That is, a real representation of the goal must be in the process's make-up so that the goal can be pursued at all, because correct navigation requires assessing actions for whether they work towards the goal or not, and any such assessment requires meaningful reference to the goal. Without such a definition to refer to, differentiation between desirable and undesirable outcomes is impossible.
This goal encoding may be explicit (ie readily understandable by observers studying the implementation) or implicit (hard to parse), but either way, it must be instantiated in the make-up of the implementation, in some medium with the capacity to hold the goal definition, ie a way of storing the requisite number of bits within the implementation itself (or readily reading it from elsewhere, or constructing it from some combination thereof). This definition of the goal must be implemented in a manner that can be read and acted upon by the rest of the algorithm implementation, so that the system as a whole can pursue states that better match the goal. ie so that it can optimize.
With regards to evolution, how could nature select without having an idea of what it was selecting for? A reference definition of fitness must be available to nature if it is to measure each individual organism's fitness and select accordingly.
For a natural-selection-as-optimization-process algorithm implementation, there would need to be a component that encodes natural selection's optimization objective into the implementation's very make-up (or a ready way to read that goal from an external source).
What is the make-up of the natural selection algorithm's implementation? It is the entirety of nature itself, in whole and in part. Nature is literally everything in the universe, and literally anything in the universe, from the most massive galaxy to the smallest particle, can participate in natural selection events. And no part of nature, save for some animal brains, seems to contain a representation of a goal for natural selection.
Is it even conceivable that everything in the universe, down to the smallest particle, could encode a common goal? Does a volcano encode the goal of maximizing reproductive fitness for the populations living around it? Can a shower of cosmic rays encode the goal of making sure the creatures who's DNA it disrupts are the ones who should be removed from the populace? They don't appear to encode any such evolutionary goals, nor do they have the capacity to maintain any goal at all beyond following the physical laws of matter -- Volcanos are disordered piles of rock and churning lava, and cosmic rays are singular fundamental particles that are subject to wholesale transformation with every impact -- neither has any way of encoding a common objective for natural selection, nor is there evidence for them being able to collectively maintain one.
We can illustrate the paradox of an optimizing nature using your water molecule analogy. A collection of water molecules acting under a gravitational field will demonstrate downwards fluid dynamics which single molecules in space would not, but no matter how much H2O you put together, it will never spontaneously develop any concept of evolutionary fitness.
And yet a flash flood is a very real natural selection event that can reshape the genepool of a coastal town, but all the same it has no means of representing any goal of optimizing the population's fitness through who it drowns and who it spares; its just water. Flowing water performs natural selection, but it isn't optimizing for any goal, no matter how you try to spin it, because it has no way of maintaining a representation of a goal in its disordered and inconstant structure. It flows, yes, but it has no goal in doing so, its not pursuing any optimization objective, all the while it is a real instance of natural selection. It doesn't have or need any way of determining who is more or less fit than another, so how could it be optimizing for it? It's just flooding.
Whether its by deluge, an erupting volcano, a congenital heart attack, or a pack of rabid dogs, the processes making up natural selection events do not possess an encoding of a goal for natural selection. They do not possess the necessary information structure required to pursue a common optimization objective, and so they cannot be optimization processes in any meaningful sense.