[mparrett]

Why

#1057 gave aget, alength and byte-array guarded direct calls to their Go entries, taken while rt.NativePrimsIntact() holds. Those entries (rt.CoreAgetf and friends) return (vm.Value, error), but typeinfer can give the call's node a native slot: float64 for aget on a ^doubles array, int64 for alength, *vm.TypedArray for a byte-array local. typed-call-result-stmts assigns the vm.Value straight into that slot, so the lowered Go doesn't build:

if rt.NativePrimsIntact() {
	arg__7, callErr = rt.CoreAgetf(arg0, arg__6)   // arg__7 is float64
	...
} else {
	tb23, callErr = ec.Invoke2(rt.CachedVarFnEC(ec, &__v_clojure_core_aget, ...), arg0, arg__6)
	...
	tv23, tk23 = tb23.(vm.Float)
	...
	arg__7 = float64(tv23)
}

The else branch is already right: it reuses trampoline-call-stmts, which unboxes through boxed-call-stmts since #1044. The direct branch never got the same step.

What

The (vm.Value, error) case of typed-call-result-stmts calls boxed-call-stmts instead of emit-call-stmts. Both branches now take the result into tb<nid>, check the error, and comma-ok assert it to the slot's vm type, returning a vm.TypeError on mismatch. They register the same temp names and types, and the prologue's distinct emits each declaration once.

A vm.Value slot or a dead result goes through emit-call-stmts unchanged, so only sites that were Go type errors emit different code.

Only the Go-AOT lowerer changes. pkg/compiler and *ir-compile* are unaffected.

Verification

• New fixture test/gogen/native_prim_typed_slot.lg: aget with an index boxed by a var, alength of an unhinted array into an int slot, and a byte-array local filled in a loop. On main its lowered Go fails to build at all three sites. With this change the gogen trampoline builds it and matches the bytecode VM ([13.0 3.0 [0 2 4 6]]).
• All 24 test/gogen fixtures pass the trampoline.
• make check-generated-manifest and make lint pass; go test ./pkg/ir/ passes. I didn't run the full make test locally.
• The 12-line repro from lg compile: aget with a boxed index emits Go that does not build #1081 builds and prints 13.0.
• phosphene's kernels, which hit this at 37 sites, compile with lg compile with no workarounds, and the native binary renders plasma and milk scenes.

Fixes #1081

🤖 Generated with Claude Code

[nooga]

Approving. This fixes the root cause: the (vm.Value, error) branch now unboxes into a natively typed slot through boxed-call-stmts, the same path the trampoline branch has used since #1044. boxed-call-stmts falls back to emit-call-stmts when the slot is vm.Value or the result is dead, so only sites that used to be Go type errors emit different code. The tb/tv/tk temps are registered the same way in both branches, so the prologue doesn't get duplicate declarations. The fixture covers all three shapes from #1081, and the .lg edit is regenerated.

Nit for the record: this touches only the Go-AOT lowerer, so pkg/compiler and ir-compile are unaffected. A line saying so in the body would satisfy the three-paths note in AGENTS.md.