How we know evolution happened
Nobody watched the first vertebrate crawl onto land. The case for common descent does not rest on having watched anything, and it is stronger for that. It rests on many independent lines of evidence that had every opportunity to disagree with each other and did not.
The structure of the argument
The strongest scientific arguments are rarely a single decisive observation. They are several unrelated methods, each with its own assumptions and its own ways of being wrong, arriving at the same answer.
That is the situation here. Fossils, anatomy, embryology, genetics, geography and direct observation are independent in the sense that matters: an error in the physics of radiometric dating would not produce a matching error in how genomes are sequenced. When independent methods converge, the simplest explanation for the convergence is that they are tracking something real.
The pattern that would be easiest to break
Common descent makes a strong and fragile prediction: living things should fall into groups within groups, with no crossing over. Mammals within vertebrates, vertebrates within animals, each nested cleanly inside the last.
This is fragile because it forbids a great deal. A single organism combining features from two separate branches, in the way a mammal with feathers would, has never been found. Species genuinely intermediate between two branches exist in abundance, and they are always intermediate along one line of descent rather than assembled from unrelated parts.
The prediction gets stronger when genetics is added, because DNA was unknown when the pattern was first described from anatomy. Building the tree again from genomes could have produced a completely different arrangement. It produced substantially the same one, and disagreements between the two are the exception, informative, and usually explained by known processes.
The awkward evidence is the most persuasive
Traits that work well are weak evidence, because a well-designed thing might be well designed for any number of reasons. Traits that are clearly suboptimal are much harder to explain away, and they are everywhere.
The vertebrate eye has its nerve fibres in front of the light-sensitive cells, producing a blind spot that octopus eyes do not have. A nerve running from the brain to the larynx in mammals loops down around a blood vessel near the heart before coming back up, which in a giraffe means a detour of several metres to travel a short distance. Whales carry the remains of hip bones connected to nothing.
Each makes immediate sense as inherited history that could not be redesigned, since evolution modifies what is already there one viable step at a time and has no access to starting over. Each makes no sense at all as a fresh arrangement.
It is also observed directly
The claim that evolution is unobservable is simply out of date, and the examples are not obscure.
Antibiotic resistance emerges in hospitals on a timescale of years and is tracked in detail. Pesticide resistance follows the same course in agriculture. Beak dimensions in Galapagos finch populations have been measured shifting in response to drought and back again, within the working life of the researchers doing the measuring. Long-running laboratory populations have been followed for tens of thousands of generations with every ancestor preserved and revivable.
What is not observed on human timescales is large-scale divergence, for the same reason nobody watches a mountain form. The mechanism is observed, and the accumulated results are inferred from the evidence above, which is the normal structure of a historical science.
What a simulation can and cannot add
A world like this one is not evidence about the history of life on Earth, and it is worth being firm about that before saying what it is good for.
What it offers is the control that historical evidence cannot. Earth's history ran once, and the counterfactual is unavailable: nobody can rerun the Cretaceous with one variable changed. Here every run is exactly reproducible from its seed, a control condition can be run alongside an experimental one on identical starting conditions, and a threshold for what counts as a result can be registered before the run that tests it.
That answers a different kind of question. Not what happened, but what a stated set of rules produces, and whether a given outcome depended on the rules or on the run. The mechanics are on determinism and reproducibility, and the discipline around pre-registration and controls is on methodology.