LocalLLMClientMLX: missing 'Tokenizers' module after mlx-swift-lm decoupling

[gsdali]

Summary

LocalLLMClientMLX fails to build against current mlx-swift-lm main because
its source imports Tokenizers (from swift-transformers), but
mlx-swift-lm decoupled from swift-transformers in
ml-explore/mlx-swift-lm#118
on 2026-04-01. The Tokenizers module is no longer transitively available.

Repro

Consume LocalLLMClient main + LocalLLMClientMLX from a Swift 6 package:

.package(url: "https://github.com/tattn/LocalLLMClient.git", revision: "<main HEAD>"),

.target(
    name: "MyTarget",
    dependencies: [
        .product(name: "LocalLLMClient", package: "LocalLLMClient"),
        .product(name: "LocalLLMClientMLX", package: "LocalLLMClient"),
    ]
),

swift build →

error: no such module 'Tokenizers'
.build/checkouts/LocalLLMClient/Sources/LocalLLMClientMLX/Context.swift:7:8
  5 | import MLXLLM
  6 | import MLXLMCommon
  7 | import Tokenizers
    |        `- error: no such module 'Tokenizers'

Same error in LocalLLMClientMLX/MLXClient.swift and
LocalLLMClientMLX/LLMSession+MLX.swift.

Fix

Add swift-transformers as a direct package dependency for LocalLLMClientMLX:

// Package.swift
.package(url: "https://github.com/huggingface/swift-transformers.git", from: "1.3.0"),

// LocalLLMClientMLX target dependencies:
.product(name: "Tokenizers", package: "swift-transformers"),

swift-transformers 1.3.0 (2026-03-23) provides the Tokenizers library
target directly.

Context

Filed while wiring LocalLLMClient into OCCTSwiftPartsAgent (in-process
on-device parser path). Shipped 0.4.6 still works against older
mlx-swift-lm tags, but adopters who pin LocalLLMClient by branch/revision
(forced by the SPM unstable-version rule, since 0.4.6 declares mlx-swift-lm
via branch: "main") get current mlx-swift-lm and hit this break.

Pairs with the Llama backend C++ stdlib issue (filed separately) — together
they leave LocalLLMClient unable to deliver any local-inference backend on
current main.