Skip to content

CHANGELOG.md

Latest commit

 

History

History
854 lines (742 loc) · 52.2 KB

CHANGELOG.md

File metadata and controls

854 lines (742 loc) · 52.2 KB
Raw

Changelog

All notable changes to MNE-CPP will be documented in this file.

[2.4.0-rc] - 2026-10-23

Theme: correctness you can check. v2.4.0 compares MNE-CPP against MNE-Python 1.11 wherever both compute the same thing. That comparison found and fixed more than 20 numerical defects in inverse, forward, beamformer, filtering and spectral code and more than 200 other bugs. The release adds sEEG/ECoG support, about 60 new MNE-Python ports, a WebAssembly renderer that matches the desktop, and a signing pipeline for the Windows and macOS installers. 970 commits since v2.3.0.

This release candidate replaces the v2.4.0 tag of 2026-10-22, whose Windows downloads could not be signed.

Highlights

  • Code signing — every Windows executable and DLL carries a version resource with product name MNE-CPP and the release version, checked before each SignPath request against the artifact configurations in .signpath/. The macOS release is prepared for Developer ID signing and notarization of the apps, the .dmg images and the installer. The code signing policy states what is signed, by whom, and what the applications send over the network.
  • Connected documentation — the manual, API reference and examples link to each other by scientific topic; the examples page can be filtered by task, recording type, level and library; a new Get Started guide builds a first program against an installed MNE-CPP, and every example states what it computes.
  • sEEG and ECoG — new mne_ieeg library (implantations, leads, bipolar montages, projection of contact values onto the cortex). MNE Inspect draws and picks depth and grid electrodes from BIDS *_electrodes.tsv and paints per-contact values onto the brain, on the desktop and in the browser.
  • Validated against MNE-Python — about 60 cross-validation test suites compare MNE-CPP with MNE-Python 1.11, SciPy, NiBabel and MNE-BIDS on committed reference values, from FIFF reading to inverse solutions. A versioned inventory lists, for 443 MNE-Python functions, whether MNE-CPP has an equivalent, a partial port or nothing yet, with the test that proves it.
  • Inverse and forward results corrected:
    • Minimum-norm kernels were 0.8 % off.
    • eLORETA was off by a factor of 2 to 4; loose eLORETA was 24 % off.
    • Loose dSPM norms were 17 % off.
    • LCMV/DICS whitening was wrong.
    • The sphere-model forward crashed.
    • Magnetic-dipole fits ended 18 mm off.
    • Details are listed under Fixed.
  • File formats corrected:
    • EDF channels were scaled by 1e6, and BDF was decoded as 16-bit.
    • MNE Browse saved EEG/EOG/STI channels 3276.8 times too large.
    • FIF event files were written scrambled.
    • Raw files with acquisition skips crashed.
    • FIFF writes used host byte order.
  • New MNE-Python ports — FIR filter design (create_filter, filter_data, firwin2), ECG/EOG detection and projectors, Welch, multitaper and Morlet PSD/CSD/TFR, muscle and amplitude annotation, EOG regression, permutation t-test, combine_evoked, apply_forward, compute_proj_evoked, estimate_snr, grade_to_vertices, grow_labels, event utilities and six digitizer formats.
  • 3D in the browser — WebAssembly and OpenGL now use the same rendering path as the desktop. A long-standing cause of garbled or blank WebGL frames is fixed. Source-estimate animation is about twice as fast and picking is about 1000 times faster.
  • MNE Scan shows the head-movement path from continuous HPI in 3D.
  • MNE Analyze gains event durations and trigger codes, averaging per condition, and saving of averages.
  • Documentation you can run — every public class has a compiled, tested example (29 new example programs). The API reference has zero Doxygen warnings. Manual screenshots are generated by CI from the current build.
  • Trustworthy releases — signed Windows installers and archives, signed and notarized macOS disk images, published SHA-256 checksums, and CI that installs MNE-CPP on clean Linux, macOS and Windows machines following the build guide.

