Running some experiments on this seems to show that the same platform (OS) returns the same results to the level that I need, however it is not the same across platforms I assume due to handling floating points numbers.
Example, multiple different 64 bit windows 10 machines are producing the same result on either Windows .exe file or HTML5 export in any browser.
Headless Linux server and HTML5 running on Linux produces the same results but those results are different to the windows builds.
Not compatible enough for the verification method I was looking for. maybe fixed point maths could do this but I think I'm heading down a rabbit hole doing that.
BUT it does seems to be reproducible enough that running the same inputs on the same architecture can be used to verify / replay results. I just need to sychronised that common server platform with the clients.
Unless there is another approach I'll run the simulation on the server so its always the same and use Zylann's video for some client side approximation to make things smooth and responsive.