A lambda parameter whose declared type is a type variable now binds it: two independent defects, one of them a fabricated empty type

[gunbai-bot]

Root 2 of the receiver-type-never-propagated lane (node://adhoc-b99a2a7b-866).

A lambda argument contributes nothing to a call's generic substitution — the argument fold skips it — yet reads its own parameter types out of that substitution. Two defects sat under that asymmetry. Neither subsumes the other, and each was measured necessary by fixing the other alone first.

1. Order

call_subst accumulates only from arguments already inferred, so a lambda saw only what the arguments to its left had bound. The fold is extracted to one function (infer_call_arguments_generic_pass) and run a second round seeded with the substitution the first ended with — gated on generic_names nonempty, a nonempty substitution, and an argument containing a lambda.

Sound rather than merely convenient: lambda arguments are skipped for unification, so deferring what they would have contributed cannot change the result, and a slot already bound substitutes to a concrete node that unify_generics no longer binds. Bounded at two rounds, and the annotation states the narrowing — a binding derivable only from a substituted expected type is still out of reach.

2. A fabricated empty type — the one the first fix did not reach

record_lit_instantiated_fields built its slot substitution with resolved_type, which answers the empty error_type for any node carrying no inferred type — which is every authored type argument. So instantiating a generic record's fields substituted its type parameter to nothing.

Harmless on a scalar field, fatal on a callable one: the parameter node was replaced wholesale, param_node_type_expr found nothing, type_node_is_established answered false, and the lambda bound to that field fell to type_variable_node("callable_param"). Measured with the type argument already concrete — instantiating Sink<List<Int>> produced a parameter node with empty name, no children and no params — so inference is not part of the question. child_type_node is the total read of the same fact.

This is why the fabricated value was worse than a refusal (§5): error_type is a plausible-looking Node, so it substituted silently and surfaced stages later as an unresolvable method on a receiver nobody could point back to a declaration.

Measured

Same probe both ways, v2 pipeline closure (--entry src/v2/compiler/00_compile.dag, 175 files emitted):

	blocking	advisory
baseline (HEAD)	0	1983
this branch	0	1892

The 91 partition exactly: 7 ReceiverTypeUnestablished + 13 undecidable-at-call in tokenize + 71 undecidable-at-list-element. No advisory class rose. The other four unresolved_method_frontier rows' modules were compiled as entries; none moved.

Regen fixed point reached — required-regen: first_generation_equal=true.

The tokenize frontier row is deleted, not lowered

Its seven apply sites measure zero observed. The budget is an equality in both directions, so leaving the row would leave headroom a reintroduction could slip into — the same deletion the object_store rows got.

Its cause: field also named the wrong idiom: LexMatchThunk { apply: fn(s) } declares its parameter String, a leaf, and a non-generic record field holding a callable compiles 0 blocking. The real receivers were the generic algebra parameters declared fn(R, R) -> R. Three carrier clauses that repeated the wrong attribution or the stale seven-sites count are corrected with it. The misattribution was found by quick-bat-372; their annotated-parameter measurement is recorded as theirs rather than restated as mine.

Evidence, enrolled rather than retired (§4b(4))

dag/test/claim/generic_lambda_parameter_binding_witness_test.dag carries the three sources that refused — record-field lambda, lambda argument before its binder, and the field-access form that reported no field 'px' on type 'R' — plus a control whose method is genuinely absent from the bound type. The control asserts MethodNotFound fires and MethodExistenceUndecided does not: a pair reachable only when the parameter is bound, so a regression that unbinds it fails both halves rather than going quietly green.

All four measured FAIL on a claim_batch built from HEAD and PASS on this tree.

v1 seed admission

src/v1/04_infer.dag is frozen-with-active-maintenance. The admission test is purpose (operator ruling 2026-08-20): this is a compiler inference repair in direct support of the v2 self-host program — the seven sites are in v2.compiler.tokenize — and it adds no growth surface to the seed.

🤖 Generated with Claude Code