New library: mne_ieeg

  • IeegImplantation:
    • Groups contacts into leads.
    • Reads and writes the BIDS iEEG group and dimension columns.
    • Transforms between coordinate frames.
    • Supports declared lead types (depth, strip, grid).
  • IeegReference builds bipolar montages along each lead, named as in MNE-Python (LA1-LA2), as a sparse derivation matrix.
  • IeegSurfaceMap projects contact values onto the cortex like mne.stc_near_sensors (sum, single, nearest and weighted modes).
  • Example ex_ieeg.

New APIs

Reading and writing data

  • FiffExplain names FIFF tag kinds, block kinds, units, unit multipliers and channel kinds; FiffDirNode::print writes the block tree to a stream.
  • Digitizer readers on FiffDigPointSet:
    • Polhemus Isotrak (.elp/.hsp/.eeg) and Polhemus FastSCAN.
    • BrainProducts CapTrak (.bvct).
    • EGI MFF.
    • Localite (.csv).
    • Neuroscan (.dat).
    • transformToHead.
  • Head → MNI → Talairach transforms (FiffCoordTransSet::readMniTransform, headToMni, mniToTalairach); mne_collect_transforms --mgh.
  • Events (FiffEvents): pick, merge, count, concatenate, find_stim_steps, make_fixed_length.
  • Channels:
    • channel_indices_by_type, pick_channels_regexp and pick_data_channels.
    • Every MNE-Python channel type, including seeg, ecog, dbs, fNIRS and eye tracking.
  • FiffEvoked::combine (combine_evoked), FiffEvokedSet::save, FiffProj::compute_from_evoked, FiffCov::read without exceptions, and covariance from epochs already in memory.
  • HPI coil frequencies are read from the measurement info.
  • MNE-C formats: layouts (.lout), event lists, derivations (MNEDerivSet), and morph maps (MNEMorphMap, used by mne_make_morph_maps).
  • BEM solutions written by MNE-CPP are readable by MNE-C and MNE-Python.

Source estimation

  • applyForward, estimateSnr, MNEForwardSolution::convert_to_surf_ori and MNESourceSpace::gradeToVertices.
  • InvSourceEstimate::writeHemispherePair writes MNE-Python-compatible -lh.stc/-rh.stc files, also for LCMV and DICS results.
  • The MNE-C minimum-norm quantities: eigenfield projection, SNR, lambda² and predicted data.

Signal processing and statistics

  • FirFilter::designMne, filterData and firwin2, with window choice.
  • ArtifactDetect: computeProjEcg/computeProjEog and peakFinderMne.
  • setEegReference can add an average-reference projector instead of changing the data.
  • Morlet wavelets (MorletTfr::buildWavelet), multitaper TFR on arrays, DPSS with sym/lowBias and concentration ratios, and Welch with DC removal.
  • Statistics:
    • Permutation t-test.
    • StatsAdjacency::combine and chain.
    • Benjamini-Yekutieli FDR (negcorr) and reject masks.
    • Numerics::chi2Isf.
  • FsLabelUtils::growCircularLabel (grow_labels). KMeans matches MATLAB for start centroids, Hamming distance and empty clusters.

Workflows (MNA)

  • The graph executor runs verification checks and records provenance: versions, host, a SHA-256 hash per input, and timing.
  • Parameter expressions support comparisons.

Display

  • BrainView:
    • Electrodes: setElectrodes, setElectrodeValues and the electrodeContactClicked signal.
    • setHeadMovementPath, takeScreenshot and setAutomatedRotation.
  • RtFiffRawView shows a crosshair with the value under the cursor. ChannelDataView::renderToImage renders traces without a GPU.
  • Events carry a duration and a trigger code (addRangedEvent, setEventDuration, setEventCode).
  • MNELogger::setLogFile. Debug and info messages go to stdout; warnings and errors go to stderr.

