fix(gate): the online gate has no key for the nvidia 27B, and a skipped model gate records as passed - #229

…, and make a model gate prove it ran

Review of #229 (issue #222) returned CHANGES-REQUIRED. The examples/vllm-server
repair and its CMake-parsing drift guard were confirmed correct; the 27n half was
not. This resolves the findings.

F1 -- the two shell-side 27n assertions had ZERO effective coverage. Each was a
substring satisfied by the OTHER line: `${model} == 27n` occurs inside the
--model guard, and `${model} == 27 || ${model} == 27n` occurs inside the
test_name dispatch. Proven RED-free both ways on the shipped tests: reverting
:243 to `27` only left 8/8 GREEN while 27n fell through to the else-arm
(test_name=mxfp4_smoke_battery, gate_target=vllm-cli -- benched with no golden,
verbatim what the docstring claims to prevent), and dropping 27n from the guard
also left 8/8 GREEN. Both assertions now PARSE the driver: accepted_model_keys()
reads the guard LINE and requires set equality with MODEL_REVISIONS, and
correctness_gate_by_model_key() replays the if/elif/else chain and asserts the
whole key->test_name mapping. Re-verified: mutation A now fails with
'test_qwen27_paged_engine' != 'mxfp4_smoke_battery', mutation B with "27n not
found in frozenset({'q3mxfp4','27','35'})".

F2 -- `if __name__ == "__main__"` sat BEFORE the second class, so
`python3 tests/tools/test_online_gate_server_binary.py` ran 4 of 8 and printed
OK. Moved to the end. The module's sibling convention imports tools.bench at
module scope, so REPO_ROOT joins sys.path or the direct invocation would now die
on ImportError instead of under-running.

F3 -- the same defect was unfixed in a sibling driver, and the guard could not
see it because it hardcoded two file paths. It now SCANS the live surfaces
(scripts, tools, tests, examples, src, include, docs, cmake, release, docker,
.github, README, root CMakeLists) for three shapes: a path component in front of
examples/server, pathlib's "examples" / "server", and a BASENAME check against
"server". That last shape found a site nobody had listed: PR #229 made
online_gate.py record-execution write examples/vllm-server while
gdn_packed_component.py:1559 still required `server.name != "server"` -- so the
component driver rejected the manifest the shared harness had just written for
it. Seven live sites fixed: dgx-gdn-packed-component.sh x3 (including the live
server command array), gdn_packed_component.py, its fixture,
run_serve_low.py's recipe manifest and docs/USAGE.md.

NEEDS_DECISION -- README.md:168 and :368 are NOT fixed here and remain user-
visible. scripts/check-doc-checkpoint.py refuses any README change whose commit
does not also touch a landing source, and the file that decides the name the
quick start tells a user to type -- examples/CMakeLists.txt OUTPUT_NAME -- is not
on that list, while examples/{cli,server}/main.cpp are. This change has no honest
reason to edit a listed file; manufacturing a cosmetic edit to one, or relaxing
the checker, would both be weakening a gate to make a change pass. `4f24ff44`
recorded this exact class ("a gate that rejects a true statement is not
protecting anything") and repaired it by adding a REAL source to the list. The
same decision is owed here. The scan therefore carries README in a one-file
SCAN_BLOCKED_ON_POLICY set that names the blocker, guarded by a case asserting
the set holds exactly that file AND that the debt still exists -- so the
exclusion cannot outlive it or quietly grow.

.agents/ and benchmarks/ are deliberately NOT scanned: they quote the command a
past campaign actually ran, and rewriting them would falsify evidence rather
than repair a consumer. scripts/dgx-online-serving.sh's `build_targets=(server)`
is correct and untouched -- that is the CMake TARGET name.

F4 -- the documented no-OUTPUT_NAME fallback was unsatisfiable: it demanded
'"examples" / "server"' present AND absent. Removing the override and setting
both harnesses to examples/server (correct for that world) yielded 2 red with
the self-refuting message "still resolves the server as examples/server ...; the
build emits examples/server". The stale-spelling assertion is now skipped when
the target name IS the output name.

