Glm4MoeLite's only vs-vLLM assertion cannot fail: the gap artifact is identically zero and the goldens match 69/128

Row: MODEL-TEXT-glm4-moe-lite-glm4-moe-lite-for-causal-lm

What

docs/FEATURES.md:155 advertises Glm4MoeLiteForCausalLM as "near-tie 8/8 vs
vLLM 0.25.0". The 8/8 is what the test prints. It is produced by a predicate
that cannot fail, on an artifact that is identically zero.

Measured, at ca07f6e94

The committed gap artifact,
tests/parity/goldens/glm4_moe_lite_greedy/neartie_gap_mnats.npy:

$ python3 -c "import numpy;a=numpy.load('tests/parity/goldens/glm4_moe_lite_greedy/neartie_gap_mnats.npy');print(a.shape,a.dtype,a.min(),a.max(),(a==0).all())"
(8, 16) int32 0 0 True

All 128 values are exactly 0. The only vs-vLLM assertion in
tests/vllm/models/test_glm4_moe_lite_paged_engine.cpp reads that array and
tests mn > kNearTieMnats with constexpr int32_t kNearTieMnats = 500; (:61,
:308). 0 > 500 is false at every position, so prompt_ok is true for all 8
prompts unconditionally and CHECK_MESSAGE(prompt_ok, ...) (:336) has no
failure mode. It is a positive control that provably cannot trip.

What the goldens actually say, comparing the committed greedy_ids.npy (the
vLLM oracle) against our_ids.npy:

positions match     69 of 128   (53.9%)
prompts exact        1 of 8      (prompt 7 only)
per-prompt mismatch [2, 6, 6, 12, 12, 11, 10, 0]

Prompts 3 and 4 differ in 12 of their 16 tokens. Nothing asserts against that.

Two further weaknesses in the same file

1. The exact-equality assertion at :283-287 compares against our_ids.npy,
   labelled in the file itself as "ANCHOR: the committed gaps describe OUR
   engine's exact sequence". That is a self-regression check, not an oracle
   check.
2. The header at :22-24 says a divergence counts as a near-tie "ONLY where
   that gap is within kNearTieMnats AND our token is inside vLLM's top-K". The
   second conjunct is not implemented: greedy_dist.npy is read only for the
   self-determinism count at :140-163, never for top-K membership.
3. .agents/model-matrix.md:120 records "vLLM K=5 self-deterministic -> STRICT
   bar". Per CLAUDE.md §Gates, a distributional gate is licensed only when the
   oracle's greedy decode is non-deterministic. It is deterministic here, so the
   STRICT bar is what is owed, and the shipped gate asserts neither.

Not fixed in the flow that found it

Found while repairing docs/FEATURES.md's dead citations (#2825). The doc repair
publishes the real numbers in place of "near-tie 8/8". Repairing the GATE is a
different unit of work: it needs a red-first argument about which bar applies,
and regenerating an artifact that is currently all zeros.

What would close it

A gap artifact that is not identically zero, or a STRICT assertion consistent
with the deterministic oracle -- and a mutation proving the predicate can fail.

Related: #2825, #2794 (goldens predate the current pin).

Start with tests/vllm/models/test_glm4_moe_lite_paged_engine.cpp, the committed files under tests/parity/goldens/glm4_moe_lite_greedy/, and CLAUDE.md §Gates. Run the existing Glm4MoeLite parity test and inspect the oracle, gap, and top-K inputs. Done means the gate uses the applicable bar, the artifact is regenerated, and a mutation proves the assertion can fail.