No production path gives an EXL3 checkpoint a paged KV cache: DeepseekV4ForwardExl3Paged has only test callers

Row: MODEL-DSV4-EXL3

DeepseekV4ForwardExl3Paged is the only DeepSeek-V4 forward that takes both a
paged KV cache and a DeepseekV4CompressorState. It has NO production caller.

$ grep -rn "DeepseekV4ForwardExl3Paged" src/ examples/
src/vllm/model_executor/models/deepseek_v4.cpp:3738   (its own refusal message)
src/vllm/model_executor/models/deepseek_v4.cpp:3740   (its own refusal message)

Every other reference is in tests/vllm/models/test_deepseek_v4_exl3_loader.cpp.

The registered forward ForwardDeepseekV4ForCausalLM
(deepseek_v4_registry.cpp:102-133) has three branches, and none of them reaches
it:

• input.gather_logits -> DeepseekV4Model::ForwardDevice, which binds
  dev_be.exl3 and neither paged_kv nor compressor.
• input.multi_kv != nullptr -> DeepseekV4ForwardGgufPaged, which refuses by
  name on !weights.has_gguf_weights. An EXL3 checkpoint has none, so this
  branch is a refusal for this arm rather than a path.
• otherwise -> DeepseekV4Model::Forward -> DeepseekV4ForwardExl3, which also
  binds neither.

So no production path gives an EXL3 checkpoint a paged KV cache. V4Backend
documents what that means: "Null = stateless full-recompute (the default / --gpu
path)". The EXL3 arm therefore recomputes the entire prefix on every decode step.

WHY THIS MATTERS MORE THAN #2442. The published 44-47 tok/s target is measured at
384k context, 1 seq. A forward that recomputes 384k tokens per step is not a
decode, and no amount of expert-tower residency or speculation changes that.
#2442 (routed experts on a CPU queue) is real and now fixed, but this sits in
front of it.

It also means the DSA composition -- the compressor and indexer work that W3
landed -- is unreachable in production, since compressor is only ever non-null
on the paged entry. Its gates are real; its only callers are tests.

AGENTS.md §"Nothing lands dead": a staged slice may land unreached only when
the commit body, the PR body and the row's ## Owed all name it. This one is not
recorded anywhere as unreached.

Found while wiring #2442, by checking whether the device-staging call sites were
on the hot path.

Start with ForwardDeepseekV4ForCausalLM in src/vllm/model_executor/models/deepseek_v4_registry.cpp and trace its three branches into the DeepSeek-V4 forward functions in deepseek_v4.cpp. Compare the production paths with DeepseekV4ForwardExl3Paged and its callers in tests/vllm/models/test_deepseek_v4_exl3_loader.cpp. Done means an EXL3 checkpoint has a production path that supplies paged KV and compressor state, with tests covering that path.