Fix Lenaig2026 data_path layout + Schrag2026Pediatric zero-annotation run crash

[bruAristimunha]

Problem

Both bugs were found re-running the Voyager convert pipeline on real downloads (not caught by synthetic tests, since neither reproduces the real archive layout / real-subject trial-CSV drift). See .pr-audits/w41/V6B-REPORT.md for the original evidence.

1. Lenaig2026

data_path() hard-coded an EEG_24Chan_AudioStim/ wrapper directory for the extracted RAR. The current Zenodo v2 archive (record 21156618) extracts EXP1//EXP2/ directly at the extraction root, so every subject's data_path() raised FileNotFoundError: Some data files are missing. before ever reading a file.

Fix: locate each run's file with a recursive glob (base_path.glob(f"**/{rel}")) that matches both the wrapped and flat layouts, instead of a single hard-coded relative path. Events/labels are unchanged.

2. Schrag2026Pediatric

_load_game_run's documented policy drops all trial labels when a game run's Trial Started marker count drifts more than 10% from its movements-CSV row count. Subject 1's personalized-stimulus game run (task-T3_acq-C3S1_M2) has 15% drift, so that run ends up with zero annotations. _get_single_subject_data returned it anyway, and bids_interface._write_file hard-requires every Raw to carry annotations, so the whole convert crashed with ValueError: Raw object must have annotations to be saved in BIDS format.

Fix: _get_single_subject_data now skips a run that ends up with zero events instead of returning it, logging a warning that names the subject and run. The class and _load_game_run docstrings document this. The high-drift policy itself (why labels are dropped) is unchanged — this only changes what happens to a run with zero resulting labels.

Tests

Two new regression tests in moabb/tests/test_dataset_fixes.py:

• test_lenaig2026_data_path_accepts_wrapped_and_flat_layouts: a synthetic tmp tree in each layout, asserting data_path() finds the right files either way.
• test_schrag2026_skips_run_with_zero_annotations_after_high_drift: mocks _load_game_run to return one zero-annotation and one labelled run, asserts the zero-annotation run is skipped and a warning names the subject/run.

Both pass locally (pytest -k "lenaig2026 or schrag2026"), plus the rest of test_dataset_fixes.py (15/15 passed).

Not in scope

Does not touch the Lenaig2026 CC-BY-NC-ND-4.0 licence question (held for a separate confirmation) or attempt any upload/publish.

[bruAristimunha]

Pushed a third fix (52a5e6e): Lenaig2026.METADATA.experiment.trials_per_class was a bare int (10), not the Dict[str, int] the schema declares. Found during the real Voyager end-to-end convert this PR enables (the layout fix got past data_path(), then convert_to_bids() crashed in bids_interface._build_readme with AttributeError: 'int' object has no attribute 'items'). Fixed to {"Stimulus": 10, "Silence": 10}, matching the documented 10 repetitions per condition. Regression test + whats_new entry added.

[bruAristimunha]

Pushed a fourth fix (5e98acd): convert_to_bids's per-subject loop had no try/except around self.get_data(). Found on the real Voyager convert of Schrag2026Pediatric: subject 16's only game recording has a single run, and that run is dropped entirely by the >10% drift policy (same mechanism as subject 1, fixed earlier in this PR), so _get_single_subject_data correctly raises FileNotFoundError for that one subject -- but with no per-subject exception handling, that aborted conversion of every subject after 16 too (would have silently lost subjects 17-47). Now caught, logged with a warning naming the subject, and the loop continues. Regression test + whats_new entry added; full re-convert of both datasets on Voyager now running end-to-end.