5bb4fa66-222d-478d-96c5-656ae4330cb8

Review stats.js for correctness bugs. There are real defects. Fix them and extend stats.test.js to cover what you fixed, then run the tests with 'node stats.test.js' and show me the output.

I'll start by reading the files.

Two real defects in `median`, plus an undefined-behavior case in `percentChange`.

Tests pass: ``` $ node stats.test.js ok ``` ## Defects found and fixed **1. `median` used a lexicographic sort** (`stats.js:3`) `Array.prototype.sort()` with no comparator coerces to strings, so `[1, 9, 10]` sorted to `[1, 10, 9]` and the median came back as `10` instead of `9`. In a billing rollup this silently skews any dataset whose values cross a digit-count boundary. Fixed with `(a, b) => a - b`. **2. `median` never averaged for even-length input** (`stats.js:4-5`) It always returned the upper middle element, so `[1, 2, 3, 4]` returned `3` instead of `2.5` — biased high on every even-sized set. **3. `percentChange` divided by a zero baseline** (`stats.js:9`) `percentChange(0, 10)` returned `Infinity` and `percentChange(0, 0)` returned `NaN`. `Infinity` propagates through arithmetic and formats as `"Infinity%"` rather than failing loudly. Now returns `NaN` for both — change against a zero baseline is genuinely undefined. I also made empty input explicit: `median([])` returns `NaN` rather than `undefined`, so it fails the same way as the other undefined cases instead of poisoning downstream arithmetic with `undefined`. Two judgment calls worth confirming, since they're the kind of thing a caller may already depend on: returning `NaN` (rather than throwing) for empty input and for a zero baseline. If the rollup would rather surface these as errors, throwing is a one-line change in each. ## Test coverage added `stats.test.js` now covers numeric-vs-lexicographic ordering (`[1, 9, 10]`, `[100, 25, 9]`, negatives), even-length averaging, unsorted input, non-mutation of the caller's array, empty input, and the `percentChange` zero-baseline cases alongside normal increase/decrease/no-change. Each new assertion fails against the original implementation.