F5 -- the 27n correctness precondition could pass VACUOUSLY and nothing recorded
it. test_qwen27_paged_engine emits a MESSAGE and returns 0 when the UNSLOTH
snapshot is absent; record_model_gate then wrote "passed": true unconditionally.
A box with @0893e160 but not @890bdef7 -- the EXPECTED shape for a key whose
whole point is a different checkpoint -- produced 27n.json marked passed with
zero tokens compared. Now MODEL_GATE_CONTRACTS carries, per dispatched gate, the
proof line it prints only after comparing tokens and the revision its committed
goldens belong to. record_model_gate fails closed without the proof line and
records golden_revision, model_revision and golden_covers_benched_checkpoint
(deliberately false for 27n); the driver runs ctest -V so a PASSING run's output
reaches the log at all; and the summary revalidates all of it. Two drift guards
parse the markers out of the gates' own sources and the revision out of
hf_snapshot.h, so a reworded MESSAGE fails here rather than silently disarming
the check.

F6 -- #222 added to the roadmap issue table, and the 27n key, its pinned
revision, its build-sanity-only scope and its stop condition recorded in the
EXISTING cuda-online-serving-gate spec. No new row.

F7 -- the trace-only refusal said "defined only for the Qwen3.6-27B dense
graph", which 27n IS. It now names the checkpoint whose graph contracts were
captured.

F8 noted, no action: 27n widens the cross-model summary default exactly as
q3mxfp4 already does; not newly broken.

Fixtures in test_online_gate_summary.py and test_online_gate_client.py are
re-derived from the real gate log shape (they carried "passed\n"), not relaxed.

GATES: 18 modules under tests/tools, 206 tests, GREEN (195 before; 11 new cases,
every one shown red first). scripts/agent-preflight.sh exits 0.

Refs #222, #213

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:claude-opus-5 [ClaudeCode]

…o four decimals

First measurement of nvidia/Qwen3.6-27B-NVFP4 @0893e160 through a reproducible
recipe. Every prior number for it came from an ad-hoc harness, because the
canonical gate could not start a server (#222) and there was no model key for it
until #229 added 27n.

vllm.cpp 9ff2742e; oracle pip vLLM 0.25.0 + FlashInfer 0.6.13 in its PRODUCTION
GRAPHED configuration, never --enforce-eager. Build guards confirmed present:
FlashAttention-2 ENABLED for [121a], VLLM_CPP_TRITON resolved ON with real
sm_121a AOT variants. Idle box, medians of 3, drop_caches between legs, noise
band measured BEFORE any delta was read.

c1 0.8289x, c2 0.8461x, c4 0.8529x, c8 0.8639x. This CONFIRMS the previously
recorded 0.843-0.861x band, so the ad-hoc numbers were not wrong; they are kept
as a superseded row rather than deleted.

The step is now attributed. A decode-only nsys two-length diff with
--cuda-graph-trace=node, BOTH arms in the same tool, node-level tracing proven by
integral launch counts and two capture pairs agreeing to 0.005%:

  +8.6414 lm_head/Marlin  +7.6068 fp8 tower  +0.0532 splitK  +1.0279 other
  = 17.3292 ms/step   vs measured 17.3292

At 99.2/99.3% GPU-busy the c1 step is GPU-bound on both arms, so the whole gap is
kernel time: no host, scheduler or launch-overhead lever exists here. Both levers
are TRAFFIC/FUSION, not kernel quality. Our Marlin reaches 207.9 GiB/s against
vLLM's 210.0 on the same combined shape, and both arms read identical fp8 bytes;
we simply move more of them, and issue 96 GDN input GEMMs where vLLM issues 48
merged qkvz.

Recorded honestly rather than flatteringly: RSS is 1.612x and BELOW FLOOR,
c16/c32 are NOT MEASURED (one attempt lost the denominator to a foreign build
contending mid-timing, the other to a host OOM-reboot), GPU memory is PENDING
because the instrument returns N/A on this unified part, and our c1 ratio is
host-memory-state sensitive in a way vLLM's is not (0.7604 vs 0.8289 across a
reboot, protocols also differing, variables not separated).

Startup is the one axis we win outright: 33.38 s vs 182.41 s, 5.46x faster.

Written as table ROWS, not prose: the checker's budget is what forced that, and
it was right that a scoreboard should stay a scoreboard.

Refs #213, #252

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:claude-opus-5 [ClaudeCode]