Model matrix is short BailingMoeV3ForCausalLM, a recipe architecture that exists in vLLM main

[localai-bot]

CORRECTION 2026-08-13. This issue originally claimed two rows. The
Qwen3_5MoeForCausalLM half is a DUPLICATE of #490, whose PR #601 was
already open when this was filed — and #601 does not merely row it, it
registers both Qwen3_5MoeForCausalLM and Qwen3_5ForCausalLM, resolves the
backbone prefix, and carries CPU 404/404 plus SACRED GPU evidence, reviewed
twice. #601 owns that architecture; two branches adding the same keyed row
would merge cleanly and define it twice.

This issue is now scoped to BailingMoeV3ForCausalLM alone — one row.
The filing error was mine: I checked existing issues with search terms that
missed #490 and did not check open PRs at all, which AGENTS.md requires.

What

Two architectures used by official recipes have no row in
.agents/model-matrix.md. Both exist in vLLM main and are absent only from
our pin 5559679229bc961848b121ccdeaa8fa5d79bec98 — this is pin-lag, not an
inventory defect, but the matrix claims exhaustive coverage and today it is short
by these two.

Architecture	Recipe	At pin	vLLM main
Qwen3_5MoeForCausalLM	Qwen/Qwen3.8-2.4T-A95B	absent	present
BailingMoeV3ForCausalLM	inclusionAI/Ling-3.0-flash	absent	present

Qwen3_5MoeForCausalLM is the higher-value one: it is a new registry entry for the
Qwen3.5-MoE family we already ship token-exact and gated as
Qwen3_5MoeForConditionalGeneration. Whether it is a pure alias or a genuinely
separate class is the first thing the row has to establish — read it in
vllm/model_executor/models/registry.py on main and diff the target module
against what we already implement. If it is an alias, the row closes for the cost
of an alias entry.

BailingMoeV3ForCausalLM succeeds BailingMoeV2_5ForCausalLM, which does have
a row (used by Ling-2.6-1T, Ling-2.6-flash, Ring-2.6-1T), so the delta is
V2.5 → V3 rather than a new family.

Scope

Add both rows to .agents/model-matrix.md with upstream anchors and INVENTORIED
state, and update the rollup counts in the same change (the checklist is
CI-enforced by scripts/check-model-checklist.py). Records-only — no engine
change. Whether either can advance past INVENTORIED before the pin advances is
noted on the row, not assumed.

Found while auditing recipes.vllm.ai coverage.