XL-2 hotfix: call arguments lower as values; nested-call/record args no longer refuse the whole module (list args still refuse, located)

[gunbai-bot]

What broke

gunbc#12145 stopped body_lower_operand_ref_optional from narrowing a sequence to its left element. That narrowing had read h(q: a) as h and [a] as [, dropping everything after the first atom. body_lower_call_arg_value read every argument with that operand reader alone. After #12145, an argument whose value is a nested call, a record literal, a caret symbol, a parenthesised group or a list literal answers Absent. The result is body_lowering_reason_call_argument_unread, and the whole module refuses normalize.

Before number: cool-crab-649's pinned sample (v2.compiler.reference_conservation_census reference_conservation_stratified_sample_paths, 315 files) on main 1b2f64fd49:

• 153/315 modules refuse normalize, against 8/315 at fec339d561, before XL-2: sequence-operand refs reach resolve — operator expressions lower whole, a sequence operand is never narrowed #12145.
• 21,389 dropped rows are module-wide refusal collateral.
• 1,696 dropped rows are silent drops.

The repair (the earliest wrong link, DESIGN §6b)

The wrong link was the argument reader, not the operand reader. A call argument is a value, and the fold already has a value reader: body_lower_value_lowered, used by field initializers and if-conditions (the bottom-up fold, then the walker).

• body_lower_call_arg_value now lowers the argument's value capture through body_lower_value_lowered. The operand reader is the fallback, and an argument neither reads is still a located call_argument_unread.
  • This departs, with a reason, from the suggestion to route list and call sequences inside the operand reader. The value reader already covers nested calls, records, ^sym and groups. Enumerating those shapes a second time in the operand reader would be a second authority for "what a value is".
• Parenthesised group. body_lower_primary_expr had no arm for the grammar's ( expr ), so body_lower_pattern_first_atom_optional answered it with (. The inner name was silently dropped on every route. A group now lowers to its inner expression, or refuses at the group (body_lowering_reason_paren_group_unread).
• List literals still refuse (ruling C). Body lowering has no lowered form for a list literal on any route: [a] becomes [. Accepting list arguments now would trade a loud refusal for a silent drop. So a list literal anywhere inside an argument value refuses at the list, with body_lowering_reason_list_literal_unlowered. "Anywhere" includes a list inside a record argument this reader now accepts. List-literal lowering on every route is the next PR, in this lane.

Evidence

The witness is v2.test.claim.namespace_xl0.call_argument_value_resolve_refusal: 12 claims over one warm front end, enrolled in v2.workflow.floor_pure_producer_share. It runs on the native-lane route: native_test_context_from_ingest and native_census_module_resolution.

claim	main 1b2f64fd49	this head
undeclared name in a nested-call argument → unbound at its atom	red (module refused)	green
… in a positional nested-call argument	red	green
… in a record-literal argument	red	green
… in a parenthesised argument	red	green
named list / positional list / list in nested call / list in record → refuses at the list	red: refused as call_argument_unread, not at the list	green
control: literal and string arguments resolve	red	green
control: nullary-call argument resolves	red	green
control: declared names in nested-call and record arguments resolve	red	green
neighbour: undeclared plain-name argument → unbound at its atom	green	green

These results come from local gunbc run over a ProcessExit entry that ANDs the 12 claims, run once against each fold.

After number: pending. The same 315-file sample on this head is running now. I'll post the refusing-module count and the silent-drop count as a comment, including how many of the remaining refusals are list-literal arguments.

Scope

• Not touched: compiler_frontend_program_status.dag, xl2_completeness, gunbc.namespace_xl2_rehearsal_census, data_initializer_fn_value, the lean_bash roster gate.
• No stage0 file changes: XL-2: sequence-operand refs reach resolve — operator expressions lower whole, a sequence operand is never narrowed #12145 changed this same fold and produced none.

🤖 Generated with Claude Code