iter bugfix-instance-body-unbound-var: GREEN — walk instance method bodies through check_fn

Fixes the bug RED-pinned in 72f3f65: an unbound identifier inside an
instance method body slipped past `ail check` (false-OK exit 0) and
surfaced only at `ail build` as the degraded internal-error diagnostic
`monomorphise_workspace: unknown identifier: <name>` with no source
location, symbol kind, or "did you mean" candidates.

Root cause: `check_def` in `crates/ailang-check/src/lib.rs:1574-1595`
early-returned `Ok(())` for `Def::Class(_) | Def::Instance(_)` without
ever invoking the body walker. The arm's comment claimed iter 22b.2
landed instance-body typechecking; that wiring was never implemented.
Only the workspace-load coherence checks (Orphan/Duplicate/
MissingMethod) touched instance defs, and they inspect schema only,
not the method-body identifier graph.

Fix: split the combined arm so `Def::Class(_)` stays schema-only at
this layer, while `Def::Instance(inst)` routes through a new helper
`check_instance` that lifts each `InstanceMethod`'s `Term::Lam` body
into a synthetic `FnDef`, applies class-method substitution (class
param -> instance type, via the existing `substitute_rigids` /
`substitute_rigids_in_term` helpers), and hands it to `check_fn`.
The reuse of `check_fn` is what gets the body walked through the
same identifier-resolution path as fn bodies, so an unbound name
fires `[unbound-var]` at `ail check` with exit 1 and the standard
structured diagnostic.

Per "minimal fix, no surrounding cleanup" constraint: no widening
to Def::Class (still schema-only); no changes to the
monomorphise_workspace diagnostic wording (defence-in-depth
internal-error path stays as backstop); no changes to mono.rs,
codegen, or any other crate.

Tests: 557 baseline + 2 new RED -> 559 green. Both RED pins in
unbound_in_instance_method_pin.rs now PASS. fn-body level
unbound-var path stays green. All 8 carve-out fixtures classify
unchanged.

