desktop(mesh-llm): let a serving node route a different model

[tlongwell-block]

Why

A user asked: why can't I use a different model than the one I'm hosting? The answer turned out to be a Sprout-side policy, not a mesh constraint.

ensure_client_node_for_model rejected any running Serve runtime with "this desktop is currently sharing compute; stop sharing before using relay mesh as a client". That made "host model A" and "use model B" mutually exclusive.

But mesh-llm's API proxy already resolves a request to a local, remote, or split target at request time (route_missing_local_model → hosts_for_model). A serving node's 9337 ingress can route a model it doesn't host locally to a peer. The serve-vs-client split was a blunt UI/runtime policy layered on top of a router that never needed it.

What

desktop/src-tauri/src/commands/mesh_llm.rs — treat a running runtime (in any mode) as the local 9337 OpenAI ingress and hand it back. If the caller selected a target, dial it first so the node joins the chosen peer's mesh.

if let Some(runtime) = runtime.as_ref() {
    if let Some(endpoint_addr) = /* selected target, trimmed */ {
        runtime.dial_endpoint_addr(endpoint_addr).await?;  // join the chosen peer
    }
    return runtime.status().await;  // agent uses 9337; router decides per request
}

Two deliberate calls:

• No /v1/models preflight probe. It races model gossip and would wrongly reject freshly-discovered remote/split models. Routability is left to the request path, which already returns a real error. (/api/model-targets exists in mesh-llm but isn't surfaced in Sprout's SDK and still isn't a per-request guarantee.)
• Preserve the dial side effect. A first cut that just deleted the guard dropped dial_endpoint_addr — a serve runtime not yet on the target's mesh would fail its first inference while the frontend had already published call-me-now to the peer. Dialing the selected target restores the join.

justfile — pass --features mesh-llm directly to tauri dev/tauri build. TAURI_CLI_EXTRA_CARGO_ARGS is not a recognized Tauri v2 env var, so the feature was silently dropped and desktop dev/staging/build ran without mesh-llm. Fixed all three recipes (dev, staging, build).

Tests

• New #[ignore] hardware regression in commands/mesh_llm.rs: a real Serve runtime + a different-model preflight returns Ok with mode: Serve and the same api_base_url — exactly the case the old guard forbade. PASS.
• Full sprout-desktop --features mesh-llm suite: 436 passed, 0 failed (the hardware test is correctly ignored by default).
• mesh_signaling: 14/14. compile + clippy clean (lib + tests).

Known gaps

1. The Some(endpoint) dial path isn't asserted to fire in a unit test — dialing a bogus endpoint is flaky, and a live second peer hit gap Initial release — Sprout Nostr relay with enterprise extensions #2. The branch is small and source-obvious; covered by the live create-agent flow.
2. The live "serve A, route B to a real peer" end-to-end isn't re-proven here. A two-process harness hit a mesh-llm downloader/readiness bug (peer's model download stalled at 0%, then a 5s startup-shutdown timeout + Metal teardown assert) — HF itself is reachable, so it's orthogonal to this change. The routing premise is proven by source + the existing mesh_serve_client_smoke. Worth a separate look (possibly relevant to split-model startup).

Reviewed twice by Max (mesh internals) before opening.