refactor: give sortEndpointStats a strict comparator

[FrameAutomata]

Small hygiene change found while reviewing the percentile work in #319. Not a bug fix — see the measurement below before spending review time on it.

What

sortEndpointStats built its descending order by negating the ascending comparison:

if desc {
    return !less
}
return less

For two rows that tie on the sort key, less is false in both directions, so !less returns true for both (i, j) and (j, i). That is not the strict weak ordering sort.Slice documents a requirement for, and the results are then formally unspecified.

Descending is now the same comparison with the operands swapped, which is strict in both directions:

if desc {
    i, j = j, i
}
switch orderBy {
case "count":
    return stats[i].Count < stats[j].Count
...

Identical change in sqlite/ and duckdb/ — the two copies of this function were byte-identical before and after. ClickHouse sorts in SQL and is unaffected.

Why this is hygiene and not a fix

I tried to make the old comparator actually misbehave and could not:

trials	2400
slice sizes	8 – 20,000
distinct values	1 – 100 (heavy ties)
input patterns	random, ascending-modulo, descending-modulo
misordered results	0
elements lost or duplicated	0

Go's pdqsort tolerates the >= comparator in practice today. So nothing is broken, and no user-visible symptom is being fixed. What the change buys is not depending on tolerance that sort.Slice explicitly does not promise — a future Go release is free to change it.

I had initially described this as tied rows reordering between page loads. That was wrong, and the probe above is what corrected it.

No test

None can distinguish the two versions: any behaviour a test could assert holds for the old comparator too, which is exactly what the probe measured. Adding one would be coverage-shaped code that pins nothing.

Precedent

sqlite/task.repository.go already uses the strict form (> for desc, < for asc) in all four of its sort sites. This brings the two endpoint repositories in line with the pattern already in the package.

Verification

go test -race -count=1 ./app/..., the DuckDB-tagged repository suite, go vet ./app/..., gofmt — all clean.

Reasonable to close this if you'd rather not carry the churn. It is 6 lines of behaviour-preserving change against a risk that is currently theoretical.

🤖 Generated with Claude Code