Clarify and possibly require MLX for quality workflows

[swernerx]

Goal

Decide and implement a clearer MLX product stance: make MLX the expected quality runtime for model-backed Morphēa workflows, while keeping any non-MLX fallback explicitly scoped to portability tests, debugging, or minimal non-model usage.

Why this matters

The code currently treats MLX as optional in packaging and README language, but the high-quality model paths already fail by default when MLX is unavailable. Fallback artifacts only appear when allow_unavailable is explicitly set.

That split creates product ambiguity:

• users may think MLX is optional for equivalent quality;
• platform differences may produce different model artifacts or reports;
• docs understate that the own-model story is strongest when MLX is present;
• a Mac-first audience may prefer one recommended runtime over broad-but-variable behavior.

Current state

• morphea train-mlx fails by default without MLX; allow_unavailable writes a centroid fallback artifact only when explicitly requested.
• morphea train-raster-targets also fails by default without MLX unless allow_unavailable is set.
• morphea retrain --backend mlx and self-learn --backend mlx already route through the MLX primitive classifier.
• README and docs still describe MLX as optional.
• self-learn and retrain default to centroid unless configured otherwise.
• MLX/SAM segmentation is a separate concern because it also depends on model weights and the mlx-sam runtime.

Proposed direction

Adopt this product stance:

Morphēa can run basic deterministic geometry without MLX, but model-backed quality workflows require MLX. Non-MLX fallbacks are diagnostic, not equivalent quality paths.

Work items

• Update README and setup docs to present MLX as the recommended/default runtime for serious Morphēa development on Apple Silicon Macs.
• Decide whether to move mlx>=0.31 from the optional mlx extra into core dependencies, or introduce an explicit quality/dev install profile that all model-backed docs use.
• Change self-learning/retraining docs and checked-in configs to default to backend: mlx where quality evidence is expected.
• Consider changing CLI defaults for retrain and self-learn from centroid to mlx, with a clearly named --backend centroid escape hatch.
• Rename or de-emphasize allow_unavailable so it reads as a diagnostic/testing escape hatch rather than a supported quality mode.
• Keep MLX/SAM documented separately as an experimental candidate-region source, not a required default, until checkpoint/runtime setup is simpler and promotion evidence proves a consistent quality gain.
• Add a status/doc section explaining platform expectations: Apple Silicon macOS is the primary target; other platforms may run non-model or reduced workflows unless MLX support is confirmed.

Acceptance criteria

• A new reader can tell from the README which install path gives the intended model-backed quality behavior.
• Model-backed smoke configs do not silently pass through non-MLX fallback artifacts.
• CI/test docs distinguish required quality tests from portability/fallback tests.
• morphea status or docs clearly reports when the environment is below the intended quality runtime.
• MLX/SAM remains opt-in until it has the same repeatability and quality evidence as the classifier paths.

References

• MLX official docs describe MLX as an Apple Silicon-focused machine learning array framework and list macOS / Apple Silicon installation requirements.
• Current Morphēa MLX training paths already fail by default without MLX unless allow_unavailable is explicitly enabled.