Track stable mflux release for MLXBits Ideogram q8 loader

[haxlys]

Context

The MLX migration currently depends on mflux PR #445 support for mlx-forge converted Ideogram 4 checkpoints, including MLXBits/ideogram-4-mlx-q8.

Current state checked on 2026-06-18:

• Upstream PR: feat: load ideogram 4 from mlx-forge converted checkpoints (bf16 + int8) filipstrand/mflux#445
• PR state: open
• Latest mflux release observed: v.0.18.0
• Current project pin: mflux @ git+https://github.com/plz12345/mflux.git@8d80b9cb53688b62a2f814604b9f8b48987c5acd

Why this exists

The MLXBits q8 model requires split_model.json / mlx-forge loader support. Stock stable mflux currently supports Ideogram 4 FP8 but not this q8 layout. We should keep the commit pin until the loader lands in an upstream stable release.

Exit criteria

• mflux PR #445, or equivalent support, is merged upstream.
• A stable mflux release includes MLXBits / mlx-forge Ideogram 4 q8 loading.
• server/requirements.txt is changed from the git commit pin to the stable release pin.
• README.md, AGENTS.md, and PR/installation notes are updated to remove the temporary PR-pin warning.
• Run verification:
  • python -m compileall server ideogram4_mlx.py scripts/doctor.py
  • rg "torch|safetensors.torch|from ideogram4|import ideogram4" server ideogram4_mlx.py
  • cd webui && pnpm lint && pnpm build
  • 256x256 V4_TURBO_12 smoke generation with MLXBits/ideogram-4-mlx-q8
  • LoRA apply -> generate -> remove -> generate smoke if local LoRA files are available