Record the MLX convergence-stop divergence

[neuromechanist]

Documentation half of #248 (the implementation stays open).

What

docs/guides/amica-differences.md's backend table covered Newton, PDF families,
component sharing, outlier rejection, precision and rank detection — but not the
convergence stops. MLX implements neither:

grep target	hits in pamica/mlx_impl/core.py
min_dll, use_min_dll, maxincs	0
min_nd, use_grad_norm, ndtmpsum	0

Its only stop_reason values are max_iter (:570), lrate_floor (:659) and the
degenerate ones (:590, :634). An MLX fit always runs the full iteration
budget, while the other backends stop on min_dll (on by default at 1e-9) at
iteration 326-1076 on the bundled sample.

Why this one needs saying

The doc previously closed with "MLX limitations raise NotImplementedError
rather than differing silently." Verified by probing the constructor — that is
true for what it covers:

passed to AMICAMLXNG(...)	result
do_newton=True	NotImplementedError
min_dll, min_nd, use_grad_norm, keep_best, do_reject	TypeError (no such parameter)

But a stop that does not exist has no parameter to reject, so nothing signals
its absence. That is the gap the section now names, with the practical
consequence: pick max_iter for MLX from where the other backends actually
converge on your data, not from their default budget.

It also scopes a claim in validation.md — "min_nd is unreachable in all
three implementations" means Fortran, PyTorch and NumPy. MLX never computes the
quantity, so it cannot be evaluated on that axis.

Also corrected

The keep_best / MIR / persistence rows were added and verified per backend
rather than assumed — NumPy has no keep_best, no mir, and no state_dict
(it persists through the EEGLAB amicaout format instead), so an earlier draft
calling it "partial" would have been wrong.

Docs-only; no code paths touched.

[neuromechanist]

Both Critical findings confirmed and fixed in f2ca8d7. They had a single root cause, correctly identified: this section was drafted against #244's validation.md rewrite, which is not on dev.

Critical 1 — cross-reference to content that does not exist on the target branch. Correct. The phrase "in all three implementations" lives only on 218-document-min-nd. Rather than sequence the two PRs, I made this section self-contained: it no longer asserts what another page says, so merge order no longer matters. Verified there is now no reference to validation.md, "three implementations", or the 326-1076 figure in this file.

Critical 2 — the 326-1076 framing implied MLX is uniquely budget-bound. Correct, and worth stating plainly: this is the same error #244 exists to fix, reintroduced one file over. PyTorch reaches its own max_iter=100 default long before min_dll can fire, so "MLX always runs the full budget" was not a contrast at all. Rewritten around the distinction that actually holds — what raising max_iter buys you: on the other backends a larger budget lets the likelihood stop end the fit once it stops improving; on MLX a larger budget is simply spent. No measurement is cited now, so nothing here depends on an unmerged source.

Important 4 — min_nd vs min_grad_norm. Added in both the table row and the prose, so a numpy_impl user grepping for min_nd is not left concluding the feature is absent rather than renamed.

Thanks for checking the per-backend cells independently rather than taking the table's word — the keep_best/MIR/persistence rows were ones I'd verified after an earlier draft got NumPy wrong, and it is useful to have that confirmed from a second read.