pub fn ba_gates_log(pick: &BaGatesPick, requested: bool, m: u32)Expand description
Say WHICH kernel runs and, when the lever asked for the twin and did not get it, WHICH guard refused — once per process PER BRANCH.
Once per branch, not once per call: this runs inside the per-layer prefill step, so on a 48-GDN-layer model an unconditional line is 48 of them per request. But the verdict CHANGES between calls — the token count is an argument — so a single flag records whichever M arrived first and then lies for the life of the process. Round 15 measured exactly that: a 27-token smoke request before the first real prefill, and the twin’s own line never printed on any of the five serve cells that ran it. With a flag per branch a serve log now carries BOTH lines, each naming the shape it was reached at.