Function gdn_init_spine_line 

pub fn gdn_init_spine_line(tc_spine_bound: bool, scalar_entry: &str) -> String

GDN state spine: … — the line qwen3_ssm::init prints ONCE PER LAYER as it binds the handles, before any prefill has run.

Why it is not simply the scalar entry’s name

H100 round 14 (h100-round14-report.md, anomaly 2). A serve with the tensor-core spine live logged both of these:

   48  qwen3_ssm::init: GDN state spine: gated_delta_rule_chunk_delta_h_vfused
14400  GDN state spine: gated_delta_rule_chunk_delta_h_tcfuse_x2 (…)

48 init lines naming the scalar parent, one per layer, ahead of 14 400 dispatch lines naming the kernel that actually ran. f9ae638 fixed the dispatch line; the init line was not touched, and it is the FIRST GDN line a reader meets in a log they opened to answer “did the lever engage?” — so it read as “the TC spine is not engaged” on a serve where it was.

The line now reads the handle the probe resolved, which is the same bit the dispatch reads: with the [defaults] gdn_prefill_tc handle bound, the spine that will launch is GDN_TC_SPINE_ENTRY and the line says so. The scalar entry stays bound underneath — the dispatch’s shape guards fall back to it, and the dispatch logs whichever one it launched — but it is no longer what this line NAMES, which was the whole defect.