Applications

MNE Scan

  • Shows the head-movement path from continuous HPI fitting in the 3D view. Coil frequencies are taken from the recording. Digitizer points appear when the HPI view opens.
  • Writes a log file. Plugin load errors are reported, and MNE_SCAN_PLUGIN_DIR overrides the plugin location.
  • The clustered-MNE plugin is renamed rtclustmne and accepts unclustered forward solutions.
  • Robustness:
    • Survives configurations naming unknown plugins.
    • Plugins stop cleanly before any data arrived.
    • Saving to an unwritable path is reported.
    • The dark theme and font are bundled again.

MNE Analyze

  • Events have a duration, a trigger code and a type. Ranged annotations are drawn as bands, and codes are shared between instances.
  • Averaging computes one evoked response per trigger code. Save average is implemented.
  • Opens BEM files under any name and .dip dipole files.
  • View ▸ Plugins menu. Automated 3D rotation works.
  • Raw and event export work in the browser build.

MNE Inspect (desktop and browser)

  • sEEG/ECoG electrodes, File ▸ Load Electrode Values…, dipoles placed in MRI coordinates.
  • View layout and zoom are restored when you reopen a session.

MNE Analyze Studio

  • Offline mode, a workflow preview, and FIFF files opening in the raw browser on Linux.
  • Kernel and skill errors are reported instead of failing silently.

MNE Browse

  • Event markers use the colours chosen for each event type.

Command-line tools

  • mne_edf2fiff converts through the corrected EDF reader.
  • mne_eximia2fiff works again; every conversion used to fail.
  • mne_prepare_bem_model and mne_setup_forward_model write BEM solutions that other tools can read.
  • mne_make_morph_maps writes real morph maps. mne_make_derivations writes the MNE-C format.
  • mne_collect_transforms collects the full transform chain. mne_transform_points accepts multi-transform files.
  • mne_dipole_fit honours --grid, the guess radius and --magdip.
  • mne_process_raw's grand average now matches MNE-C.
  • Every tool returns a non-zero exit code on failure and writes errors to stderr.

Real-time

  • mne_rt_server writes the UTC offset correctly, and the real-time client calibrates CTF compensators like the file reader.
  • LSL streams are found on the same machine without multicast routing.

Fixed

Source estimation

  • Minimum-norm kernels:
    • 0.8 % off: the wrong noise eigenvalues were dropped.
    • eLORETA read its eigenfields transposed (off by a factor of 2 to 4).
    • Fixed-orientation dSPM/sLORETA had an empty noise normalisation.
    • The depth prior summed per channel instead of per source.
  • Loose inverses on a Cartesian forward silently became loose = 1 (dSPM norms 17 % off). Loose eLORETA lacked a renormalisation step (24 % off).
  • Fixed-orientation solutions:
    • Forward reads used vertex instead of patch normals (1-2 % error).
    • Inverses built from free-orientation forwards were wrong.
    • force_fixed reads overwrote left-hemisphere rows.
  • The sample-covariance whitener kept rank 58 of 360. CMNE whitening dropped every MEG direction.
  • Beamformers:
    • Loading factor wrong.
    • pick_ori="normal" gave an empty filter.
    • NAI normalisation and rank handling wrong.
    • The whitener was transposed.
    • Max-power orientation signs were flipped.
    • LCMV used the wrong channels when the evoked channel order differed.
  • Label-restricted inverses offset right-hemisphere vertices and kept one orientation instead of three.
  • Sparse solvers:
    • MxNE stopped 43 % above the optimum.
    • Gamma-MAP pruned too early.
    • TF-MxNE converged short.
    • TRAP-MUSIC missed sources.
    • RAP-MUSIC read out of bounds.
  • CMNE re-weighted the wrong quantity.

Forward modelling and dipole fitting

  • Every sphere-model forward computation crashed.
  • MEG/EEG gradients were stored in the wrong matrix.
  • Magnetic-dipole fits computed fields at EEG electrodes and fitted current dipoles (ending 18 mm off).
  • Icosahedron downsampling used the wrong base: only 63 of 2562 vertices matched.
  • Volume source spaces listed excluded points.
  • Volume centre of mass swapped x and z.
  • Dipole fit:
    • Skipped or doubled the SSP projection.
    • The baseline used the end time for both bounds.
    • --grid and the guess radius were ignored.
    • MNE Analyze defaulted the sphere radius to 0.09 mm.
  • On Windows, the Berg-Scherg sphere fit converged to a worse minimum.

