record(MODEL-TEXT-nemotron-h): the row is blocked on a kernel that landed, and a file glob is why nobody saw it

[localai-bot]

The NemotronH model-matrix row still says the architecture is BLOCKED on
KERNEL-SSM-MAMBA (#496)
because "the Mamba2 SSD core is unported". That kernel landed on 2026-08-13.
Thirty-nine Nemotron commits have landed on main since the row was written,
including the paged forward, the device MoE arm and the ABI driver, and the row
has read INVENTORIED with ☐ required for its spec and none for its
evidence throughout.

FOLLOWING_AGENTS_PROTOCOL

Why this was rewritten rather than merged, and why it was not closed

This branch was cut on 2026-08-16 and main moved 57 commits past it. Main's
version of this row is NEWER than this branch's — A2-R 598226e96, A2-P
a6df72777, A2-Q2a and the A3 driver c83b96934 all appended to the same cell
after the branch was cut. A three-way merge would have carried this branch's
"the step still REFUSES paged or batched decode because the reached forward is
the HOST reference" over the top of a paged forward that has since landed.

So no keyed record was auto-merged. Every file was taken from main
whole, the tree was asserted byte-identical to it before any edit, and the
scoped edit was written again against what main actually says. The
## Records rule names this shape; the branch's merge commits are squashed
away by squash_merge_commit_message = PR_BODY.

main moved twice more during this run. The third merge (10fe7f475, #1248)
did touch .agents/model-matrix.md, docs/BENCHMARKS.md and docs/STATUS.md,
and Git auto-merged all three. That was verified rather than trusted: the diff
against 10fe7f475 is 13 lines in the matrix and one line each in the two docs,
every one of them this branch's own, with #1248's DSpark row, benchmark row and
status paragraph byte-for-byte equal to main's.

Closing this as superseded was the other candidate and it is wrong, because
the claim it exists to retire is still on main. Verified at 10fe7f475,
.agents/model-matrix.md:285:

Cell	On main today	Verdict
BLOCKED on KERNEL-SSM-MAMBA (#496), "the Mamba2 SSD core is unported"	still there	FALSE
non-gated relu² MoE "exists nowhere locally"	still there	FALSE
ModelOpt MIXED_PRECISION loading "exists nowhere locally"	still there	FALSE
the DeepSeek-style MTP head "exists nowhere locally"	still there	TRUE, owed as W5
State	INVENTORIED	stale
Spike/spec	☐ required	stale: five specs are committed
Our code + tests/evidence	none	stale

docs/FEATURES.md and docs/STATUS.md on main are already current and carry
none of this, so FEATURES is not touched at all and STATUS gains one clause.
The stale record IS the work, which is the case AGENTS.md ## Work happens in a worktree explicitly admits for a record-only pull request.

The blocker, and why a directory listing hides it

#496 landed its host arm at
47960a009 (W1) and its CUDA arm at 43a6c5518 (W2). The kernel is
src/vt/cuda/cuda_mamba2_ssd.cuh — a 692-line header, #included at
cuda_gdn.cu:48 and registered as kMamba2ChunkScan at cuda_gdn.cu:6669, not
a translation unit of its own. A src/vt/*mamba* FILE GLOB therefore returns
nothing and reads as absence, and that glob is the search both the row's text and
#1074 rested on.
nemotron_h.cpp:597 calls vt::Mamba2ChunkScan today, and
nemotron_h.cpp:354 calls vt::MoeRelu2. #496 stays open for its GENERIC
MambaSpec producer, which this architecture does not need, because A1 made the
runner read the model's own KV spec.

Worth carrying: #496's own issue TITLE still asserts "Mamba2 SSD core is
unported", so the false claim has a second live copy that this change cannot
reach.

The state, and why it is not ACTIVE

INVENTORIED -> PARTIAL, mark 🚧, with the rollup (INVENTORIED 324->323,
PARTIAL 22->23), the engaged-architecture checklist entry, and the two prose
counts (53->54, 324->323) in the same file.

ACTIVE is refused on a rule. scripts/check-agent-record.py:1257-1266 requires
an ACTIVE row to name a CLAIM-* row that a claim source carries. Measured:
no file under .agents/claims/ (15 of them) names
MODEL-TEXT-nemotron-h-nemotron-hfor-causal-lm, and authoring one for another
session's in-flight work would be a fabricated record. PARTIAL is in
EVIDENCED_STATES, whose contract is exact code and test anchors, which this row
now has.

The A3 gate passed, and my first version of this row said PENDING

I got this one wrong and the correction is instructive, so it stays in the
record. I searched origin/main and five row/MODEL-NEMOTRON-H-* branches for
a GB10 96/96, found none, and concluded none existed. Then I took
#1221's pull request body as
the current state. Its "Still owed: the sm_121a green-after re-run of the full
96-token gate under this fix" was true when written and stale by the time I read
it. A body is a snapshot of its author's knowledge at write time, not a live
field. No amount of branch searching corrects for that, because the artefact was
never in git.

