| name | member-resolution-soundness |
| description | The recurring soundness traps in `borzoi-sema` code that walks base chains or compares member signatures through `AssemblyEnv` โ type identity ignoring the referencing assembly, name hiding across static/instance and kind, and structural `TypeRef` comparison. Use before writing or changing any member-access, inheritance, completion-through-inheritance, or overload-resolution logic in `crates/sema`, and when a reviewer flags one such soundness hole (audit all three at once instead of patching the one). |
AssemblyEnv member resolution: the three traps
Any sema code that walks base chains or compares member signatures through
AssemblyEnv hits the same family of soundness holes. On the inheritance slice
a reviewer found one facet per round for four rounds; they are all instances of
the three below. Audit all three in one pass rather than fixing the flagged
one.
1. AssemblyEnv resolves types by (namespace, name, arity) โ it ignores TypeRef::Named.assembly
Only the first-enumerated definition per key is kept. So when two referenced
assemblies define the same full type name โ or when a base's assembly is absent
while a same-named type from another assembly is present โ the by_type slot
can hold the wrong type, and a base-chain walk silently walks into it.
The shipped fix in resolve_base: resolve a base only when the candidate's
assembly name matches the base's referenced assembly (or, when
assembly: None, the declaring type's own assembly).
Name โ not the full (name, version, public_key_token) identity. A base is
compiled against a possibly-different version than the one loaded, and the
compiler binds by name with version redirection; full-identity matching
over-defers the ordinary case.
2. Name hiding is by nearest declaring level, across kind AND static/instance
A derived public static (or a non-method) named M hides an inherited
public instance M, but is unreachable through a value receiver โ FCS
leaves d.M / d.M() as obj. So the "who declares this name?" check must:
- include statics โ a same-name static at the owning level blocks, so defer;
- exclude non-public members โ cross-assembly, an inaccessible derived member
is removed before hiding applies, so a private derived
M does not shadow a
public inherited M;
- ignore mere coexistence โ a static that simply shares the owning level with
an instance method as part of one overload set (e.g.
Object.Equals(object)
alongside static Equals(object, object)) does not hide. Don't defer on it.
(The static-hiding claim is FCS-probed: HDerived().M() โ call:function / obj.)
3. Comparing two TypeRef signatures to decide "same overload" is a rabbit hole โ don't
Structural TypeRef equality is unsound here for at least three independent
reasons: cross-assembly meaning (assembly: None is relative to the declaring
assembly), byref/out param flags being stripped off ty onto
Parameter::is_byref / is_out, and nested-type ambiguity.
The inheritance slice originally deduped overrides this way and a reviewer found
a hole every round. The resolution was to delete the dedup entirely: collect
the group without deduping and defer any group of โฅ 2, so an overridden method
spanning levels simply defers. No differential coverage needed the dedup (all
typed cases are single-level).
Lesson: this was speculative complexity. When signature comparison keeps
generating soundness holes, delete it rather than keep hardening it.
Working notes
- Probe FCS before trusting a review claim about FCS behaviour โ the traps above
are stated as probed facts, and claims that sound equally plausible have
turned out false.
- The tests read the reference pack, not the SDK runtime BCL; see
assembly-surface-check before choosing a BCL member to assert on.
- Deferral is the safe direction. Prefer "defer this group" over a clever rule
that is right in the cases you thought of.
