[Narek-Khachikyan]

Upper-bound-only ranges such as <2.0.0 incorrectly fail subset checks against *, and <1.0.0 fails against 0.x.

Use the lowest admitted version for the omitted lower bound and normalize exclusive stable upper bounds when prereleases are excluded. Track empty subset branches independently so they do not change a union's result.

Verified with npm test: all 54 files pass with 100% coverage. Separate lint, template, and commit-message checks also pass.

References

Fixes #949.
Related to #919.

[1AbdullahButt]

I read this against ranges/subset.js on main. Three notes:

1. gt is now assumed to be set inside the dom loop. The if (gt) guard around the >/>= handling is gone, so higherGT(gt, c) and gt.operator run unconditionally. As far as I can tell that's safe: eqSet.size === 1 returns before the loop, lt-only now synthesises gt, and a sub with no comparators is the * special case handled earlier. But that's an implicit invariant spread over ~60 lines. A one-line comment at the loop (or an assert-style early return if !gt) would stop a future change to the EQ branch from turning into TypeError: Cannot read properties of undefined (reading 'operator').
2. The -0 normalisation isn't in the algorithm comment. You added "If LT is set without GT or EQ, use the lowest admitted version for GT" to the header, but the rewrite of an exclusive stable <X into <X-0 (when !includePrerelease) also changes semantics and is what makes ['<1.0.0', '0.x', true] pass. A matching bullet would keep the header comment accurate, since it is effectively the spec for this function.
3. sawNonNull moved into the outer loop. I agree with this: on main, a non-null earlier branch could make a later, entirely-null branch return false. It's a separate behaviour change from #949, though. The ['<0.0.0 || <0.0.0', '1.0.0', true] style tests cover it, but it'd be worth a sentence in the PR description so it isn't missed in the changelog.

The test table is thorough, especially the includePrerelease mirror cases and the loose <01.0.0 ones.

Static review; I didn't run the suite here.

Reviewed with help from Claude Code.

[Narek-Khachikyan]

@1AbdullahButt Addressed in fd4f520:

1. Added a comment immediately before the dom loop explaining why gt is always set: EQ cases return earlier, and ANY and upper-bound-only ranges receive a lower bound.
2. Added the <X to <X-0 normalization to the algorithm comment, under the upper-bound-only case with prereleases excluded.
3. The PR description already says, "Track empty subset branches independently so they do not change a union's result." That covers the sawNonNull change, so I left the description unchanged.

This commit only adds comments. Locally, npm test passed all 54 files with 100% coverage on Node 26.3.1, including ESLint and template-oss-check. All three PR commits passed commitlint. GitHub currently reports action_required for the new head's workflows; these local results do not establish a CI pass.

[1AbdullahButt]

I read this against ranges/subset.js on main. Three notes:

1. gt is now assumed to be set inside the dom loop. The if (gt) guard around the >/>= handling is gone, so higherGT(gt, c) and gt.operator run unconditionally. As far as I can tell that's safe: eqSet.size === 1 returns before the loop, lt-only now synthesises gt, and a sub with no comparators is the * special case handled earlier. But that's an implicit invariant spread over ~60 lines. A one-line comment at the loop (or an assert-style early return if !gt) would stop a future change to the EQ branch from turning into TypeError: Cannot read properties of undefined (reading 'operator').
2. The -0 normalisation isn't in the algorithm comment. You added "If LT is set without GT or EQ, use the lowest admitted version for GT" to the header, but the rewrite of an exclusive stable <X into <X-0 (when !includePrerelease) also changes semantics and is what makes ['<1.0.0', '0.x', true] pass. A matching bullet would keep the header comment accurate, since it is effectively the spec for this function.
3. sawNonNull moved into the outer loop. I agree with this: on main, a non-null earlier branch could make a later, entirely-null branch return false. It's a separate behaviour change from #949, though. The ['<0.0.0 || <0.0.0', '1.0.0', true] style tests cover it, but it'd be worth a sentence in the PR description so it isn't missed in the changelog.

The test table is thorough, especially the includePrerelease mirror cases and the loose <01.0.0 ones.

Static review; I didn't run the suite here.

Reviewed with help from Claude Code.