Resolve relative imports by their level (#369); accept complete enum-literal unions (upstream a490662)

[Wamwea]

TLDR

Fixes relative-import resolution (from ..sub import mod no longer false-positives imports_unresolved, Refs #369) and teaches assignment_compatibility the enums-expansion equivalence that python/typing@a490662 now requires, keeping conformance at 141/141 (100%) with 0 false positives.

What Was Added?

• ImportInfo.relative_level: u32 (crates/basilisk-resolver/src/scope/import_types.rs) — the leading-dot count of a from-import, populated from StmtImportFrom.level in the visitor (class_info_ext.rs) for both named and star forms; always 0 for plain imports.
• resolve_relative_import_cached (crates/basilisk-checker/src/imports/resolve.rs) — FsCache-aware variant of the existing (previously dead) relative resolver, so the per-file import loop shares directory listings.
• Enum-member knowledge in assignment_compatibility (rules/assignment_compatibility/mod.rs): SkipNames.enum_members, collect_enum_members (members = valued class-body attributes, excluding nonmember(...) and _-prefixed names), and enum_complete_union_assignable implementing the typing-spec enums rule "a complete union of all literal members is equivalent to the enum type".
• New integration tests:
  • crates/basilisk-checker/tests/relative_import_resolution_tests.rs — 8 tests driving parse → visitor → resolve_module_imports over the issue Relative "from ..mylib.some import smth" was broken several versions ago #369 package layout.
  • crates/basilisk-checker/tests/checker/assignment_compatibility_2_tests.rs — 5 tests for the enum complete-union equivalence.

What Was Changed or Deleted?

• resolve_module_imports (crates/basilisk-checker/src/imports/apply.rs) now dispatches imports with relative_level > 0 to the relative resolver instead of the absolute search path. Previously the level was discarded, so from ..sub import mod was resolved as absolute sub and failed; single-dot imports only worked by accident via the importer-directory fallback.
• Relative imports are exempted from uv-registry classification, package-metadata enrichment, and typeshed stub-distribution lookup — a relative import can never be a PyPI package, so the misleading "sub is not installed" help text is gone.
• check_vars in assignment_compatibility accepts an enum-typed value assigned to a Literal[...] union naming every member of that enum (incomplete unions and other enums' members still mismatch). This closes the false positive that python/typing@a490662 ("Conformance: broaden enum-expansion tests", merged 2026-08-04) introduced against main — upstream's own committed results graded basilisk (and mypy, and pyright) as failing the new enums_expansion.py test4.
• Generated/ratcheted artifacts from the gates: conformance stamps re-graded at a490662 (READMEs, docs/specs/CHECKER-ARCHITECTURE-SPEC.md, website/src/_data/conformance_report.json), and the mutation baseline ratchets up (138/145 → 139/146 caught, kill rate 100%).

How Do The Automated Tests Prove It Works?

• relative_import_resolution_tests.rs: double_dot_bare_module_target_resolves_to_parent_sibling_package, double_dot_dotted_module_target_resolves_through_parent_package, bare_dot_import_resolves_to_own_package_init, bare_double_dot_import_resolves_to_parent_package_init, and double_dot_star_import_resolves_to_parent_sibling_package all failed with resolution == Unresolved before the fix and now assert the exact resolved paths (src/pkg/sub/__init__.py, src/pkg/sub/mod.py, …). Guards single_dot_sibling_module_resolves, missing_relative_target_stays_unresolved, and absolute_import_still_walks_search_paths pin that the dispatch neither breaks the accidental single-dot path nor blanket-accepts unresolvable or absolute imports.
• assignment_compatibility_2_tests.rs: enum_assigned_to_complete_literal_union_no_diagnostic, single_member_enum_assigned_to_its_literal_no_diagnostic, and nonmember_attribute_does_not_count_toward_completeness failed before the fix; enum_assigned_to_incomplete_literal_union_fires and literal_union_of_wrong_enum_still_fires prove incomplete/wrong unions still error.
• The real python/typing harness (fresh clone of a490662, unmodified src/main.py --only-run basilisk, clean release build): 141/141 graded files pass, 0 false positives — including the broadened enums_expansion.py. Without the enum fix the gate fails at 99% with 1 FP.
• Full local CI-equivalent run: workspace tests + coverage (all 8 crate thresholds met), make lint, Zed (97 tests + standalone mirror build), VS Code e2e (633 tests + corpus suites), Neovim suite (44.13% ≥ 44%), Shipwright binary gate, website build + 52 Playwright tests, and mutation testing (146 mutants, 139 caught, 0 missed, 100% kill rate — the one new mutant introduced by the check_vars change is killed by the new tests).

Spec / Doc Changes

README/README.zh/README-pypi, docs/readme/* sources, docs/specs/CHECKER-ARCHITECTURE-SPEC.md, and website/src/_data/conformance_report.json carry only the regenerated conformance stamps (graded commit 0dc9b5d → a490662), produced by the scorer's own stamping pass.

Breaking Changes

• None

🤖 Generated with Claude Code