Reject non-finite weights in sample_weighted

[teddytennant]

• Added a CHANGELOG.md entry

Redo of #1812 along the lines you suggested there:

If anything, non-finite weights should be considered an error; we certainly don't want this approach.

So this drops the separate-reservoir approach entirely and just rejects non-finite weights.

The problem

seq::index::sample_weighted accepts infinite weights but cannot sample them correctly. In A-ExpJ an infinite weight gives the key ln(r) / inf == -0.0 for every r, so infinite-weight items cannot be ordered against each other, or against finite ones. Once the reservoir's minimum key is -0.0, the skip value x = ln(r) / -0.0 is inf and the replacement branch is never taken again:

let weights = [1.0, 1.0, f64::INFINITY, f64::INFINITY, f64::INFINITY];
let v = rand::seq::index::sample_weighted(&mut rng, 5, |i| weights[i], 2).unwrap();
// always [0, 1]: the two *finite*-weight items — i.e. the least likely ones

So the failure mode is a silently wrong result, not an error.

The change

Reject any weight that is not finite and non-negative, which is where NaN and negative weights are already rejected. -0.0 is still accepted (case 8), and sampling among finite weights is untouched.

Behavior change

test_multiple_weighted_edge_cases case 6 previously asserted the opposite — that a +inf weight is accepted and its item always selected. That assertion is now WeightError::InvalidWeight. I've also added a case 6b for an infinite weight appearing after the first amount items, which is the path that produced the wrong answer above.

This is scoped to seq; WeightedIndex is untouched.

cargo test --features std, cargo fmt --check and cargo clippy --features std,alloc --lib are clean.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[dhardy]

Note first that f64::INFINITY is really just a placeholder for a value "too large to represent". Mathematically speaking, 0 / 0 and inf - inf could be anything.

Our current approach breaks down if more than amount infinite weights exist, ignoring all items beyond the first amount with non-finite weight. But since two positive non-finite weights cannot be compared, sampling from all non-finite weights uniformly is no more correct than ignoring all non-finite weights beyond the first amount.

Rejecting any non-finite weight is the simplest solution, but a regression since the current approach is fine with up to amount non-finite weights. I believe there is a better compromise available (PR incoming).