L4: the kernel gate is a tracked exception, and the oracle denominator is now gated (#607, #414)

[localai-bot]

L4 of the multimodal track had one line of spec: mirror the kernel gate, or
record an explicit tracked exception with the #414 cross-reference. This takes
the exception arm, argues it in .agents/specs/multimodal-track.md §1.6, and
then repairs the benchmark denominator that the exception leaves exposed.

The decision: tracked exception, not a mirror

vllm/model_executor/models/qwen3_next.py:324-331 conjoins text_only into
use_fused_qk_norm_rope_gate. We do not mirror that conjunct, and the reason is
that mirroring is not representable at our seams rather than a preference.

Upstream conjoins it because its fused Triton kernel indexes cos_sin_cache by
a 1-D positions vector and cannot express MRoPE. qwen3_next.py:323 says so
in as many words, # TODO: support MRoPE, and the fused branch at :344-352
restates it: the T-row shortcut is exact only because fusion is text-only. The
arm it falls back to is complete, because upstream's eager self.rotary_emb IS
the MRoPE module. So the conjunct is a limitation guard, not a behaviour.

Our two arms are not those two arms. vt::AttnQkNormRopeGate takes no positions
argument at all; it consumes a precomputed per-token cos_sin cache
[T, rotary_dim], which qwen3_5.cpp::BuildMropeCosSinHost fills with the
interleaved 3-section MRoPE axis selection for the M3-b image and M3d video
paths. Our FUSED arm is the MRoPE arm. Our eager arm calls vt::RopeNeox on the
1-D positions vector and has no MRoPE spelling. Conjoining text_only would
therefore select 1-D RoPE on exactly the configuration the conjunct exists to
protect, silently, since every tensor shape still agrees, and would break the
landed STRICT 32/32 image and video gates in the name of mirroring a TODO.

Owner ENG-MM-INPUT-PIPELINE, cross-referenced to #414. If upstream lands MRoPE
support and drops the conjunct, the exception is retired by upstream converging
on us, and the record is deleted rather than migrated.

What the exception does not excuse

#414's defect is a benchmark configuration, not a kernel.
vllm/config/multimodal.py:78 defaults language_model_only to False, so a
Qwen3.6 checkpoint hands the oracle text_only == False and four ops per
full-attention layer where its own production configuration issues one, against
our single fused launch. That compares two algorithms, and the error runs in our
favour, hardest on TTFT.

The correct denominator is vLLM WITH the flag, on two grounds that agree: it is
a production configuration, passed by 43 of the 157 official recipes, and
AGENTS.md §Gates requires vLLM's production configuration rather than a
handicapped one, the same rule that forbids --enforce-eager.

The 2026-08-13 clock-controlled series repaired the RUN. It did not repair the
HARNESS. scripts/dgx-online-serving.sh still launched the oracle without the
flag while tools/bench/run_serve_low.py passed it, so the two disagreed about
the oracle's own configuration and the next canonical campaign would have
produced another flattered set. A grep for the flag read as coverage while the
canonical driver lacked it, which is why this ships as a checker.

The 27/27n/35 arm now passes the flag. The q3mxfp4 arm is marked exempt with its
reason: a dense Qwen3-8B checkpoint has no multimodal_config, so
qwen3_next.py:325 already resolves text_only == True and the flag decides
nothing there.

The gate

scripts/check-oracle-denominator-flags.py, wired into agent-preflight.sh and
CI. Scoped to CLI serve launches, the surface the published ratios come from,
with its docstring naming what it deliberately does not cover. It reads harness
sources and holds its expectation in its own code, so it is not a surface every
pull request must edit. --json exists so the test pins the discovered SET
rather than the violation count, and finding nothing exits 2 rather than 0,
because a broken detector and a clean tree are otherwise indistinguishable.

RED before, on the tree as it stood: exit 1, naming dgx-online-serving.sh:487
and :498 of 3 discovered launches, with run_serve_low.py:552 already green.
GREEN after: 3 launches, 2 carrying the flag, 1 exempt.
tests/scripts/test_check_oracle_denominator_flags.py 13/13, including the
mutation that reconstructs the pre-L4 state and requires exactly one violation.

The first cut of _ORACLE_CLIENT named the two variables this tree happens to
use, ${client} and ${VLLM_ORACLE}, which made the detector a list of
spellings: a later harness writing "$ORACLE" serve would have launched the
oracle unscanned, and the exit-2 arm cannot catch that, because the other
launches keep the count above zero. It now matches the binary spelled out OR any
$-expansion, since what identifies a launch is serve sitting directly after
the binary rather than the name of the variable holding it. bench serve stays
a client, because the token before serve there is bench. RED first:
test_any_variable_spelling_of_the_client_is_still_a_launch failed 4/4 of its
spellings on the unwidened pattern, each reporting exit 2 and "found NO oracle
server launch at all"; test_bench_serve_stays_a_client_under_any_variable_spelling
holds the widening off the timed client; the real tree discovers the same three
launches, so no false positive was added.