Preprocessing, epochs and averaging

  • SSP vectors from raw data spanned the wrong subspace. Covariance removed the grand mean instead of the evoked response, and ignored rejection limits.
  • Epochs started one sample early. Baselines dropped the stimulus sample.
  • Epoch rejection with a limit for only some channel types (e.g. gradiometers) rejected every epoch.
  • Averages dropped their last sample, and the real-time average mixed stale pre-stimulus data across trigger codes (12 % error).
  • Saved filtered raw files were shifted by half the filter length, and trigger channels were low-pass filtered with the data.
  • SSS swapped its internal and external bases. Fine-calibration files now use the Elekta format.
  • Spectra:
    • Morlet power was about 88 times too large.
    • PSD/CSD scaling and DC handling were wrong.
    • Real-time noise PSD used |X| instead of |X|².
  • The surface Laplacian was centred on the electrodes instead of the head sphere.
  • Standard montages used invented positions; they now use MNE-Python's standard_1005.
  • Statistics:
    • The one-tailed cluster threshold was wrong.
    • The TFCE grid was wrong.
    • p-value tails below 1e-16 were wrong.
    • FDR ignored alpha.
    • Permutation p-values left out the observed statistic.
  • Connectivity: correlation did not return Pearson r, and node positions came from the wrong channel.

Files

  • FiffDirNode::print split every fragment into its own log line.
  • EDF/BDF:
    • Every EDF channel was divided by 1e6.
    • Padded header fields read as 0.
    • BDF was decoded as 16-bit.
    • Channel kinds were wrong.
  • BrainVision: unknown units read as zero, channel types were wrong, and ASCII files were read as binary.
  • BIDS:
    • Electrode positions were attached to the wrong channels.
    • Units were wiped.
    • recording was written as rec.
    • Run numbers were padded.
    • The digitizer frame was wrong.
  • FIFF:
    • Writes used host byte order.
    • UTF-8 strings were truncated.
    • Saving picked channels twice and calibrated evoked data twice.
    • Event files were scrambled.
    • The UTC offset and the inverse source unit were lost.
    • Old-style evoked files could not be read.
    • Corrupt tags were read out of bounds.
    • Files with acquisition skips or no data crashed.
  • CTF compensation: the inverse transform was wrong, calibration was re-applied on write, and partial reads were returned silently.
  • FreeSurfer:
    • Surface normals were overwritten instead of summed (affects mne_surf2bem).
    • Quad surfaces were scrambled.
    • Curvature or truncated files crashed the reader.
    • Asking for a hemisphere that was not loaded crashed.
  • MRI: truncated .nii.gz files looped forever, and the Talairach transform frames were wrong.

Applications

  • macOS (Apple silicon): the apps of the dynamic package and the installer were killed at launch because packaging invalidated their signatures.
  • macOS and Linux dynamic packages: resources (layouts, styles, server configuration) was a link to the build machine instead of the files.
  • macOS packages only started on the build machine's macOS (26); they now run on macOS 14.4 and later, the minimum of the bundled Qt.
  • MNE Browse:
    • Saved EEG/EOG/STI channels 3276.8 times too large.
    • Could overwrite its own input.
    • Wrote empty annotation and event files.
    • Annotation durations grew on every save.
  • MNE Analyze and MNE Scan: the butterfly plot showed an empty (on some screens magenta) area instead of the averages, and its screenshots were empty.
  • MNE Analyze:
    • Events: event files never opened, trigger detection found no onsets, and reselecting a recording discarded its events.
    • Coregistration paired fiducials by position, ran ICP widgets from a worker thread, and read weights as 0 in decimal-comma locales.
    • Crashes fixed: adding an average without a recording, jumping to an event with nothing selected, channel selection, and a 13 GB allocation from an uninitialised pointer.
  • MNE Scan:
    • Crashed on shutdown and on unknown plugins.
    • Reported 4294967295 channels before the channel info arrived.
    • Took screenshots of the wrong view.
    • Did not restore scalings.
  • The real-time raw view drew markers offset after the buffer wrapped. The butterfly view compressed long averages. The spectrum view drew upside down. Several settings views lost their values without a settings file, and colour dialogs turned items black on Cancel.
  • 3D view:
    • Thresholded source estimates showed the colormap's zero colour instead of the cortex.
    • Real-time sensor colouring never reached the head.
    • Dipoles were about 4 cm off in MRI coordinates.
    • Screenshots were written to /.
  • Coregistration: the ICP nasion pin pulled exact alignments 25 mm off; optical calibration ended 24 mm RMS off.
  • FiffFileSharer never delivered files (MNE Scan → MNE Analyze hand-over).

