record(MODEL-MATRIX): row the 8 recipe architectures that had none (#609, #610)

[localai-bot]

Records only. No src/, include/ or tests/ change.

The matrix claims exhaustive coverage. An audit of the 157 official model recipes
at vllm-project/recipes 86c7777a, joined to each model's HF config.json
architectures and then to .agents/model-matrix.md, found architecture strings
with no row at all. This adds them at INVENTORIED.

The 8 rows

Pin-lag — registered on vLLM main, absent only at 555967922 (#609):

Row ID	Architecture	Recipe
MODEL-TEXT-qwen3-5-qwen3-5-moe-for-causal-lm	Qwen3_5MoeForCausalLM	Qwen/Qwen3.8-2.4T-A95B
MODEL-TEXT-bailing-moe-v3-bailing-moe-v3-for-causal-lm	BailingMoeV3ForCausalLM	inclusionAI/Ling-3.0-flash

Out-of-repo — vllm-project/vllm-omni, following the MiniMaxH3DiTModel
precedent (#610):

Row ID	Architecture	Recipes
MODEL-MM-moss-tts-moss-tts-delay-talker-for-generation	MossTTSDelayModel	MOSS-TTS, MOSS-TTSD-v1.0, MOSS-SoundEffect, MOSS-VoiceGenerator
MODEL-MM-moss-tts-moss-tts-realtime-talker-for-generation	MossTTSRealtime	MOSS-TTS-Realtime
MODEL-MM-qwen3-tts-qwen3-tts-talker-for-conditional-generation	Qwen3TTSForConditionalGeneration, Qwen3TTSTalkerForConditionalGeneration	Qwen3-TTS-12Hz-1.7B-CustomVoice
MODEL-MM-higgs-audio-v3-higgs-audio-v3-talker-for-conditional-generation	HiggsMultimodalQwen3ForConditionalGeneration, HiggsAudioV3TalkerForConditionalGeneration	higgs-audio-v3-tts-4b
MODEL-MM-voxtral-realtime-voxtral-realtime-for-conditional-generation	VoxtralRealtimeForConditionalGeneration (target-pending)	Voxtral-Mini-4B-Realtime-2602
MODEL-MM-bailing-mm-native-bailing-mm-native-for-conditional-generation	BailingMMNativeForConditionalGeneration (target-pending)	Ming-omni-tts-0.5B

Qwen3_5MoeForCausalLM: alias or separate target?

Neither a pure alias nor a new port, and the row says so plainly.

The class already exists at our pin — qwen3_5.py:381 (main :443) — and is
exactly what the multimodal class builds as its own text tower:
Qwen3_5MoeForConditionalGeneration.__init__ does
self.language_model = Qwen3_5MoeForCausalLM(...) at qwen3_5.py:643 (pin) /
:724 (main). What main added is a registry entry at registry.py:203
in _TEXT_GENERATION_MODELS, exposing that text-only backbone standalone: same
module, different class, different task category (text generation vs
multimodal), no vision tower and no MM processor. Under the matrix's alias rule —
group only on an exact (category, module, class) match — it earns its own row
rather than folding into MODEL-MM-qwen3-5-qwen3-5-moe-for-conditional-generation.

But it is a cheap close, not a port: we already ship that tower token-exact. We
bind only the string Qwen3_5MoeForConditionalGeneration
(src/vllm/model_executor/models/qwen3_5_moe.cpp:215) and nest the text fields
under a text sub-config (src/vllm/transformers_utils/hf_config.cpp:396), while
Qwen/Qwen3.8-2.4T-A95B declares Qwen3_5MoeForCausalLM with those fields at
top level. Registration plus a flat-config descent, once the pin advances.

BailingMoeV3ForCausalLM succeeds BailingMoeV2_5ForCausalLM but is a new
module, not a class rename: V2.5 (bailing_moe_linear.py, 821 lines at the pin)
pairs MLA with Bailing linear attention; V3 (bailing_moe_v3.py, 1495 lines on
main) keeps MLA and swaps the linear arm for Kimi Delta Attention, adds
per-projection causal conv1d, grouped-topk MoE via config.n_group, a clamped
SwigluStepAndMul, an fp8 quant-config hook, and refuses a checkpoint without
no_kda_lora=True. Line anchors are on the row.

