Remove allocations from ALPM, Conan and Gentoo comparators

[abhinavgautam01]

Part of #31. Closes three of the seven checklist boxes: ALPM, Conan and Gentoo/APK.

Approach

Same scanner plus stack-buffer treatment #30 applied to SemVer, RubyGems and NuGet, extended to the three comparators that live in ecosystem_extra.go. Each one now walks both versions in lockstep through string views instead of materializing a component list per side.

Comparator	What was allocating	Replacement
ALPM	splitTypedSegments grew a []typedSegment by append, four times per comparison	nextTypedSegment, a lockstep iterator over runs of same-kind bytes
Conan	strings.Split plus a recursive parse whose &pre and &build pointers escaped to the heap	splitConanVersion returns views; the recursion runs over those views, costing stack frames rather than allocations
Gentoo / APK	four strings.Split calls per comparison	splitGentooBase, nextGentooSuffix and nextDotPart

Conan's trailing numeric-zero rule, where 1.2.0 and 1.2 are equal, is handled by counting significant components with conanMainLen instead of building a slice and truncating it.

One shared change in schemes.go: nextSemverIdentifier is plain dot-splitting with nothing SemVer-specific in it, so it is renamed nextDotPart and reused by the Gentoo and Conan paths rather than copied. No behavior change, just the two call sites.

Exported APIs and scheme ordering are unchanged.

Results

Measured with Go 1.26 on darwin/arm64. The before column comes from running the identical new benchmark file against main in a throwaway worktree, so the two runs are directly comparable.

Benchmark	Before	After
ALPM	362.2 ns, 768 B, 10 allocs	88.5 ns, 0 B, 0 allocs
ALPM epoch	84.0 ns, 48 B, 2 allocs	31.2 ns, 0 B, 0 allocs
Conan	665.3 ns, 520 B, 11 allocs	163.2 ns, 0 B, 0 allocs
Conan build	703.2 ns, 672 B, 19 allocs	212.4 ns, 0 B, 0 allocs
Gentoo	173.7 ns, 144 B, 4 allocs	70.9 ns, 0 B, 0 allocs
Gentoo revision	198.3 ns, 160 B, 4 allocs	86.7 ns, 0 B, 0 allocs
APK	172.7 ns, 144 B, 4 allocs	76.1 ns, 0 B, 0 allocs

Between 2.3x and 4.1x faster, with every allocation removed. BenchmarkCompareEcosystem is new in bench_test.go so these paths stay covered, since BenchmarkCompareWithScheme did not reach any of them.

Parity coverage

The issue asks for differential coverage against current behavior, so ecosystem_extra_reference_test.go holds the pre-optimization implementations under ref* names. It was produced by a rename script rather than retyped, so it is a faithful copy of what was on main.

ecosystem_extra_parity_test.go then asserts the rewrite returns an identical result for every ordered pair across two corpora:

• a 100-entry adversarial set covering separator-only strings, trailing separators, empty components, leading and trailing zeros, trailing letters and stacked suffixes
• 688 unique version strings harvested from the univers conformance fixtures for alpm, conan, gentoo and apk, which is 473,344 ordered pairs per comparator

A testing.AllocsPerRun test pins the zero-allocation property so a future change cannot quietly reintroduce a heap escape.

Mutation testing

I checked that the parity suite actually bites rather than assuming it. Each of these deliberate breakages produced a failure naming the exact offending input:

Mutation	Caught by
Drop Conan trailing-zero trimming	compareConan("", "0")
Extract Conan prerelease before build	compareConan("1", "1.0+build-alpha")
Treat a trailing _ as having no suffixes	compareGentoo("", "_")
Split Gentoo on the last _ instead of the first	compareGentoo("1.2.3", "1.2.3_alpha_p1")
Swap ALPM segment-kind precedence	compareALPM("1.0", "1a.2b")

One mutation survived, changing the Gentoo no-letter sentinel from -1 to 0. That one is an equivalent mutant rather than a coverage gap: every ASCII letter exceeds both values, so cmpInt returns the same answer either way.

Testing

• gofmt -l . clean, go build ./... and go vet ./... pass
• go test -race -count=1 ./... passes on all three packages, conformance suite included
• golangci-lint run ./... reports 0 issues

[andrew]

The new BenchmarkCompareEcosystem duplicates the harness in BenchmarkCompareWithScheme, so the extended golangci-lint check fails with dupl. Please add the ecosystem cases to the existing benchmark table or extract the shared runner.

[Copilot]

Pull request overview

This PR removes heap allocations from the remaining ecosystem-specific version comparators (ALPM, Conan, Gentoo/APK) by switching from split/slice-based parsing to allocation-free, lockstep scanning over string views, while preserving comparator behavior via differential parity tests.

Changes:

• Reworks ALPM, Conan, and Gentoo/APK comparison to iterate over runs/components directly (no strings.Split, no slice materialization).
• Generalizes the shared dot-splitting helper (nextSemverIdentifier → nextDotPart) and reuses it in the ecosystem comparators.
• Adds reference implementations plus parity + zero-allocation tests, and introduces a benchmark to keep these comparators covered by -benchmem.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
schemes.go	Renames/generalizes the dot-part iterator helper for reuse outside SemVer.
ecosystem_extra.go	Rewrites ALPM/Conan/Gentoo comparison logic to be allocation-free via scanning/string views.
ecosystem_extra_reference_test.go	Adds mechanically-copied pre-optimization reference implementations for differential testing.
ecosystem_extra_parity_test.go	Adds exhaustive parity checks across corpora + AllocsPerRun tests to pin 0 allocations.
bench_test.go	Adds an ecosystem-specific benchmark to keep these comparators visible in perf/alloc regressions.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[abhinavgautam01]

Thanks, fixed. golangci-lint run --enable dupl ./... reported 2 issues on the previous commit; it now reports 0.

I went with folding the cases into the existing table rather than extracting a runner, so the duplication is removed instead of restructured around. BenchmarkCompareEcosystem is gone and its seven cases now sit in BenchmarkCompareWithScheme with a short comment on why they are there. Net 14 lines fewer. I also dropped the underscores from the sub-benchmark names so they follow the existing Semver and NuGet convention instead of adding a second one.

Since you mentioned the extended check, I also diffed the full linter-set profile of this branch against main to be sure nothing else of mine was hiding in there. No linter category appears on the branch that is absent on main. The only count that moves is nonamedreturns, 6 to 13, which is existing idiom here: main already has it in splitALPMVersion and splitGentooRevision and nextSemverIdentifier used (part, rest string, more bool). I kept that style, so I have left those alone. Happy to change them if you would rather they went.