reference: self_attention.t27 passes or fails by host -- its test iterates an undefined W585 scaffold slice (t27b ledger STALE) #8051

Refs #6063. Related: #7733 (same W585 scaffold mechanism; this is the undeclared-helper case, and its effect on the t27b ledger).

What the ratchet reports

The t27b lab run at master 7a07828 (https://t27b-lab-production.up.railway.app/runs/7a07828fc8d65c9626f76018c0884a8f01e87ab7.json) reports

STALE  specs/ml/recurrent/self_attention.t27: the reference does not pass it any more (fail: 1 of 1 tests fail)

with reference_tests: {"forward_basic_case": false}. The ledger row (blocked, StmtAssign(undeclared), unimplemented) was added by the batch bless #7476, which recorded the reference as passing this spec.

It is not a new regression

• The spec has not changed since a67f7df (2026-09-16).
• The lab reference (t27c test-report <file> --specs-dir specs, zig 0.16.0, Linux x86_64) already failed it in the lab runs at ba161bb (2026-10-07 09:56Z, the master t27b ledger: batch bless from a full lab run on master (Closes #7432) #7476 says it blessed from) and at b94f486 (2026-10-07 11:54Z): same 1 of 1 tests fail.
• A t27c release binary built on 2026-10-06 on macOS aarch64, with the same zig 0.16.0, passes it: t27c test-report specs/ml/recurrent/self_attention.t27 gives 1 pass (vacuous, 0 runtime asserts), and zig test on the generated file passes.

So the reference verdict for this spec depends on the host, not on the commit.

Why

The test is

test forward_basic_case
    given input = default_input()
    when forward(input)
    then true

default_input is declared nowhere in the spec. The W585 scaffold in bootstrap/src/compiler.rs lowers the local to undefined, and t27c gen prints

test "forward_basic_case" {
    const input = undefined;
    forward(input);
    if (!(true)) @panic("assertion failed");
}

forward(input: []const f32) runs for (input) |value| { sum += value; }, so it iterates a slice whose pointer and length are undefined. That is illegal behaviour in Zig: whether the test passes, crashes or fails depends on the host, the build cache and the optimiser. t27b refuses it by name (StmtLocal(scaffold read by callee) is among the blockers in this run), which is the right answer.

Expected

• The reference should not give a verdict on a test that reads an undefined scaffold: when the helper is undeclared and the callee reads the local, t27c gen should refuse the test (or the file) by name rather than print const x = undefined;. reference W585: a declared default_input() is never called, so a callee that reads the scaffold reads dead memory and a false assert passes #7733 asks for the declared-helper half (call the helper).
• Until then the spec is not reference-passing on the lab, so the t27b ledger drops it; the ledger-sync PR for the lab run at 6b6386e (which fails it the same way) does that through tri t27b ratchet --bless.
• The spec itself is a hollow converter test (then true, see default_input has no author: the converter generated 169 specs' worth of hollow tests #5166); giving it a real default_input() and a real assertion would also end the flapping.