Repository navigation
Resolve instance declarations and instance signatures #90
Description
Part of #75 (Phase 2). Instance declarations are not part of resolution, which costs two M2 failing cases and blocks the library work in Phase 3.
Failing cases
| File | Expected | Why |
|---|---|---|
failing/881.purs |
DuplicateValueDeclaration |
Two foo definitions in one instance; we report nothing. |
failing/InstanceSigsOrphanTypeDeclaration.purs |
OrphanTypeDeclaration |
A type declared in an instance signature; we report nothing. |
Why this is on the critical path
Phase 3's Prelude needs Eq, Ord, Functor, Apply, Applicative, Foldable, Semigroup, Monoid as classes with instances. The class and instance elaboration is done (#63, closed; commits 7b39ca8, c983b32, 814e04b, fe82bb0, and I verified a class + superclass + instance program runs under Wasmtime with the superclass-projected method). What is missing is instances in the resolver: duplicate member detection, instance signatures participating in orphan checks, and instance names in the module namespace.
Scope
- Instance declarations lower into AST and HIR and enter the resolver.
- Instance member names join the value namespace for duplicate detection (
DuplicateValueDeclaration). - Instance signatures are checked for orphan types (
OrphanTypeDeclaration). - Instance names participate in module-level name conflicts.
- Export/import of classes carries instance visibility consistently.
Do not narrow this to the two failing cases: the resolver must model instances the way it models classes, or Phase 3 will build the library on a partial model.
Acceptance
Both cases agree, instances resolve across modules, and stdlib can declare classes with instances.
Activity
biuld commentedon Oct 2, 2026
Closed by #117 (merge commit
0c7544d), two commits:98e46e0attributes every program diagnostic to a namedDiagnosticOrigin,cb52868resolves instance declarations and instance signatures.Evidence, re-measured on the merged tree with the annotations oracle against the vendored v0.15.16 corpus:
failing/881.purs(DuplicateValueDeclaration) andfailing/InstanceSigsOrphanTypeDeclaration.purs(OrphanTypeDeclaration).passingresolution 59/413, unchanged, with 354 first-stage blockers: 337 missing library modules, 8 at P3, 5 at P2, 4 at P0, none on assembly.The scoreboard numbers were not trustworthy until diagnostic attribution was fixed, so the first commit repairs it before the second re-measures. Two defects, both confirmed by tests that fail without the fix: a lenient pass enumerated the modules that survived resolution and used that position as the source index, so a diagnostic landed on whichever module took the failing module's slot and an imported instance was looked up in the importing module's own slot; and the trusted library prefix was subtracted with saturation, mapping every library diagnostic and every module-less backend diagnostic onto the caller's first source.
DiagnosticOriginnow names a caller source, the library, and the program as a whole. That commit is scoreboard-neutral againstmaster: the full board output is byte-identical.Corpus casesstays 2, the twofailingcases this issue recovers, one pererrorCode; nopassingcount moved.Validation:
cargo fmt --all --check,cargo clippy --workspace --all-targets -- -D warnings,cargo test --workspace(1126 passed, 0 failed),PSRS_REQUIRE_WASMTIME=1 cargo test --workspace(1126 passed, 0 failed, 5 ignored, Wasmtime 49.0.1),PURESCRIPT_REPO=… cargo test -p psrs-driver --test upstream(4 passed), full scoreboards run twice with identical output, and both required Wasmtime CI checks.