Record impact

One published number carried the defect unmarked, and finding it corrects this
branch's own spec commit, which claimed nothing currently binding was affected
because it only looked at the two gate models.

docs/BENCHMARKS.md's Qwen3.5-4B 1.0283x ran its vLLM arm through
vllm_closed_loop_metrics.py, which never passes the field; upstream registers
Qwen/Qwen3.5-4B as Qwen3_5ForConditionalGeneration
(tests/models/registry.py:1322-1324); and git log -S dates the conjunct to
16282a9c4, 2026-06-10, two months before that 2026-08-07 run. The oracle ran
unfused. The direction is known and it flatters us; the magnitude is NOT
measured and is not estimated. The row keeps its values and gains the
attribution, with the chain in .agents/benchmark-record.md.

Keeping its values means all of them. The first cut of this branch rewrote that
row and dropped the VRAM +118.7 MiB axis while marking the other three, and
docs/STATUS.md's row cell lost Atomic pretoken exact, the direction word
slower and the GDN-stack note with it. Annotating evidence is the rule;
deleting it is not, whichever direction it points. Every one is restored and
marked. The denominator column now also records that the oracle carried a vision
tower: multimodal.py:326 zeroes every modality limit under the flag, and
interfaces.py:293 then builds the tower under no_init_weights as a
StageMissingLayer, which qwen3_5.py:422 wraps the Qwen3.5 visual tower in.
The axis is marked not like-for-like and stays OPEN rather than being given a
sign: vLLM sizes its KV cache to fill --gpu-memory-utilization, so memory a
tower does not take is memory the cache does, and this campaign's evidence file
records no utilization setting. #1345's re-measurement is what settles it.

Not affected, checked rather than assumed: #414 reaches a figure only where the
full-attention layers are Qwen3NextAttention AND the checkpoint loads as a
*ForConditionalGeneration. The Qwen3.8-27B rows ran through run_serve_low.py,
which has always passed the flag. The Qwen3.5-4B GDN prefill kernel row is
conv and post-conv timing on the linear-attention path.

Owed, filed in flow

Both name ENG-MM-INPUT-PIPELINE in .agents/issue-index.md and are listed
under ## Owed in the spec. Neither is fixed here, and neither is deferred
silently.

• VT_FUSE_ATTN_PREAMBLE=0 on the Qwen3.6 MRoPE path silently applies 1-D RoPE instead of refusing #1340 is the same seam asymmetry read the other way: VT_FUSE_ATTN_PREAMBLE=0
  on the MRoPE path substitutes 1-D RoPE instead of refusing, so the model emits
  wrong tokens with no diagnostic. Its gate is a VL token-exactness run through
  ModelRegistry::Forward on a GPU.
• Three in-process bench harnesses construct the vLLM oracle with language_model_only left at False and no way to set it #1345 is Canonical grids measure vLLM WITHOUT --language-model-only: the oracle runs the UNFUSED QK-norm+RoPE+gate path while we run the fused one — our published ratios are FLATTERED #414 on the in-process surface: profile_vllm_online_gate.py:209,
  vllm_closed_loop_metrics.py:160 and dump_vllm_tokens.py:34 construct the
  oracle with language_model_only at its default and expose no way to set it.
  They take --model as a path, so no static rule can decide them, and changing
  what they hand the oracle moves future denominators.

Gates

No product code path changed, so no token gate and no measurement is claimed.
scripts/agent-preflight.sh all gates green, on an idle box (load 1.3) with
15G free, so neither the contention nor the disk-pressure failure mode applies.
check-commit-trailers.py and check-commit-style.py over origin/main..HEAD
both OK.

The review-repair commit was gated on a loaded box and reports its condition per
result, because a full-disk or contended run produces false failures and false
passes alike: check-public-doc-tables.py exit 0 (59G free, load 43);
check-oracle-denominator-flags.py exit 0, 3 launches (59G, load 48); its suite
Ran 13 tests/OK (59G, load 47); check-agent-record.py exit 0.

Two upstream anchors are corrected in the same pass: # TODO: support MRoPE is
qwen3_next.py:323 at the pin, not :322, and fused_qk_rmsnorm_rope_gate is
def-ed at fused_qk_norm_rope.py:117, not :119. Every other anchor in the
§1.6 table was re-read at 555967922 and resolves.

FOLLOWING_AGENTS_PROTOCOL

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