[IonDen]

mlx-dfloat could only run FLUX.1. This adds Z-Image-Turbo and Z-Image, two current image models with DFloat11 checkpoints, and the layer that lets more mflux model families follow.

The FLUX.1 integration hooks mflux's per-block methods, and Z-Image's transformer has none: it calls its blocks inline. The new layer swaps each block's class for a subclass that installs that block's weights from the compressed set just before it runs and puts placeholders back afterwards, so every parameter keeps the path mflux expects. Weight names come from mflux's own Z-Image weight mapping, including the renames outside the blocks. The one compressed group that is not a block, the caption embedder, is decoded once when the set loads. mflux compiles Z-Image's denoising step on M1/M2 Max and Ultra and on M3 or later, and evaluating after each block cannot happen inside a compiled function, so the adapter takes mflux's uncompiled path on every chip and refuses to run if mflux hands back a compiled step anyway. A model registry backs mlx-dfloat generate --model z-image-turbo / --model z-image and a new DFloatModel(name), which returns a ready model. FLUX.1 keeps its own hook and behaves as before.

Measured on an M1 Max with 32 GB (macOS 27.0.1, mlx 0.32.2, mflux 0.20.0):

• Z-Image's DFloat11 checkpoint decodes to exactly its BF16 original: all 271 compressed matrices and the 250 tensors stored uncompressed.
• Z-Image-Turbo's original is FP32, so there is no BF16 original to compare with. The GPU kernel matches the CPU reference on all 35 groups, and five sampled groups (32 of 271 matrices) equal the FP32 original rounded to BF16 (round to nearest, ties to even). Turbo's latents have not been compared with a reference yet.
• Z-Image's final latents at 1024², 4 steps, guidance 4.0 are bit-identical to the same transformer streaming its BF16 weights, through the same uncompiled path on both sides.
• At 1024², Turbo (9 steps) peaked at 13.11 GiB and Z-Image (50 steps, guidance 4.0) at 13.15 GiB, both during the VAE decode. The memory estimate is calibrated to reproduce those peaks in that phase.
• A 16 GB Mac was not measured. On the 32 GB Mac under a 16 GB Mac's limits, with the fit check turned off (it predicted 12.2 GiB and refused the run), the watchdog stopped Turbo after 6.7 s while encoding the prompt, at 9.29 GiB against a 9.17 GiB ceiling, so that size is marked at risk.

FLUX.1 is unchanged: compared with the commit this branch started from, schnell's 1024² image is pixel-identical and the watched peaks are 0.019 GiB apart, and schnell's latents are still bit-identical to the BF16-streamed run.

Smaller changes users will see: the default Z-Image checkpoints are fetched at fixed Hub revisions; generate says on stderr when a model spec is fetched from the Hub, warns when a negative prompt is ignored because guidance is 1 or less, and records elapsed_seconds in its --report; a watchdog stop records the phase that was running.

Also in this PR: scripts/verify_remote_group.py --cast-fp32-to-bf16 for checkpoints whose original is FP32, Z-Image support in scripts/verify_image.py (Turbo is refused there, since it has no BF16 original), and tier-table rows for runs the watchdog stopped, shown as lower bounds with the time they ran.