Counts

Rollup INVENTORIED 314 → 322 and Total 362 → 370; check-agent-record's MODEL
row count moves with it.

The at-the-pin model inventory is deliberately unchanged — 324 rows / 373
memberships / 356 architectures / 310 targets / 261 modules. None of the eight is
at the pin, so like the MuseGlimmer, KimiK3 and MiniMaxH3DiT rows before
them they carry no pinned-registry module-colon-colon-class token and
contribute nothing to that invariant. Existing convention, not a relaxation: no
checker expectation was widened to make this pass.

The header prose distinguishes at-pin from beyond-pin, so it is reconciled the
same way: 326 at the pin plus 3 beyond becomes 326 plus 11 = 337, long tail
284 → 292, engaged 45 unchanged. One pre-existing error in that same block is
corrected while it is open: the caption read "the 49 non-INVENTORIED rows"
where the file has, and had, 48.

Two notes before closing the issues

1. Model matrix is short BailingMoeV3ForCausalLM, a recipe architecture that exists in vLLM main #609 and Six vllm-omni recipe architectures have no model-matrix row: the TTS/audio-generation modality is entirely unplaced #610 say two and seven. Their own tables enumerate two and six.
   Reproducing the audit independently (157 recipes → 157 config rows → join)
   returns exactly eight, matching the enumerations. Eight is what landed.
2. That reproduction could not read 20 of the 157 recipe configs (gated or
   404 — including Voxtral-4B-TTS-2603, GLM-TTS, IndexTTS-2.5,
   stable-audio-open-1.0). Coverage for those is unverified, not clean, and
   several vllm-omni TTS architectures visible in the omni docs still have no
   row. Worth its own audit issue.

Evidence

python3 scripts/check-model-checklist.py:

OK: architecture-support checklist matches the detailed row states.

scripts/agent-preflight.sh --staged → exit 0, all 24 record gates, 40 mutation
suites and both staged gates ok.

test_cpu_x86_llamacpp_floor was RED on the first staged run at loadavg 63 and
GREEN on the same tree at 21.9: it races the harness's quiet-window gate and
fails toward a verdict on whatever diff is in flight. Filed as #618 and added to
the intake table — not fixed here, since it changes a test's semantics.

Closes #609.
Closes #610.

[localai-bot]

Correction to the commit message, recorded here rather than force-pushed (the branch is already published and this protocol takes no force variant).

The commit body says "Two pre-existing errors in that block are corrected" and then names only one. One is correct: the caption read "the 49 non-INVENTORIED rows" where the file has, and had, 48. Everything else in that paragraph — 326 + 3 = 329, engaged 45, tail 284 — was internally consistent before this change; the 284 → 292 move is this PR's own reconciliation for the 8 new rows, not a repair of prior drift.

The PR body above states it correctly as one. Worth fixing in the squash message at merge time.

[localai-bot]

Superseded by #625 — same tree, byte-identical, clean history.

Why this one could not be repaired in place. #620's first commit (5775f744a) changed .agents/model-matrix.md (a feature_surface) without docs/FEATURES.md in that same commit. check-doc-checkpoint is per-commit by design and carries no exemption mechanism — its own source says escape hatches are what it was rebuilt to remove. I verified directly that a later commit in the range does not clear it:

$ python3 scripts/check-doc-checkpoint.py --base origin/main --head HEAD
ERROR: commit b326899a9: changed feature_surface but did not update docs/FEATURES.md.

So the only remedies were rewriting published history or clean history on a fresh branch. This protocol takes no force variant, so #625 is the fresh branch.

#625 also carries the two corrections that arrived after #620 was opened: the Qwen3_5MoeForCausalLM row is dropped (already owned by #490 / PR #601 under the identical stable ID — two branches adding the same keyed row merge with no conflict and define it twice), taking this to seven rows and counts to INVENTORIED 321 / Total 369; and the pr-size mutation evidence in tests/scripts/test_agent_record.py that #620 was red on.

Branch row/model-matrix-recipe-backfill is left in place for history and can be deleted with this PR.