kolibri1 TT: the P150 teacher-forced adjudication re-run on the corrected walk — the residual is structural

[lu-zero]

The teacher-forced instrument's earlier counts were computed against a shifted chain (fixed in #3453); this re-runs the adjudication on the corrected walk and delivers the decision input for the P150 band-gate standard. Adjudication table (reproduced exactly, both arms, cold-cache recipe): host-free ON 170 free divergences, 69 tf flips (29 near-tie, 40 HARD, worst 10.6 nats); host-free OFF 158 free divergences, 44 tf flips (27 near-tie, 17 HARD, worst 8.6). The zero-artifact stop condition did not fire — the residual is real.

Findings (corrected walk):

1. Drift origin unchanged: the L0 MoE input (dh2) at 1-bf16-ULP class, every decode step from step 0, all three diverging prompts.
2. The router-flip mechanism HOLDS decisively: top-6-of-384 expert-set overlap vs CPU averages 3.05-3.71 of 6 (full agreement only 6-7% of layer-steps); top6-top7 margins below 0.05 nats at 26-44 of 50 layers at every hard-flip step. A new route_scores probe (RouteDump, STAGE_DUMP-gated) supplies the margins.
3. Distribution: SPREAD (structural) — hard flips on 7 of 8 prompts, BOTH arms, opening bursts AND isolated late flips, first-divergent-layer L0 everywhere. No localized defect site.

Recommended gate-standard input: keep the 2.5-nat/top-5 adjudication; record the P150 band as STRUCTURAL (spread), not fixable-by-a-patch. The levers, in order: (1) host-free shadow reconciliation (halves the hard count: 40 -> ~20 ON), (2) f32-enabling the device attention/norm chain (removes the origin — a perf/substrate question), (3) router-margin widening is a model change, out of scope for this row. The flip step is decided by the LM-head margin at that step, not by where the drift begins.

Raw evidence in the doc; probes committed behind VT_KOLIBRI1_TT_STAGE_DUMP. Issue: ISSUE-LOCAL-01M4ER0E9HHM95YZYJB7T5FECN (Resolution updated with the corrected analysis).

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:zai/glm-5.3-flash [maki]

[localai-org-maint-bot]

@mudler @richiejp Focused source review at 1a905450ce4e7995afa9a205c26e35953ef180ac, compared with current main 50d425bec:

The new evidence needs an explicit mapping between the scratch diagnostic's step labels and the gate's decode positions. In tests/vllm/models/test_kolibri1_tt_b2ii.cpp:988-1016, the final prompt token produces decode-position-0 logits but prints TFstep -1. The next iteration feeds generated_ids[0], produces decode-position-1 logits, and prints TFstep 0. Thus the new document's correlation of SCRATCH dbg TF prompt stages with hard flips at the same numbered decode step is not established by those labels. Please include the mapping used by the parser, or correct the labels and regenerate the affected tables. This finding concerns the new stage/step interpretation; it does not establish that the separate 69/44 gate totals are wrong.

The cited raw evidence is only under /tmp/readj/. Please attach durable logs and the parsing commands so the mapping and derived overlap tables can be checked. The conclusion that no single-op repair can clear the residual also needs evidence beyond its spread across prompts; the report itself identifies a shared L0 origin and further implementation hypotheses.

This PR currently conflicts with main, which already contains the memo/residency fixes and corrected gate indexing. Reconcile that base before requesting acceptance. I have not reproduced the P150 runs or completed the mutation/hardware gates: Python, the native build toolchain, and device assets are unavailable in this session. No merge recommendation yet.