fix(meg): correct M100 measurement (standards-only events + in-window peak search) + study_status

[juniperbevensee]

Confirmed root cause (live, real data)

The meg_validation study reported the auditory M100 peak at ~15 ms (refuted) on the Brainstorm auditory dataset, where the real M100 is ~100–120 ms. A direct mne probe on the real data isolated two compounding defects:

1. measure_peak searched too wide. It searched window_ms ± 100 ms (≈ −20 ms to 220 ms in abs mode for an 80–120 ms window), so it grabbed the large early stimulus artifact at ~15 ms instead of the M100.
   • Probe A — all-events + wide search → peak 15 ms.
2. epoch did not filter events. It used mne.find_events(raw) and epoched all triggers (in a 30 s crop: standards=16, deviants=3, other/button=3), muddying the average.
   • Probe B — standards-only (event_id=1) + search within 0.08–0.12 s → peak 120 ms, a real in-window auditory response → would be supported.

So the fix is: filter to the standard-tone event and search within the expected window (no ±100 ms widening).

Fixes

• In-window peak search. Extracted a pure, mne-free _peak_search_bounds(window_ms, times_lo, times_hi) -> (lo, hi) that keeps the search within the expected window (small symmetric ±10 ms margin), clamped to available sample times. _MneIO.measure_peak calls it then ev.get_peak(tmin=lo, tmax=hi, ...). No more ±100 ms widening.
• Standards-only epoching. _MneIO.epoch filters to a single trigger code. event_id semantics: None → auto = most-frequent code (the standard tone) [default]; an int (via bounds.event_id) selects explicitly; "all" is a diagnostic escape hatch (legacy unfiltered behavior). The resolved event_id threads through the step result + artifact meta so a fake io can assert filtering is applied.
• More trials, still bounded. Raised default crop_tmax 30 → 90 s so the standards-only average has enough trials for a clean peak (a 30 s crop yielded only ~16 standard tones). Still clearly memory-safe: the channel-subset + resample keep the loaded array to a few MB, and per-step checkpointing means an OOM resumes from the last completed step. Documented inline.
• study_status envelope fix. On a real run, study_status returned an error because json.dumps(..., default=str) emits bare NaN/Infinity for non-finite measured floats — invalid JSON, which a strict downstream parser rejects (the output began {"study_id":2,... then broke). Tool envelopes (_tool_result/_tool_error) now recursively sanitize non-finite floats to null and dump with allow_nan=False, so every envelope is strictly valid JSON.

New tests (stdlib-only — CI has no mne/numpy)

• test_peak_search_bounds_stay_in_window_not_widened_by_100ms — the regression test that would have caught the ±100 ms bug: window (80,120) over wide times → ~(0.08, 0.12), not (−0.02, 0.22); the ~15 ms artifact is outside the search.
• test_peak_search_bounds_clamped_to_available_times — clamps to available samples.
• test_epoch_applies_standards_event_filter_by_default — a fake io records the event_id; the study applies the standards filter by default (not "all").
• test_epoch_event_id_overridable_via_bounds — bounds.event_id overrides.
• test_study_status_returns_strict_json_for_completed_study — completed study with NaN/Infinity in finding evidence returns a strict-JSON-parseable structured result (non-finite → null), not an error envelope.

Verification

• .venv/bin/python -m pytest -q → 136 passed, 7 skipped (opt-in MATILDE_LIVE=1 test stays skipped in CI).
• No top-level mne/numpy imports added; import matilde_plugin.engine.meg_study works without mne.
• Deterministic sanitization layer clean.

Do not merge — left open for independent audit.

🤖 Generated with Claude Code