Three of Qwen3.5's four new expert-stream step guards sit in parity-only entry points that no production path calls

[localai-bot]

Split out of #1106 finding 2 so the debt stays open after that repair closes.

#1091 finding 3 added a Qwen35ExpertStreamStep guard to four MoE entry points that reach ExpertMlpKq -> KqExpertSlice and marked no step. The guards are correct. Exactly one of the four is reachable from a production entry point:

Entry point	Production caller
Qwen3_5MTPModel::ForwardPaged	YES — src/vllm/v1/worker/gpu/runner.cpp:2183 -> src/vllm/v1/worker/gpu/spec_decode/mtp/speculator.cpp:107,262
Qwen3_5MTPModel::Forward	no — only tests/vllm/v1/spec_decode/test_mtp_speculator.cpp, and only through ForwardLogitsHost, which qwen3_5_mtp.h:135 documents as "standalone parity convenience" and which is itself called from nothing outside tests/
Qwen3_5Model::ForwardDense	no — qwen3_5.h:234 calls it the parity reference; callers are tests/parity/test_op_parity.cpp:1107, tests/vllm/v1/worker/test_runner.cpp:1278, tests/vllm/models/test_qwen35_paged_forward.cpp:293,320,385,403
Qwen3_5ReplayLayer	no — qwen3_5.h:322 calls it per-layer parity replay; only caller is tests/parity/test_op_parity.cpp:1050

Per .agents/reachability.md a call site inside a test is not reach, so "every added path is reached from a production entry point at this commit" was true of one guard in four.

Nothing needs deleting. The three guards are right where they are and cost nothing, and they become live the moment any of those entry points gains a production caller — the alternative, adding the guard later along with the caller, is how this row acquired the missing boundary in the first place. What is owed is the record: this is a staged slice that lands unreached, and it is now named as one rather than claimed as reached.

Two ways it closes. Either one of these three entry points gains a production caller, at which point its guard is exercised by whatever drives it; or the entry points are retired as parity references and the guards go with them. Neither is scheduled, and neither should be forced by this issue.

Row ENG-EXPERT-STREAM, listed under ## Owed in expert-streaming.md.

[localai-bot]

The fourth guard is reached by a production caller, but not on the default configuration

Raised by the fresh review of #1100 at 4ada1fb8d and recorded here rather than expanding that pull request, because this issue already owns the debt.

This issue's premise is that one guard of the four is reachable and three are not. That is accurate as far as it goes, and the reachability trace behind it holds: Qwen3_5MTPModel::ForwardPaged is called from src/vllm/v1/worker/gpu/spec_decode/mtp/speculator.cpp:107,262, reached from src/vllm/v1/worker/gpu/runner.cpp:2183.

The residual is what that one caller is gated on. GPUModelRunner::propose_drafts, which contains runner.cpp:2183, carries its own header note:

// Stashes the resulting per-request drafts in pending_drafts_ for the engine
// core's out-of-band pull. UNREACHABLE unless a speculator is configured.
— src/vllm/v1/worker/gpu/runner.cpp:2120

.agents/reachability.md defines a production entry point as include/vllm.h, the loader, ModelRegistry::Forward, or "a registered server or command-line path on its default configuration". So two statements that sound alike come apart here:

• "ForwardPaged's guard has a production caller" — TRUE, and it is what fix(ENG-EXPERT-STREAM): the liveness line could not print the zero the docs told an operator to read, and then its comments outran the code twice (#912, #1091, #1106, #1108) #1100 claims.
• "a default-configuration run reaches any of the four guards" — FALSE. Spec decode is off unless a speculator is configured, so on a default run the count reached is zero of four, not one of four.

This is not the "flag with no default path" shape from that guide's table, and no guard should be deleted over it. A speculator is a supported, documented, user-selected configuration, not a branch nothing turns on; the guard is correct where it sits, and it runs whenever spec decode does. The point is narrower and is about the record: the difference between "one of four is reached" and "one of four is reached, and only under a non-default configuration" is exactly the size of gap this row has now been reviewed for four times running, and it should be written down before it has to be found again.

What has already been done, in #1100 at 3ef9d023c, is to carry the qualifier in the three sentences that were being rewritten anyway — leaving a knowingly imprecise clause standing in the middle of an audit for imprecise clauses would have been the same failure:

• src/vllm/model_executor/models/qwen3_5.cpp — the ForwardLayers note now says that caller is "UNREACHABLE unless a speculator is configured (runner.cpp:2120), so a DEFAULT-configuration run reaches none of these four guards; one of them has a production caller, which is not the same claim".
• tests/vllm/model_executor/test_expert_stream_steps.cpp — the same qualifier at the head of the step-clock cases.
• .agents/specs/expert-streaming.md — the ## Owed row for this issue now records it rather than rounding it up.

What stays owed here is unchanged and is still what closes this issue: one of those entry points gains a production caller, or they are retired as parity references. Neither is scheduled. This comment adds no new work; it makes the existing claim precise so the next reader does not have to re-derive it from runner.cpp:2120.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5-1m [claude-code]