Performance

  • Source-estimate animation on two pial surfaces: 4.4 → 1.9 ms per frame. Hover highlighting: 10 ms → under 1 µs. Ray picking: 1.6 ms → about 1 µs.
  • FIR filtering uses power-of-two FFTs (a test filter went from 13 s to 18 ms). The Hilbert envelope is FFT-based instead of O(n²).
  • The browser build no longer re-uploads geometry every frame.

For developers

Breaking changes

  • Renamed:
    • macOS and Linux deploy with scripts/deploy/deploy.sh; deploy.bat is Windows only.
    • disp3D types live in the DISP3DLIB namespace; PickResult is now RayHit.
    • MNE Scan plugin rtcmne is now rtclustmne (RtClustMne).
  • Changed defaults:
    • FiffCoordTrans::read returns the first transform in a file; use readTransform(from, to) to choose one.
    • MNEForwardSolution::read keeps bad channels by default.
    • WelchPsd removes the DC by default, and CSD no longer returns the DC bin.
    • KMeans uses the "singleton" empty-cluster action.
    • StandardMontage positions are now MNE-Python's.
  • Changed values:
    • ECoG/sEEG/DBS channel kinds use the FIFF-standard values 902/802/803.
    • The BIDS entity is written as recording.
  • Changed signatures:
    • FIFF writers, checkIntegrity() and FiffEvokedSet::compensate_to return bool.
    • Ownership-transferring factories return std::unique_ptr.
    • CircularBuffer cannot be copied.
  • Removed:
    • ElectrodeArray/ElectrodeContact/ElectrodeLayout (use IeegImplantation).
    • EDFReader scale factor and mne_edf2fiff --scaleFactor.
    • MNEHemisphere::getTriCoords.
    • FiffDirNode::explain/explain_block/get_tag_explanation/ get_unit_name and FiffExplainEntry (use FiffExplain).
    • mne_browse MnxProject (use the MNA project container).
    • Several unused internals.
  • MNE Scan's bad-channel resource is renamed babymeg_bad_channels.txt. The MNE Analyze event shared-memory key is versioned.

Quality

  • Every compiler warning is an error on all platforms, including the WebAssembly build.
  • The source tree is formatted with clang-format 21.1.8 (one reformat commit, ignored by git blame), and clang-tidy findings may not grow.
  • No raw new, NULL, C-style casts, printf or numeric #defines remain. Ownership uses smart pointers.
  • Tests:
    • 34 tests that had never run in CI now run, and 26 unregistered tests were found.
    • Every test is registered with a label and a timeout.
    • Tests run against both the oldest and the newest supported Qt.
  • Line coverage: libraries 89 %, applications 56 %, tools 61 %. Library branch coverage: 65 %.
  • A layering check keeps the libraries free of dependency cycles.

Build and release

  • Qt 6.10.0 to 6.12.0, Emscripten 6.0.12, ONNX Runtime 1.31.0. Versions are kept in src/external/external_deps.env.
  • doc/wiki/Release-Process.md describes:
    • A release-candidate check before every release.
    • Windows Authenticode signing through SignPath, and macOS signing and notarization.
    • SHA256SUMS with per-file signing status.
    • Verification of the published downloads on all three platforms.
  • A weekly onboarding job follows the build guide on clean Linux, macOS and Windows machines. It found three setup bugs, fixed in this release.
  • SECURITY.md added.

