fix!: algorithm for oneOfWeighted has better distribution

[ctholho]

fixes #46

[justinvdm]

hash() gives us back a 52-bit int, so I think we might be truncating the range with the / 1_000_000, which could affect the quality of the results. Could we do id / Number.MAX_SAFE_INTEGER here instead?

edit Turns out I was wrong here, reply below this one has more details.

Suggested change

	const prob = (id % 1_000_000) / 1_000_000
	var prob = id / (2 ** 52)

[justinvdm]

It turns out my suggestion fails the tests (the diff threshold is greater than 6.5%), whereas yours doesn't. I can't understand why. So lets go for your change - scratch my suggestion :) Can you help me understand why 1M works better?

Only suggested change then is to use es5 (context here).

Suggested change

	const prob = (id % 1_000_000) / 1_000_000
	var prob = (id % 1000000) / 1000000

[ctholho]

I tried today again many different settings and don't understand it at all. I figured in the end it doesn't really matter because only the probabilities the user defined in the array are important. It's unlikely they would define values like 0.000004 and want to run the function more than a million times. Right now, somehow 1 million gave the best results with many different things I tried.

On the other hand: If we break backwards compatibility we should do it right the first time. So I'll take some more time to look into alternatives.

[justinvdm]

When I created this project 5 years ago, I had the brainwave of writing it only in es5 to avoid a build step. I'm not sure if that decision was the right one, but unfortunately it means for now we need to write the code as es5 for linting to pass. I'm sorry about this, I realise it is absurd to ask someone to write es5 in 2025 :)

Suggested change

	let cumulative = 0;
	for (const [probability, value] of samples) {
	var cumulative = 0;
	for (var i = 0; i < samples.length; i++) {
	var probability = samples[i][0];
	var value = samples[i][1];

[ctholho]

alright. happy to comply.

[justinvdm]

Suggested change

	t.assert(diffBetween(sums.blue / n, 0.3) <= DIFF_THRESHOLD)
	t.assert(diffBetween(sums.green / n, 0.2) <= DIFF_THRESHOLD)
	t.assert(diffBetween(sums.blue / n, 0.1) <= DIFF_THRESHOLD)
	// maybe yellow and fuchsia here too?

[justinvdm]

Looks great :) Just a few minor comments.

[justinvdm]

@ctholho I should really add contribution guidelines and CI for this project. For now, can you do these to run checks for your changes and update the docs and tests?

# checks
yarn lint
yarn test

# update snapshot tests
yarn test -u

# update docs (when code examples yield different results, e.g. for breaking changes)
yarn build:docs