Vendor packse scenarios and cross-check against a SAT oracle

[jonyoder]

Vendors astral-sh/packse's scenario fixtures and runs them against resolver.Resolve, cross-checked by an independent SAT oracle. Test-only; no production code changed.

---

What this does

• Vendors 147 non-universal-example scenario TOML files from astral-sh/packse at commit 18be7766b14aa3e03db37bc575959b3f4c6fbacb (2026-08-18, after tag 0.3.59) into resolver/testdata/packse/, plus its LICENSE-APACHE/LICENSE-MIT. See resolver/testdata/packse/README.md for the exact copy command and how to refresh.
• resolver/packse_scenario_test.go: parses the scenarios with BurntSushi/toml (test-only import), failing on any undecoded key so a re-pull that adds a field can't silently drop it.
• resolver/packse_test.go: runs each non-universal scenario against the real resolver.Resolve, classified into three named lists (see below).
• resolver/satoracle_test.go: builds a CNF encoding of the same scenario data (never from provider/candidate output) and solves it with github.com/crillab/gophersat (test-only import), cross-checking the resolver's own answer in both directions.

Summary line

packse: 147 scenarios = 94 pass + 8 known-fail + 4 unsupported + 41 out-of-scope
SAT oracle: cross-checked 102 scenarios, 7 disagreement(s)

Classification

• outOfScope (41): resolver_options.universal = true (all 32 of fork/, 5 tag_and_markers/, 2 backtracking/, 2 wheels/). go-pyresolver resolves one concrete marker environment at a time; universal resolution is deferred by RFD 0001.
• unsupportedOption (4): wheels/no-binary, no-build, no-wheels-no-build, only-wheels-no-binary — resolver.Options has no no_build/no_binary equivalent. The harness refuses to run these rather than silently ignoring the option.
• knownFail (8): run against the real resolver, asserted to still disagree with packse's expected result. Never t.Skip — if one starts passing, the assertion flips to "unexpectedly passed, remove it from knownFail".

Scenario	Reason
extras/missing-extra, extras/extra-does-not-exist-backtrack	go-pyresolver models name[extra] as a virtual package requiring a candidate that declares the extra; a candidate that omits it is excluded rather than the extra being silently dropped, which is what uv does
prereleases/package-only-prereleases, -boundary, transitive-package-only-prereleases	candidate.PrereleaseSet admits a pre-release only via a specifier or an explicit opt-in; it does not implement pip/uv's further fallback of admitting one when a package publishes no final release at all
requires_python/python-less-than-current	go-pyresolver enforces a Requires-Python upper bound per PEP 440; uv deliberately ignores one
yanked/transitive-package-only-yanked-in-range-opt-in, transitive-yanked-and-unyanked-dependency-opt-in	FilteredIndex.ExcludeYanked drops a yanked version outright; PEP 592 (and uv) still allow it when a requirement pins it exactly

None of these are fixed in this PR (test-only, per the issue's scope). Each is a real, reproducible gap for a later PR to pick up.

Three-way disagreement table (packse expected / resolver / oracle)

The oracle implements the PEP 592 exact-pin exception directly (the resolver's known limitation above), so it agrees with packse and disagrees with the resolver on the two yanked scenarios. On the other 6, the oracle reuses the same admission input as the resolver (candidate.PrereleaseSet, and the same strict extras model), so it agrees with the resolver and disagrees with packse — i.e. the oracle's disagreement count (7) is a subset of the resolver's knownFail (8), minus 2 yanked scenarios where it goes the other way:

scenario	packse	resolver	oracle
extras/missing-extra	sat	unsat	unsat
prereleases/package-only-prereleases	sat	unsat	unsat
prereleases/package-only-prereleases-boundary	sat	unsat	unsat
prereleases/transitive-package-only-prereleases	sat	unsat	unsat
requires_python/python-less-than-current	sat	unsat	unsat
yanked/transitive-package-only-yanked-in-range-opt-in	sat	unsat	sat
yanked/transitive-yanked-and-unyanked-dependency-opt-in	sat	unsat	sat

Shared pre-release admission input

Per the brief's one deliberate exception: the oracle takes its admissible pre-release version set from candidate.PrereleaseSet (not from provider), plus Requires-Python and wheel/sdist availability it computes itself. Everything else in the CNF is built from the scenario's own TOML data.

Mutation proofs (run locally, not committed)

1. Flipped candidate.Newest.Less to oldest-first: 13 package-list scenarios went RED.
2. Removed requires_python/python-less-than-current from knownFail: RED (expected-fail scenario now reported as a real failure). Added local/local-simple to knownFail: RED ("unexpectedly passed, remove it from knownFail").
3. Dropped the oracle's at-most-one clauses: disagreement count went from 7 to 16 (new false SATs), RED. Corrupted one resolver pin before the direct-model check: 18+ scenarios went RED ("does not satisfy the oracle's CNF").
4. Copied a scenario into the vendored corpus with an injected unknown key: the loader's strict-decode check fired (undecoded keys [...]), RED.

Consumer impact (PPM, not this PR)

PPM pins go-pyresolver and vendors its dependency tree. Verified by experiment: a test-only import (sat in this case) does not enter a consumer's go.mod or vendor/ tree under go mod tidy/go mod vendor, but does appear in go list -m all under graph pruning. Expect at most a github.com/crillab/gophersat v1.4.0/go.mod h1: line in PPM's go.sum the next time it bumps this module — nothing changes in the vendored code, and module-graph license/SBOM scanners may list gophersat (MIT, harmless).

Not in this PR

• The weekly re-pull job (RFD §14 "Weekly CI") — needs network access; follow-up.
• Random-universe SAT cross-checks beyond the packse corpus — the issue scopes the oracle to packse; follow-up.
• Fixes to anything a scenario exposes (see the knownFail table above) — each gets its own later PR with the scenario as the red test.

NEWS / CHANGELOG

No PPM NEWS.md entry (sibling repo, test-only, not customer-visible). No go-pyresolver CHANGELOG.md entry (test-only, per precedent ee1a650).

Refs rstudio/package-manager#18658

🤖 Generated with Claude Code