v2: every type declaration lowers — plain aliases, bodyless and coproduct declarations leave the parse shell

[gunbai-bot]

Replacement migration for gunbc.recurring_failure_mode plain_type_alias_left_as_a_parse_shell_in_v2. This PR is lowering-only; the side chat ruled it option B, and declaration grounding is a separate infer change.

The defect

v2.compiler.body_lowering_fold body_lower_type_decl_by_body accepted any non-generic type declaration that was neither a record nor where-refined as its raw parse shell, through NoTypeParams => outcome_accepted(value: shell). body_lower_type_decl did the same for a declaration it could not read.

So type B = T | F, type I = Int and type W reached infer as grammar tokens. Measured through the closure route on main:

Declaration	Nodes	Not grounded after infer
type B = T | F	31	29
type I = Int	28	22

About 930 non-generic = declarations take that path.

The change

• One lowering for every declaration.

  • body_lower_eq_type_decl, formerly the generic-only arm, now serves every = and bodyless declaration, generic or not.
  • A plain alias becomes type_alias_wrapper, and a plain bodyless declaration becomes type_opaque_wrapper; each carries an empty binder Conj (body_lower_decl_binders).
  • A plain coproduct becomes its Disj member.
  • The shell-accepting arms are removed, and an unreadable declaration now refuses with body_lowering_reason_type_decl_unreadable.

• Shared contract. v2.std.node type_binder_conj_conforms now admits zero binders; the identity and uniqueness checks are unchanged.

• The view. Its arms are renamed TypeAlias and OpaqueTypeDecl, which are cardinality-neutral, so no reader equates "wrapped" with "generic".

  • symbol_index_fill is unchanged in behaviour: an alias or opaque declaration still binds without declaring members.
  • semantic_decl_emission refuses an alias or opaque declaration at the struct and enum surfaces under its own reason, where it previously used a "generic declaration" reason.

• Readers of the old representation are deleted. v2.compiler.namespace_graft's type_decl naming arm read a unit's shell. With that arm answering UnitResidual, all of these still held:

  • declaration_graft_assemble 16/16
  • type_param_binder_frame 37/37
  • the new claims 5/5

  It is deleted, along with namespace_graft_type_decl_name, _target and _named_edge.

Evidence

All numbers come from local claim_batch runs on this branch.

v2.test.claim.body_lowering.plain_type_decl_lowering asserts the lowered forms positively on real assemblies. It has five nullary verdict producers, each WARM in floor_pure_producer_share.

Row	Result on this branch	Result with the pass-through restored
type Decision = Allowed | Blocked is a Disj with exactly [Allowed, Blocked], and the index descends into it	PASS	FAIL
type Id = Int is the alias wrapper with 0 binders and an atom right-hand side, and the module assembles with keep(x: Id) -> Id	PASS	FAIL
type Ref = Id is an alias with 0 binders that declares no members	PASS	FAIL
type PointerWidth is opaque with 0 binders, not an empty record, and declares no members	PASS	FAIL
type Wrapped = Only { a: Int } stays a member	PASS	PASS (preservation control; it always took the record arm)

type_param_binder_frame's 37 existing claims pass on the change.

Not claimed

• Grounding. After this change, infer leaves 7–8 structural nodes ungrounded per declaration: the module spine, the Disj, childless variants and the empty binder list. This is because the product-introduction rule has no Disj rule and keeps a childless Conj on the frontier. That frontier is left intact here and is the next, separate infer capability.
• Emission. The walls are unchanged: the module wall renders only bodied functions, and a bodied-function signature still spells Bool as bool_node_symbol.

🤖 Generated with Claude Code

[briansrls]

APPROVE-MERGE at exact head c314f4c3c85bbc8dbfcf0b13b272ea5dff0b5711.

This satisfies the lowering-only / option-B boundary we agreed.

The replacement is at the producer: every readable non-record/non-where type declaration now reaches the shared body_lower_eq_type_decl path; plain aliases and bodyless declarations use the existing alias/opaque wrappers with a zero-cardinality binder collection; coproducts remain members; unreadable/uncarried declarations refuse located rather than returning the parse shell. I found no compatibility arm preserving the old accepted shell.

The consumer side is migrated rather than dual-read: TypeAlias / OpaqueTypeDecl are cardinality-neutral in type_decl_view; symbol_index_fill preserves the no-members semantics for alias/opaque forms; semantic declaration emission refuses those forms explicitly; and the obsolete namespace-graft type-declaration shell reader/helpers are deleted.

The five new real-assembly claims positively discriminate coproduct/member vs alias vs opaque shapes and member-index behavior. The required-floor log shows all five planned-and-passed, their five WARM shared producers populated, and enrollment-margin admission. Existing type-parameter claims remain covered by the floor. The reported local mutation restoring the NoTypeParams shell pass-through makes the four intended discriminating rows false; I did not independently rerun that local mutation.

The type_binder_conj_conforms relaxation from nonempty to zero-or-more binders is acceptable here: identity/uniqueness checks remain, and the binder collection's cardinality is now representation data rather than an implicit 'generic' discriminator. I found no new source producer that uses an empty binder list to fabricate a generic member or arrow.

The recurring-failure-mode row closes only the representation debt and explicitly leaves structural grounding/emission unclaimed. No infer rule, grounding suppression, or emission bypass is introduced. That is the correct boundary; the remaining Disj/childless-Conj grounding frontier should be a separate inference capability.

Exact-head CI is green across witnesses, generated, compiler, clippy, floor, and emit-build. GitHub reports the PR mergeable. No remaining merge-blocking finding.