The neartie gap artifacts are identically zero in 12 of 18 golden dirs while our_ids diverges: not measurements

Row: - (owned under ## Owed in .agents/specs/glm4-moe-lite-router-f32.md)

neartie_gap_mnats.npy is the artifact every near-tie-banded model gate reads to
decide whether a divergence is admissible. In 12 of the 18 golden directories
that carry one it is identically zero at every position, while our_ids.npy
diverges from greedy_ids.npy at 13 to 83 positions in the same directory.
Measured at c796fea41:

dir                          gap min/max  nonzero   our_ids vs greedy_ids
deepseek_v2_greedy               0/250        1/128        36/128 divergent
glm4_moe_lite_greedy               0/0        0/128        59/128 divergent
internlm2_greedy_1_8b              0/0        0/256        42/256 divergent
internlm3_greedy_8b                0/0        0/256        22/256 divergent
llama_greedy_1b                    0/0        0/256        13/256 divergent
minicpm3_greedy_4b                 0/0        0/256        20/256 divergent
minicpm_greedy_2b                  0/0        0/256        34/256 divergent
mistral_greedy_7b                  0/0        0/256         6/256 divergent
olmo2_greedy_1b                   0/94        2/256        15/256 divergent
phi2_greedy_2_7b                 0/250        6/256        75/256 divergent
phi4_14b_greedy                    0/0        0/256        37/256 divergent
phi4_mini_greedy                0/1000        9/256        83/256 divergent
qwen35_greedy_0_8b               0/125        2/256         8/256 divergent
qwen3_32b_nvfp4a16_greedy          0/0         0/96        43/96  divergent
qwen3_greedy_0_6b                0/125        1/256        55/256 divergent
qwen3_greedy_4b                  0/250        2/256        43/256 divergent
qwen3coder_greedy                  0/0         0/96         1/96  divergent
yi_greedy_coder_1_5b             0/125        2/256        38/256 divergent

Reproduce with numpy over tests/parity/goldens/*/; no checkpoint and no GPU
are needed.

Why zero is not a possible measurement at a FIRST divergence

The generating scripts (scripts/*-neartie-gap.py) teacher-force the pinned
oracle on OUR token sequence and record
max(0, argmax_logprob - our_token_logprob) in milli-nats. At a prompt's first
divergent position our prefix is byte-identical to the oracle's, so the oracle's
teacher-forced argmax there is by construction the token its own free-running
greedy decode emitted — greedy_ids[i,j] — and ours is a different token. The
gap must therefore be strictly positive unless the two logprobs are exactly
equal in the oracle's own float output.

For glm4_moe_lite_greedy it reads 0 at all seven first divergences (prompts
0..6, positions 14, 6, 6, 4, 4, 4, 6), and the scripts' 99_999_000
outside-top-20 sentinel appears nowhere in the file. our_ids.npy is written by
the same script in the same call as the gap array, so the script demonstrably
ran to completion; the zeros are what it computed, not a file it never wrote.

The same argument applies to every all-zero directory above.

Consequence

Wherever a gate's admissibility test is gap > band, an all-zero artifact makes
that test unconditionally true and the assertion has no failure mode. #2839
measured exactly that for Glm4MoeLiteForCausalLM, where the artifact is
(8,16) int32, min 0 max 0, against a 500 mnat band.

This issue is about the ARTIFACT CLASS. It does not claim any particular gate is
wrong beyond #2839's, and it does not propose widening or narrowing a band. The
first thing it asks for is a re-capture on a host that has the checkpoints, with
the script's own diagnostic output preserved, so that whether the zeros come
from the script's use of prompt_logprobs or from the artifacts being written
without a real run can be settled by evidence rather than by argument.

Start with scripts/-neartie-gap.py and inspect the artifacts under tests/parity/goldens/; reproduce the reported zero gaps with numpy. Re-capture on a host with the checkpoints while preserving the script's diagnostic output, then determine whether the zeros come from prompt_logprobs or from artifacts being written without a real run.