It is on the NAS, at /usr/local/nas_share/rc/nh1157/. Verbatim from
gate_fixed.out, read directly rather than taken on report:

TOKEN MATCH: 96/96 over 3 prompt(s) (full rows=3, short rows=0, mode=decode)
STRICT PASS

against the pinned oracle vllm=0.23.1rc1.dev1511+g555967922 on
nemotron-3.5-lightning-30b-nvfp4 at revision 29f2d1746d8f41e316523194b19018707749b1b1.

It is the device leg. The binary is libvllm 0.0.3+cuda, and cfg.log
records fp4-mma, cutlass-nvfp4 and cutlass-fp8 as ENABLED for [121a]. The
run logs Asynchronous scheduling is enabled (max_concurrent_batches=2), which
is precisely where device_token_ids is non-null; on the host queue it is always
null and the #1157 defect
cannot arise. The decisive evidence is neither of those: the same binary on the
same checkpoint with only nemotron_h_device.cpp reverted to the fix's
parent scores 4/24 (full rows=0, short rows=3) and bails at 8 generated
tokens. A host-leg run would have been unmoved by that revert. The delta is the
proof; the pass on its own is not.

The pass belongs to #1221, not to main. That branch is
row/MODEL-NEMOTRON-H-ABI-A2P-1157-fix at 6e9e8955, OPEN and CONFLICTING.
main's last touch of nemotron_h_device.cpp is a6df72777 (A2-P). A gate that
passed and a tree that is gated are different facts, and every row here keeps
them apart: the gate PASSES, the merge is owed.

No throughput, latency or memory figure is claimed anywhere in this change.
gate_fixed.out carries 264.4s to load and 327-343s per 32-token prompt. That is
a correctness run whose lm_head and 46 FP8 mamba projections still execute
host-side. Both public rows say "NOT a benchmark" in those words, so nothing can
later lift them into a grid. One config caveat travels with the run:
--gpu-memory-utilization 0.92 did not size the KV pool, which fell back to 256
blocks (#83).

Anchors were re-derived, not carried

The branch's own cited anchor for the MTP deferral, nemotron_h_weights.cpp:989,
now points at a config comment; the deferral is at :1123. Every anchor in the
new cells was re-read against 10fe7f475, and
scripts/check-symbol-anchors.py reports in-repo checked 104 (fresh 104, stale 0), buckets sum 630 vs 630 citations, floor 85.

.agents/issue-index.md

GitHub does not run the merge=union driver .gitattributes:7 sets, so the
union auto-merge was discarded, main's file taken whole, and only the rows
whose KEY main lacks re-appended: #1074 and #1080. main already carries a
#1068 row, so that one is not re-appended. Asserted: main's blob is a
strict BYTE prefix of the result (346246 of 349242 bytes), 369 rows, zero
duplicate keys. This was redone from scratch on each of the three main
merges this branch needed, because a union interleave destroys the prefix
property that makes the file an append-only log.

#1080 is re-measured rather than carried: after this change repairs
nemotron-h-model.md's ## 7. Now to ## Now, 15 specs still write the
numbered spelling, up from the twelve #1080 was filed against. A growing
population is the argument for the checker-semantics fix over a rename sweep, and
both need their own spec and a red-before test. It is listed under ## Owed in
nemotron-h-model.md.

#1217 is deliberately not appended. At 10fe7f475 it is recorded nowhere
in this tree, and both writes ride in the open #1221; appending it here would
give two branches the same append-only key and the union driver would merge them
into a DUPLICATE. It is named under ## Owed in nemotron-h-model.md instead,
where the claim is true whatever #1221 does.

Gates, run bare on this head

agent-preflight.sh --staged 0 (every suite ok, including
test_cpu_x86_llamacpp_floor, which did not hit #618 on this run).
check-agent-record.py 0 (MODEL=377), check-model-checklist.py 0,
check-doc-checkpoint.py --base 10fe7f475 --head HEAD 0,
check-public-doc-tables.py 0, check-supported-models.py 0,
check-issue-index-append-only.py 0, check-symbol-anchors.py 0,
check-now-current.py 0, check-oracle-pins.py 0,
check-commit-style.py --range 0, check-commit-trailers.py --range 0,
check-pr-size.py 0.

Worth naming again: check-doc-checkpoint.py's STATES tuple (:56-65) still
contains neither INVENTORIED nor PARTIAL, so it is structurally blind to this
move and demanded none of the projections. docs/STATUS.md, docs/BENCHMARKS.md
and the spec's ## Now are written because AGENTS.md owes them, not because a
gate asked. The docs/STATUS.md edit first tripped the oversized_cells ratchet
at 45 over 44 and was shortened to fit rather than the ratchet raised.

Inherited, not introduced

windows-msvc-cpu and windows-msvc-vulkan fail on every pull request
(#584,
#968).
test_cpu_x86_llamacpp_floor exits NO_QUIET_WINDOW under box load
(#618).

Closes #1074

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