feat(study): MEG data-sample validation study (M100, memory-bounded, resumable)

[juniperbevensee]

What it does

Adds a meg_validation study type — the first heavy / real-data study built on the stateful study pipeline. It validates the auditory M100 evoked-response claim (a magnetic field peak ~80–120 ms after an auditory stimulus) against a bounded sample of an open MEG dataset (bst_auditory), as five resumable, checkpointed steps:

fetch_sample → preprocess → epoch → evoked → validate_finding

validate_finding emits a finding: supported if the measured peak latency is within the expected window (default 80–120 ms), refuted if clearly outside, inconclusive if the sample is degenerate (no measurable peak). Evidence = measured latency + amplitude.

Lazy-mne + injected-I/O testing approach

• Lazy heavy deps. mne/numpy are imported only inside the real backend's methods (_MneIO), never at module top. matilde_plugin.engine.meg_study — and the whole plugin — imports cleanly with neither installed (verified in a venv with no mne/numpy). grep confirms no top-level heavy imports anywhere in engine/.
• Injected I/O. Every step calls through an injected MegIO handle. Production passes none and gets the real backend; tests pass fakes — no mne, no numpy, no network, no data files, fully deterministic and stdlib-only.
• The tools layer uses a small _TEST_IO hook so the offline dispatch test stays stdlib-only; production leaves it unset (real lazy-mne).

Memory-bounded by design

The real backend never full-preloads a recording: preload=False, one run, a gradiometer channel subset, crop to a stimulus-locked window, downsample — then persists small intermediates. Each step reloads the prior step's on-disk artifact rather than holding live objects, so a memory kill or container rebuild resumes from the last completed step instead of re-fetching the whole chain.

Resumability evidence

A unit test forces a failure at the evoked step: the study goes blocked with fetch_sample/preprocess/epoch done and no finding emitted. On resume with a working backend, call counters prove the earlier steps are not re-run (fetch/preprocess/epoch = 0 calls), only the failed step and the remaining one execute, and the study completes with the expected finding.

Tests

• New: tests/test_meg_study.py (step transitions, verdict boundaries 100ms→supported / 300ms→refuted / empty→inconclusive, custom window, resumability) and tests/test_meg_study_tools.py (create+run dispatch through the tools).
• Opt-in live test (MATILDE_LIVE=1 + mne importable, via pytest.importorskip): runs the real bounded pipeline on bst_auditory and asserts a peak in a plausible window. Skipped without the flag/mne — that's expected.
• Full suite: 131 passed, 7 skipped (the live test is one of the skips).

Do not merge — leaving open for independent audit.