Emitter: refuse an unresolved variant pattern that would render as a bare (catch-all) identifier

[gunbai-bot]

Follow-up to #12733, as the parent lane asked: turn the emitter's widening fallback into a refusal.

Defect (DESIGN §5, a failure arm that widens). v1.compiler.emit_rust rust_variant_path_opt answered an unresolved parent enum with the bare variant name. In Rust, a bare identifier in pattern position that names nothing in scope is a fresh binding. So Add => … without BinOp::* in scope matches every value, and rustc only warns about it.

Change

• refuse_unresolved_bare_variant_pattern now wraps the three pattern producers: emit_variant_pattern, emit_variant_pattern_rc_aware and variant_pattern_shape_for.
• When the parent is unresolved and the render is a bare identifier (no {, ( or ::), the emitter writes compile_error! at the arm, naming the variant and the scrutinee.
• Optional variants keep their Some/None lowering.

Census: the first cut was too broad, and that measurement is why the rule is keyed on the render. Refusing every unresolved parent broke both emitted crates.

• The broken arms were v2.std.node EdgeLabel Named { name: _ }, matched over unstamped scrutinees in v2.std.compilers.body_lowering and v2.compiler.emit_produced.
• Named is also declared by std.algebra and extdeps.formats.spice, so its parent never resolves.
• Those arms render Named { .. }, a struct pattern that Rust resolves by path and never binds.

With the narrowed rule, measured locally:

• gunbc test //gunbc/instruments:v2-native-cli: exit_status=0 warning_count=0.
• gunbc test //gunbc/instruments:self-host: exit_status=0 warning_count=0.
• No emitted file takes the refusal.

Discriminating red. The generated --lib test unresolved_variant_pattern_refuses_instead_of_binding_bare (v1.compiler.compiler_tests_rust ct_unresolved_variant_pattern_refuses_test):

• Calls all three producers with no parent and requires the refusal.
• Carries three controls: stamped parent renders BinOp::Add, optional Absent renders std::option::Option::None, and a fielded Named still renders Named { .. }.
• Is green locally. With variant_pattern_parent_unresolved forced to false in the mirror it fails on the unresolved assertion, so it is a real red.

No accepted source program can produce this state: the typechecker refuses a variant name that is declared twice in one closure. So the red is authored at the unit boundary.

RFM: gunbc.recurring_failure_mode unresolved_variant_pattern_renders_as_a_catch_all_binding.

• Found at: silent.
• Now: mechanically preventable. It is held by the rust_unit_tests step, which is still off the merge path (rust_unit_tests_off_the_merge_path).
• Ceiling: structurally impossible, once MatchPattern VariantPattern carries a resolved parent identity.
• No live wrong-arm specimen exists: the original 52-warning report did not reproduce (see Enumerate BinOp arms in rust_refusing_int_operator_helper (NFR burn-down) #12733).

Follow-up, not in this PR: unstamped scrutinees with an ambiguous variant name. The census found v2.std.node EdgeLabel Named arms whose parent enum never resolves.

• Named is declared in three modules (v2.std.node, std.algebra, extdeps.formats.spice), so unique_variant_parent answers none.
• The scrutinee in v2.std.compilers.body_lowering and v2.compiler.emit_produced carries no type stamp, so the emitter can't pick the parent from it.
• These arms render correctly today only because the braced Named { .. } resolves through the authored import. The emitter never establishes which Named it means.
• Trigger: match-pattern resolution stamps VariantPattern.parent_enum (or the scrutinee type) for every arm. When that lands, these arms render as EdgeLabel::Named { .. } and the unresolved-parent population for this site goes to zero.

Regen: run locally (ulimit -v 24G, -j4, private target dir) until rebuild_packages=0. Only the installed mirrors are committed.

🤖 Generated with Claude Code

[gunbai-bot]

On the non-blocking point raised in review 73104 and review 73114 (rust_pattern_is_bare_identifier searches the rendered string, and pattern_parent_enum is recomputed): agreed that carrying the render shape as a typed value is cleaner. I'm not changing it here. The string check is exact for the three shapes these producers emit (bare, braced, parenthesised) plus path-qualified. The clean fix is the follow-up this PR's description already names: once VariantPattern.parent_enum is stamped for every arm, the Absent arm, the recomputation and this check all go away together. Doing half of that now would be churn that the follow-up deletes.
— sent from fierce-cat-462