fix(MODEL-TEXT-GLM4-MOE-LITE-GATE-2839): the near-tie predicate cannot fire, so apply the bar the oracle capture licenses -- and it FAILS 69/128

[localai-org-maint-bot]

THIS PULL REQUEST MAKES A GATE FAIL, AND THAT IS THE RESULT. Glm4MoeLiteForCausalLM
matches the pinned vLLM 0.25.0 oracle at 69 of 128 positions and on 1 of 8
prompts. Nothing asserted that before, because the only vs-vLLM assertion in the
tree had no failure mode and the case that carried it reported a pass with zero
assertions on every host without a 31.2B checkpoint. Do not soften the new case
to merge this; the two admissible exits are named below.

What was true at 3e246b34f

D1 — the only vs-vLLM assertion cannot fail. The gate tests
mn > kNearTieMnats with kNearTieMnats = 500 against
tests/parity/goldens/glm4_moe_lite_greedy/neartie_gap_mnats.npy, which is
(8,16) int32, min 0 max 0. 0 > 500 is false at every position, so
prompt_ok is unconditionally true and CHECK_MESSAGE(prompt_ok, ...) is a
positive control that cannot trip.

D2 — the licensed bar is STRICT, and nothing asserted it. The gate derives
vLLM's own self-determinism from greedy_dist.npy over K=5, CHECKs that it is
zero, prints DETERMINISTIC -> the STRICT bar is well-posed — and then applies
the near-tie band anyway. CLAUDE.md §Gates permits a distributional gate "only
when the oracle's greedy decode is non-deterministic".

D3 — a skip wearing a pass. With no snapshot the body emitted
MESSAGE("SKIP ...") and returned. Measured on this host, before:

[doctest] test cases: 1 | 1 passed | 0 failed | 0 skipped
[doctest] assertions: 0 | 0 passed | 0 failed
[doctest] Status: SUCCESS!

A SACRED gate reporting SUCCESS off zero assertions. The same shape hid two more
early returns, and both of their conditions are impossible: all five artifacts
are committed.

The change

The checkpoint predicate is a doctest::skip, so absence reads as skipped. The
two golden-absence returns are REQUIREs, because a committed artifact going
missing is a broken checkout and not an environment. greedy_dist.npy is
required, and its verdict selects the bar in code: zero multi-valued cells means
STRICT, and only a non-zero count admits the band — which now also applies the
second conjunct this file's header states and nothing implemented, our token
inside vLLM's top-K.

The third piece is what makes any of it observable here. Everything the SACRED
case compares is committed, including this engine's own ids, which that case
REQUIREs the live engine to reproduce token for token before comparing anything.
So a new case applies the licensed bar to those artifacts with no checkpoint and
no GPU, and runs wherever the suite runs.

GREEN-to-RED, both counts

First measured on 3e246b34f, then re-measured after rebasing onto 3f12c617d,
31 commits later: identical counts, identical numbers, so nothing here was
falsified by the merge.

[doctest] test cases:  1 | 0 passed | 1 failed | 1 skipped
[doctest] assertions: 10 | 8 passed | 2 failed

MESSAGE: vLLM self-determinism over K=5 batch=1 runs — multi-valued cells = 0
MESSAGE: committed teacher-forced gap max = 0 mnats against a 500 mnat band,
         so `gap > band` has no failure mode on this artifact
MESSAGE: STRICT vs vLLM 0.25.0: positions 69/128, prompts exact 1/8,
         per-prompt mismatch [2,6,6,12,12,11,10,0]
ERROR:   CHECK( exact_positions == N * T ) -> CHECK( 69 == 128 )
ERROR:   CHECK( exact_prompts   == N )     -> CHECK( 1 == 8 )

Mutations

Both are artifact mutations on the goldens the case reads, each restored and each
restore verified by sha256 against
d9b2e31a…/77da140b…/65e3c580…/adfc42ab….

#	Mutation	Result
G1	perturb one cell of greedy_dist.npy so the oracle reads non-deterministic	the bar-selection REQUIRE fires and aborts before the comparison: 1 failed, 3 assertions, 1 failed
G2	replace our_ids.npy with greedy_ids.npy (positive control)	the case goes GREEN, 1 passed, 10 of 10 assertions — so the failure is data-driven, not an unconditional assertion

G2 is the one that matters: it proves the 69/128 is being read out of the
artifacts rather than asserted.

A finding recorded but deliberately NOT gated

The gap artifact is very likely not a measurement. At a prompt's FIRST divergent
position our prefix is identical to the oracle's, so vLLM's teacher-forced argmax
there IS greedy_ids[i,j] by construction and our token is a different one, so
max(0, arg_lp - our_lp) must be strictly positive. It is 0 at all seven first
divergences (prompts 0..6 at positions 14, 6, 6, 4, 4, 4, 6), and the generating
script's 99_999_000 outside-top-20 sentinel appears nowhere. Every divergent
token of ours would have to be inside vLLM's top-20 AND within 0.0005 nats of its
argmax, seven times out of seven at the one position where the prefixes provably
agree.

This is not gated, because a "gap at a first divergence is positive" assertion
could in principle be defeated by genuine sub-millinat ties, and the STRICT bar
settles the same question without that risk.

What is NOT verified

The engine case is UNRUN. No zai-org/GLM-4.7-Flash snapshot exists on this host
and none is on the NAS, so nothing here can execute a 31.2B forward. It compiles
clean; it has not run. docs/USAGE.md also carries no checkpoint row for this
architecture — no revision, no sha256 — which the spec records as owed rather
than inventing values off an artifact this host does not have.

The two admissible exits

Either a forward repair that makes the new case 128/128, or a re-capture of
neartie_gap_mnats.npy on the checkpoint plus an explicit ratification argument
for a distributional bar over a deterministic oracle. Widening the assertion is
neither, and docs/FEATURES.md already says so: the shipped 8/8 is not a measurement.

The shape is repository-wide, and is now owned

D3 is one instance of MESSAGE(...SKIP...); return;, which occurs at 228 sites
in 72 files under tests/. Measured at 3f12c617d,
test_deepseek_v2_paged_engine -- the SACRED DeepSeek-V2 gate -- reports
1 passed | 0 failed, assertions: 0, Status: SUCCESS! on this host. That is
the same defect, unfixed, on a different row. Repairing 228 sites is a campaign,
so it is filed as #2918 and
listed under this spec's ## Owed O4. This pull request repairs the one instance
it touches.

Closes #2839.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]