providers: --lex-gpu, wired for when its tools land

[alpibrupa]

Draft: depends on alpibrusl/lex-llm#64 for providers.lex_gpu_model / lex_gpu_local. CI here fails until that merges.

lex-gpu serves OpenAI chat completions from its own compiled Metal and CUDA kernels rather than llama.cpp or MLX. lex-llm drives that shape unchanged, so this is the ordinary nine-module wiring every other backend gets: a lex_gpu_agent() beside vllm_agent() in each agent, an arm in session.lex, a --lex-gpu flag in main.lex, and a README section.

It is chat-only today, and the README says so

lex-gpu accepts a tools list and silently drops it, so the model is never told the tools exist and its replies carry no tool_calls. The loop therefore never sees finish_reason: "tool_calls" and never dispatches — for a coding agent that means it will describe the work rather than do it. Asked to use an add tool, the model says as much itself:

"I don't see any add tool available in my environment... I should just answer the question directly since no such tool exists in my available tools."

So this is wiring landed ahead of the capability, on purpose. It starts working the moment lex-gpu does two things, neither of which needs a change on this side, because the wire shape does not move:

1. render tools into the prompt in the form the chat template expects;
2. split the <think> block out of content into reasoning_content.

It is a draft for that reason as much as the dependency — merge it when you want the flag present, or leave it until lex-gpu can actually hold a tool. It is placed in the README's "implemented in code" list, not the "run end-to-end against this repo" table, which is where it honestly belongs.

Checks

Run against the lex-llm branch locally (path override, reverted before commit):

• all nine agent modules, src/server/session.lex and src/tui/main.lex type-check with no errors
• lex fmt --check clean across the fourteen touched files
• lex doc-sync --check: all targets current

🤖 Generated with Claude Code

[alpibrupa]

CI failure here is exactly the dependency and nothing else: all 36 errors are unknown_identifier: providers...lex_gpu_model, from lex-code resolving lex-llm at main. Nothing else in the run fails. It goes green once alpibrusl/lex-llm#64 merges — that PR's own CI is passing.