vllm-serve aborts on --enable-auto-tool-choice and --trust-remote-code, so 89 of 157 official recipe commands fail to start

[localai-bot]

What

vllm-serve hard-errors on an unrecognized argument
(src/vllm/entrypoints/openai/server_main.cpp:440). Two flags that appear in
almost every official recipe are unrecognized, so copy-pasting a published
vllm serve line aborts before the model loads.

Measured over vllm-project/recipes @ 86c7777a (157 model recipes):

Flag	Recipes using it	Our engine
--enable-auto-tool-choice	89 / 157	unrecognized → abort
--trust-remote-code	82 / 157	unrecognized → abort

Both are no-ops for us, which is what makes this cheap:

• --enable-auto-tool-choice is upstream's opt-in for automatic tool choice. We
  already behave as if it is always set whenever --tool-call-parser resolves to
  a parser, so accepting it changes nothing.
• --trust-remote-code authorizes executing Python from the checkpoint. We have
  no Python, so there is nothing to authorize — it is N/A by construction,
  not unimplemented.

A recipe's own example is the reproducer:

# recipes/models/Qwen/Qwen3.5-27B.yaml, features.tool_calling
vllm serve Qwen/Qwen3.5-27B-FP8 --enable-auto-tool-choice --tool-call-parser qwen3_coder
#                               ^ server: unknown argument '--enable-auto-tool-choice'

Why not just ignore every unknown flag

We should not. Rejecting unknown arguments is correct — silently ignoring
--tensor-parallel-size would let a user believe they got TP. The fix is an
explicit accepted-and-inert list with a one-line reason per entry, plus a
startup notice naming what was accepted and why it does nothing. Anything genuinely
unimplemented keeps failing loudly.

Scope

Add the accepted-and-inert seam to server_main.cpp with per-flag justification,
a notice on use, and docs/USAGE.md entries stating explicitly that these are
accepted for recipe compatibility and have no effect. RED-first: a test that
asserts each listed flag starts the server and each unlisted unknown flag still
aborts.

Needs a row — no engine-matrix row covers serve CLI recipe compatibility today.
Found while auditing recipes.vllm.ai coverage.