mlx backend silently drops weight_decay and seed despite the ignored-with-a-log-line convention

[thedataengineer]

The mlx backend is the only one that implements the "ignored with a log line" convention (backends/mlx.py:349-358), but its ignored-fields list is incomplete and two fields are dropped silently:

1. weight_decay is never applied. The optimizer is plain optim.Adam(learning_rate=...) at mlx.py:347; weight_decay has zero occurrences in the file. Since TrainConfig.weight_decay defaults to 0.01 (training.py:48), every mlx run silently discards a value the user can reasonably believe is active. Either switch to AdamW or add it to the ignored list.

2. seed is honored only on the more_plus path (mlx.py:700). The standard SFT/DPO/GRPO paths never seed, so mlx runs are non-reproducible and the field is dropped with no log line.

3. The optim check at mlx.py:351 compares against != "adamw_8bit", but that IS the dataclass default (training.py:51), so a user who explicitly requests the default gets no note even though mlx uses plain Adam.

4. The adapter-kind catch-all at mlx.py:292 ("mlx backend has no attach path for adapter kind {spec.adapter!r}") is fail-closed but not actionable, unlike the soft-prompt error at mlx.py:249-253 which correctly points at backend='torch'. Any custom adapter kind registered via methods.register(...) hits the bare message.

Related, different file: training.py:74 more_plus_tau is accepted and persisted but the per-token gating it configures does not exist yet (comment says "v1.1"). Setting it is silently inert; worth a log line or a docstring note.

[thedataengineer]

Fixed in 6a1b82c: all six mlx optimizer sites now build AdamW with config.weight_decay, finetune() seeds mx.random before dispatching to any trainer path, the ignored-fields list moved to a unit-testable _ignored_fields() whose optim check accepts any adamw variant, and the unknown-adapter-kind error names the supported kinds plus the torch fallback.