Construct orcarouter Flash-Next Uncensored-MLX via qwen4_exp shim

[okwithit9-debug]

Motivation

The required Flash-Next line is orcarouter only:

• EXO MLX: orcarouter/Qwen3.8-Flash-Next-Uncensored-MLX
• Spark vLLM companion (same family, not an EXO MLX card): orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4

Do not promote mlx-community generic Flash-Next, lychee888, or other quants as the default. Uncensored is a weight pack (abliteration), not a new training pipeline. EXO needs a working qwen4_exp / Qwen4Exp* construct path so orcarouter Uncensored-MLX can load without waiting solely on mlx-lm#1788.

Changes

• First-class catalog card: orcarouter/Qwen3.8-Flash-Next-Uncensored-MLX only (repo-root 4-bit MLX). No mlx-community Flash-Next card. No NVFP4 EXO MLX card.
• Allowlist Qwen4ExpForConditionalGeneration / Qwen4ExpForCausalLM.
• Temporary mlx-vlm>=0.6.17 shim (src/exo/worker/engines/mlx/vendor/qwen4_exp.py + qwen4_exp_shim.py) so load_mlx_lm_model can construct the pack:
  1. Native mlx_lm.models.qwen4_exp if present (after mDNS peer discovery broken when running from git source (uv sync + uv run) exo-explore/exo#1788).
  2. Else EXO adapter around mlx_vlm.models.qwen4_exp (LanguageModel + VLM sanitize, logits mx.array for EXO generate).
  3. Else Qwen4ExpUnavailableError with install + “remove when mDNS peer discovery broken when running from git source (uv sync + uv run) exo-explore/exo#1788 lands” steps.
• Qwen3.8-27B cards remain secondary (qwen3_5 path).
• README: install/shim steps, how to delete the shim after mDNS peer discovery broken when running from git source (uv sync + uv run) exo-explore/exo#1788, and an honest remaining-gap note.

Why It Works

orcarouter Uncensored-MLX is an mlx-vlm conversion (Qwen4ExpForConditionalGeneration / qwen4_exp) with no model_file. The EXO pin of mlx_lm already accepts get_model_classes; the shim supplies Model / ModelArgs that match load_model (from_dict → construct → sanitize → load_weights). Native mlx_lm wins when exo-explore#1788 ships; the vendor module is marked for deletion then.

Honest remaining gap: this unblocks construct/load, not a verified full MoE / hybrid GDN+QSA / PLE n-gram generate pass, and not typed Mac+Spark auto_parallel TP/PP. Those still need Apple-silicon (or CUDA MLX) runtime, the gated ~163 GiB pack, and Qwen4Exp layer handlers.

Test Plan

Manual Testing

• pip install 'mlx-vlm>=0.6.17', accept the HuggingFace gate, set a Hub token, download on Mac, launch orcarouter/Qwen3.8-Flash-Next-Uncensored-MLX.
• Confirm /models shows that id (not mlx-community Flash-Next). NVFP4 is Spark vLLM only.
• After construct, treat generate / disagg as still unverified until run on Metal with the real pack.

Automated Testing

• Registration tests: orcarouter card exists; mlx-community Flash-Next / NVFP4 cards do not; 27B qwen3_5 still registers.
• Load-hook / shim tests (offline, no weight download): native mlx_lm wins; mlx-vlm fallback; documented error when shim missing; install does not overwrite native; adapter construct + logits + vision-key sanitize; EOS ids.