The C ABI has no multimodal request path for any model, and it needs a row: two models serve images and audio through the server, neither is reachable from include/vllm.h

Row: -

include/vllm.h has no multimodal request path, for any model. This is the C ABI's one
named open capability gap, it is tree-wide rather than any model row's, and it needs a row
of its own.

Measured

The header exposes three multimodal engine-CONFIGURATION fields — language_model_only
(:649), limit_mm_per_prompt (:669), mmproj_path (:683) — and zero multimodal
request fields. No vllm_mm_* symbol exists.

The header says so itself, twice, in contract language:

• :204-216 — "THIS ABI HAS NO MULTIMODAL CHAT REQUEST PATH YET ... the part is dropped,
  its text siblings still form the prompt"
• :288-293 — "THIS ABI STILL HAS NO MULTIMODAL REQUEST PATH"

No model sets a precedent, Qwen3-VL included. Eight architectures declare
supports_multimodal = true; only two carry encode_mm/embed_mm and register a chat seam
— Qwen3-VL (qwen3_vl_registry.cpp:55, :505-506) and dots3-note
(dots3_note_registry.cpp:97). Both reach their tower only through
ApiServer::handle_chat_completions. The seam interface itself
(REGISTER_VLLM_MM_CHAT, include/vllm/entrypoints/openai/mm_chat_registry.h:143-151) is an
in-tree C++ seam and is not shipped: CMakeLists.txt:2969 installs include/vllm.h alone,
and :2893-2894 states a consumer gets that header only.

Media DOES cross the ABI — but into separate handles, never into a language model:
transcription PCM (:1018-1021), video-generation PPM/WAV paths (:1160-1164), speech
reference audio (:1299-1301).

The docs are honest about it, which is why this is a gap rather than a defect:
docs/FEATURES.md:406 reads | Multimodal input (image/audio/video) | none | embedder-unreachable |, docs/guides/multimodal-input.md:83-85 says an image_url sent
through the C ABI is dropped, and tests/capi/test_capi.cpp:1457-1493 pins that behaviour.

Why this is not a model row's job

AGENTS.md §"Shared seams" requires every shipped capability to be reachable through
include/vllm.h, with examples and servers as thin clients. Two models now serve images and
audio through the server and neither is reachable from the ABI.

A first multimodal request path must serve Qwen3-VL as well as dots3-note, or it is a
per-model ABI — which is the thing the seam registry (#2481) was built to prevent. It also
owns scripts/abi-capability-allowlist.txt and scripts/check-surface-coverage.py, and it
changes docs/FEATURES.md. None of that belongs to a model port.

Two record repairs it should carry

• scripts/abi-capability-allowlist.txt cites vllm.h:440,451 for
  vllm_chat/vllm_chat_stream; they are now include/vllm.h:988 and :999. Its reason
  text ("even the server's Qwen3-VL image seam does not run the vision tower in the live
  engine step") is falsified by tests/vllm/entrypoints/openai/test_api_server_mm_forward.cpp.
  Its VERDICT (ABI-unreachable) remains correct; only the anchors and the reason are stale.
• include/vllm.h:290 names vllm_generate in a permanent contract paragraph. No such
  symbol exists — the completion entry point is vllm_complete (:866).

Unrelated, found while reading, recorded so it is not lost

examples/tokenize/main.cpp:20-21 includes internal headers, violating AGENTS.md's rule
that examples are thin clients of include/vllm.h. Not this row's work, but nothing else
tracks it.

Start with the multimodal contract sections in include/vllm.h, the Qwen3-VL and dots3-note registry seams, and include/vllm/entrypoints/openai/mm_chat_registry.h. Read tests/vllm/entrypoints/openai/test_api_server_mm_forward.cpp and tests/capi/test_capi.cpp to understand current server behavior and ABI behavior. Done means both models have an ABI request path, the allowlist anchors and docs are updated, and the existing coverage reflects the new contract.