Fix saved model path collisions across paradigms

[lindicaphxag-tech]

Summary

Fixes #1182 by namespacing saved evaluation artifacts with the paradigm and optional run suffix. Two benchmark runs that share an evaluation type, dataset, subject, session, and pipeline no longer overwrite one another when their paradigm or suffix differs.

The layout is now:

<Models|GridSearch>_<evaluation>/<Paradigm>/<suffix-if-set>/<dataset>/<subject>/<session>/<pipeline>/

The model path change is documented as an API compatibility change. Calls to the private path helper that omit paradigm retain their previous layout.

Validation

• git diff --check
• python -m py_compile moabb/evaluations/base.py moabb/evaluations/utils.py moabb/tests/test_evaluations.py
• uv run --active --no-sync python -m pytest moabb/tests/test_evaluations.py -q -k 'create_save_path or test_within_session_evaluation_save_model' — 11 passed.
• Added path tests covering paradigm and suffix isolation, plus an integration test that runs a fake EEG evaluation and checks the paradigm-scoped output directory.

[lindicaphxag-tech]

@bruAristimunha this is ready for maintainer review when convenient. It fixes #1182 by namespacing saved evaluation artifacts by paradigm and optional suffix and targets develop. Exact-head validation on 6e9a24f0cb3d8c5f748ce11f1e4146af82fe9a99 is green: touched-file pre-commit passes and all 11 focused path/integration regressions pass, including the fake EEG evaluation save path. Evidence: https://github.com/lindicaphxag-tech/lindicaphxag-tech/actions/runs/37136756045. I’ll hold the current head rather than make further scope changes.

[bruAristimunha]

Thanks @lindicaphxag-tech — the collision in #1182 is real and namespacing by paradigm/suffix is the natural fix. The code reads fine and the private helper keeps its old layout when paradigm is omitted. Because this changes the on-disk layout of saved models for everyone, I want to think about the migration story (existing Models_* trees become orphaned) before approving; I've approved the Actions run so CI can speak in the meantime. Merge-order note for the series: this one goes first, #1211 last.

[lindicaphxag-tech]

Migration follow-up on current head 38d4be23: I don't think an automatic move of existing Models_* / GridSearch_* trees can be made safe, because the legacy path is exactly missing the two provenance fields needed to choose a destination (paradigm and suffix). Any automatic migration would have to guess which namespace an existing pickle belongs to.

I tightened the migration story instead:

• legacy trees are left untouched at their historical paths; this PR only changes the destination for new saves;
• the API-change note now says explicitly that legacy artifacts are not auto-migrated and remain manually readable in place;
• the official noplot_load_model.py example now points to the canonical output produced by plot_benchmark.py: Models_WithinSession/LeftRightImagery/benchmark/Zhou2016/..., and notes that older trees remain at their original path.

This keeps the collision fix non-destructive and avoids assigning false provenance to old artifacts. If you prefer a warning when a legacy tree is detected during a new save, I can add that, but I would avoid moving/copying it automatically.

[lindicaphxag-tech]

Rebased onto current develop (ae47249, including #1210). I resolved the changelog overlap by retaining both API notes. The migration note now explicitly states that existing Models_* and GridSearch_* trees remain at their historical paths, are not auto-migrated because their paradigm/suffix provenance is unavailable, and remain manually readable there.

The current head is 845939818aea97df21843c7d96bb6e90fa8232a6. Pre-commit CI, Python syntax compilation, and git diff --check pass. The full GitHub test matrix is running; this Windows environment cannot execute pytest because the test dependencies (starting with nemar) are not installed.

[lindicaphxag-tech]

Added two regression checks on current head c471dd8: the fake WithinSession evaluation now verifies that a model artifact is actually written beneath the paradigm/suffix namespace, and the path test covers the GridSearch namespace as well. pre-commit.ci, link-check, Python syntax compilation, and git diff --check pass; this environment still lacks pytest, and the fresh upstream matrix is running.

[bruAristimunha]

Thanks @lindicaphxag-tech — the migration story on this head is the right call: legacy Models_*/GridSearch_* trees stay where they are (no false-provenance auto-move), the API-change note spells it out, and noplot_load_model.py points at the new canonical path. The two new regressions (paradigm-scoped save on the fake WithinSession eval, GridSearch namespace isolation) make the contract load-bearing. CI green on c471dd8. Approving and merging.