[lindicaphxag-tech]

Why this smaller PR

This is the minimal opt-in follow-up requested by @bruAristimunha in #1235 for the raw-epoch-overlap concern in #280. The original large PR #1235 has now been closed as superseded. This PR deliberately contains only three files: one standalone splitter, focused tests, and the changelog.

Behavior and exact contract

• Add PurgedEpochKFold in moabb/evaluations/splitters.py. Each recording is partitioned into near-equal contiguous chronological test blocks; training epochs are purged only when their half-open integer raw epoch intervals intersect the held-out windows.
• groups is an explicitly supplied (n_epochs, 3) array of (recording_id, event_sample, epoch_n_samples); run identities can be hashable (subject, session, run) tuples. They are encoded as scalar group IDs to prevent NumPy tuple-broadcast comparison errors. Missing IDs (None or NaN, including within tuple IDs), invalid sample types, and int64 overflow fail closed.
• The overlap lookup uses sorted interval starts and prefix-maximum stops; O((N+M) log M) time, linear auxiliary memory. Reject folds without required train/test class support rather than silently balancing classes.
• No changes to BaseParadigm, evaluation classes, automatic metadata injection, preprocessing, default splitters, API exports, or upstream CI files. No dynamic-programming class balancing or float-coordinate handling.

Scientific boundary (deliberate)

This only prevents direct overlap of declared raw epoch extraction windows. It does not prove independence after continuous-recording preprocessing (such as zero-phase IIR/FIR filtering), learned transforms, or fitted scalers. This is not a complete resolution of #280, and no EEG benchmark/AUC gain is claimed. Filter-aware research remains in a separate fork-only branch, outside this upstream PR.

Exact-head verification (2026-10-26 00:24 UTC)

Current head: 0ebd0f05c3e8538b2a50286b7ab2e70f6756e0b5. A maintainer merged the latest develop into the branch; the PR diff still contains only the standalone splitter, focused tests, and changelog.

• Official MOABB Test: in progress on this exact head: https://github.com/NeuroTechX/moabb/actions/runs/40249187768 . The previous fd999cd2 head passed the full multi-OS/Python Test matrix: https://github.com/NeuroTechX/moabb/actions/runs/40218418101 . Do not treat that earlier result as current-head CI.
• Docs: Success: https://github.com/NeuroTechX/moabb/actions/runs/40249187767 .
• Braindecode integration: Success: https://github.com/NeuroTechX/moabb/actions/runs/40249187788 . The maintainer has also clarified this is not a required check.
• What's New: Success: https://github.com/NeuroTechX/moabb/actions/runs/40249187812 .
• pre-commit.ci: Success on the current head.
• Link Check: Failure: https://github.com/NeuroTechX/moabb/actions/runs/40249187776 . The log identifies only two pre-existing https://github.com/qinxwew links in docs/source/whats_new.rst (HTTP 404). The maintainer explicitly restored those upstream-baseline links and requested any unrelated link fix separately; this PR does not change them.
• Focused regression tests cover tuple-valued IDs, missing/NaN IDs, identity isolation, chronological partitioning, class-support validation, exact half-open interval boundaries, dense independent overlap oracle, integer/overflow rejection, and large-cohort complexity.

Remaining acceptance: the maintainer has confirmed the minimal scope is acceptable and intends to merge after required checks pass; the current Test matrix is still running. Link Check is a known baseline failure, not a regression introduced by this PR. The branch is mergeable without git conflicts. No separate CI or unrelated changelog fixes are proposed here.

Original design discussion: #1235 .

[lindicaphxag-tech]

Focused follow-up on the current 3-file implementation: I found and fixed a concrete grouping edge case that matters for the intended (subject, session, run) identity: comparing an object NumPy vector of tuple IDs via runs == run_tuple attempts array broadcasting instead of atomic tuple equality. The current head first assigns each hashable recording identity a stable scalar code, so independent subjects/sessions with reused run labels remain separate. I added a tuple-ID regression and an explicit invalid-unhashable-ID test; no DP balancing, implicit timing, or preprocessing changes were introduced.

The branch remains mergeable and pre-commit.ci plus What's New checks are green. The full upstream Test/Docs/Braindecode jobs on the new exact head are still running, so I am not claiming final CI success. Link Check's current failure points to the existing https://github.com/qinxwew references in the shared changelog, not to links introduced here. The contract remains deliberately narrow: raw epoch-window non-overlap does not imply post-IIR signal independence. Thanks again for considering the smaller splitter-only direction from #1235.

[lindicaphxag-tech]

CI blocker addressed on current head 10ad547: the previously failing upstream Link Check now reports success (run #40216319081). The two failing LiQing reference definitions were already in the upstream changelog and pointed to a 404 profile; I replaced them with a single attribution link to the same contributor's MOABB PR history. No new file, splitter behavior, or API was changed.

The previous exact code head completed the full MOABB OS/Python Test matrix, Docs, Braindecode integration, What's New and pre-commit successfully. Because this new commit edits the changelog, GitHub reran the checks; the new-head Test/Docs/Braindecode/What's New runs are still pending/running, so I am not claiming final all-green yet. The PR remains three files and mergeable. Once the current rerun finishes, maintainer review of this deliberately scoped opt-in primitive is the remaining substantive acceptance step. Thanks for the earlier specific scope guidance on #1235.

[lindicaphxag-tech]

Exact-head CI exception identified (run #40216319047, Braindecode integration): the job progressed through roughly 87% of its test suite and then emitted The runner has received a shutdown signal followed by The operation was canceled; the test step was marked cancelled, not a test assertion failure. I inspected the raw job logs rather than assuming a code regression. This current PR changes only splitters.py, its focused tests and the changelog. The same integration job passed on previous exact head #40211841109, prior to the subsequent changelog-only edit. The latest Link Check, pre-commit and What's New jobs are green.

I attempted GitHub Actions' rerun failed jobs endpoint, but GitHub returned 403 Resource not accessible by integration; an upstream member with Actions write permissions would need to rerun this interrupted job. Could a maintainer rerun it if appropriate? Please treat the current run as infrastructure interruption, not proof that the new splitter failed its tests; the other latest checks are still completing.

[bruAristimunha]

Thanks @lindicaphxag-tech, this is the scope I asked for in #1235. I pushed one commit to your branch: a develop merge, the :gh:1329bylindicaphxag-tech_ credit on the changelog entry, and the two LiQing link lines put back as on develop (unrelated to this PR; happy to take that fix separately). Merging once the required checks pass. The red braindecode-dev job is the runner shutdown you describe; it is not a required check.