Known debt (out of scope per carrier; queued in journal):
`check_instance` does not yet cross-check the substituted method
signature against the class's `ClassMethodEntry.method_ty` — a
malformed instance declaring wrong types in its Lam would still
typecheck cleanly. The body-walk catches unbound vars and
effect-mismatches against the Lam's own declared types, but not
class-method-signature mismatch.
This commit is contained in:
2026-05-13 12:09:39 +02:00
parent 72f3f6541b
commit 77f584abbb
3 changed files with 204 additions and 11 deletions
+98 -11
View File
@@ -1580,20 +1580,107 @@ fn check_def(
Def::Fn(f) => check_fn(f, env, out_warnings),
Def::Const(c) => check_const(c, env, out_warnings),
Def::Type(td) => check_type_def(td, env),
// Iter 22b.1: schema-only landing for class/instance defs.
// Class-schema validation (InvalidSuperclassParam,
// ConstraintReferencesUnboundTypeVar) and instance-body
// typechecking (with class-method substitution) land in 22b.2.
// ctt.3: KindMismatch retired — the malformed-class-param
// shape now fires BareCrossModuleTypeRef from the canonical-
// form validator instead.
// Workspace-load coherence checks (Orphan / Duplicate /
// MissingMethod) already fire from `workspace::build_registry`,
// so a malformed instance never reaches this point.
Def::Class(_) | Def::Instance(_) => Ok(()),
// bugfix-instance-body-unbound-var (2026-05-13): each instance
// method body is walked through `check_fn` via a synthetic
// `FnDef` lifted from the `Term::Lam` the InstanceMethod body
// carries. This routes instance bodies through the same
// identifier-resolution path as fn bodies, so an unbound
// identifier inside an instance method fires `[unbound-var]`
// at `ail check` instead of slipping past and surfacing as a
// degraded "monomorphise_workspace: unknown identifier"
// diagnostic at `ail build`. Class-method substitution
// (class param → instance type) happens inside
// `check_instance` before the body-walk. Class-schema
// validation (InvalidSuperclassParam, etc.) and the workspace-
// load coherence checks (Orphan / Duplicate / MissingMethod)
// fire from `workspace::build_registry` independently, so a
// malformed instance never reaches this body-walk. `Def::Class`
// remains schema-only at this layer.
Def::Class(_) => Ok(()),
Def::Instance(inst) => check_instance(inst, env, out_warnings),
}
}
/// bugfix-instance-body-unbound-var (2026-05-13): route each
/// `InstanceMethod` body through `check_fn` by wrapping it in a
/// synthetic `FnDef`. The InstanceMethod body is conventionally a
/// `Term::Lam` (the post-mq.3 shape; see iter mq.3 journal
/// "instance-method body shape ... existing-convention is `Term::Lam`
/// with paramTypes/retType"); we extract its `param_tys` / `ret_ty`
/// / `params` / `body` into an `FnDef` shape and reuse the existing
/// fn-body-check machinery.
///
/// Class-method substitution: the on-disk Lam's `param_tys` / `ret_ty`
/// reference the class's type parameter directly (e.g. the prelude
/// `instance Eq Int` body declares `(typed x a) (typed y a)` with `a`
/// being `Eq`'s class param). Before handing to `check_fn` we look up
/// the class's `param` name via `env.class_methods` and substitute it
/// to the instance's concrete `type_`, using the existing
/// `substitute_rigids` helper on type positions and
/// `substitute_rigids_in_term` on the body (which targets any embedded
/// Type annotations on nested Lams / LetRecs). Without this step
/// `check_fn` would fire `polymorphic-not-supported` on the class
/// param appearing in non-Forall position.
///
/// The synthetic fn name is `"<class>::<method>"`, used only for
/// diagnostic strings.
///
/// Non-Lam method bodies are not part of the current schema (all
/// fixtures in the corpus and the convention reaffirmed in iter mq.3
/// use `Term::Lam`); they are silently skipped here. If a future
/// schema admits non-Lam shapes, this arm must be extended to walk
/// them — silent skip is the conservative choice today because
/// the pre-fix behaviour was also silent.
fn check_instance(
inst: &InstanceDef,
env: &Env,
out_warnings: &mut Vec<Diagnostic>,
) -> Result<()> {
let qualified_class = qualify_class_ref_in_check(&inst.class, &env.current_module);
for im in &inst.methods {
// Look up the class's param name. Workspace-load's MissingMethod
// gates this — if the (class, method) pair isn't in the table,
// the instance was structurally rejected upstream and we never
// get here in practice. Defensive fallback: skip substitution
// (body walks without class-param resolution).
let subst_map: BTreeMap<String, Type> = env
.class_methods
.get(&(qualified_class.clone(), im.name.clone()))
.map(|entry| {
let mut m = BTreeMap::new();
m.insert(entry.class_param.clone(), inst.type_.clone());
m
})
.unwrap_or_default();
let Term::Lam { params, param_tys, ret_ty, effects, body } = &im.body else {
// Non-Lam method body: skip (see fn-level doc-comment).
continue;
};
let subst_param_tys: Vec<Type> = param_tys
.iter()
.map(|t| substitute_rigids(t, &subst_map))
.collect();
let subst_ret_ty = substitute_rigids(ret_ty, &subst_map);
let subst_body = substitute_rigids_in_term(body, &subst_map);
let fn_ty = Type::fn_implicit(
subst_param_tys,
subst_ret_ty,
effects.clone(),
);
let synthetic = FnDef {
name: format!("{}::{}", inst.class, im.name),
ty: fn_ty,
params: params.clone(),
body: subst_body,
doc: None,
suppress: Vec::new(),
};
check_fn(&synthetic, env, out_warnings)?;
}
Ok(())
}
fn check_type_def(td: &TypeDef, env: &Env) -> Result<()> {
// Iter 13a: a parameterised ADT (`vars` non-empty) installs its
// type parameters as rigid vars while checking the ctor field