Documentation

  • 29 new example programs, compiled and run by CI. Every public class has an example snippet in its API page; internal types show the examples that exercise them.
  • The API reference builds with zero Doxygen warnings (1,726 fixed).
  • Manual screenshots are generated from the current build in CI and compared against reference images.
  • The website build uses a patched braces (nesting-depth bound for GHSA-vfj7-8cjw-p6xm, which upstream has not fixed), follow-redirects 1.16.1 and launch-editor 2.14.2.
  • The website fails to build on broken links or missing images, and has a generated quality page.
  • The build guide is complete for all three platforms; CI runs exactly the commands it shows.

Contributors

Christoph Dinh, Eric Larson (leaner CI on staging, #1013), ag-fieldline (FIFF files without raw data no longer crash, #1003); dependency updates by Dependabot.

[2.3.0] - 2026-06-13

Highlights

  • MNE Inspect multimodal visualization — Unified MultimodalScene integrating sEEG depth electrodes, ECoG grids/strips (ElectrodeObject), MRI orthogonal slices (SliceObject), cortical surfaces with STC overlays, dipoles, connectivity arcs, and sensor locations; cross-modality picking with shared PickResult and timeline; data-driven colormap controls.
  • MNE Scan production hardening — Plugin audit, real-time CMNE wiring (MNE/dSPM/sLORETA/CMNE method selector), recording/replay/resume, quickstart/hardware/deployment documentation, packaging refinements.
  • MNE Analyze cortical workflow — FreeSurfer surface plugin, STC overlay with threshold/mid/max colormap and time slider, vertex picking and timecourse display.
  • MNE Align coregistration wizard — 7-step wizard-guided coregistration tool with Polhemus FASTRAK serial driver, fiducial/HSP/EEG-cap digitization, Kabsch SVD + ICP alignment, -trans.fif export; session persistence; optical calibration engine with constrained solver.
  • Video overlay & probe tracking — Tracker-driven video overlay projection on head surface with billboard fallback and vertex-displaced depth relief; catheter probe visualization (station 3) with tip-only crosshair, glow aura, and ethereal instrument shader.
  • MRI slice rendering — 3D orthogonal MRI slice rendering with per-viewport visibility, intensity windowing controls, voxel-to-world transforms, and session persistence.
  • MNE-Python algorithm parity (~92%) — Covariance estimation (auto/shrunk/OAS/FA/PCA), inverse convenience API, preprocessing & artifact detection, standard montages (10-20/10-10/10-05), Maxwell movement compensation, multitaper TFR, Picard ICA, surface Laplacian, auto ICA classification, depth priors, resolution metrics, SourceMorph, TRAP-MUSIC, TF-MxNE, CSP/SPoC/SSD decoding, simulation API.
  • Skigen integration — ML algorithms (StandardScaler, MinMaxScaler, LOF, LedoitWolf, OAS, EmpiricalCovariance, FactorAnalysis, CSP, SPoC, SSD) delegated to skigen; API provenance registry with CI validation.
  • SPDX license-header normalization — 1013 files migrated to unified SPDX+Doxygen headers; CI gate enforces compliance on every PR.
  • Doxygen-to-Docusaurus API docs — Complete automated pipeline (doxy2mdx.py) generating 355 MDX pages (336 classes + 19 library indexes) with Mermaid inheritance diagrams and cross-reference hyperlinks.

New Features

  • Covariance estimation parity (TASK 12): Auto/shrunk/OAS/FA/PCA/regularize/whitener methods; apply_inverse_epochs/raw/cov, beamformer epochs, source PSD/power convenience API
  • Preprocessing & artifact detection parity (TASK 13): Maxwell bad-channel detection, EOG regression, EEG re-referencing, cHPI filter, stim artifact removal, muscle/amplitude annotation, bridged electrode detection, LOF bad-channel detection, fine calibration, annotation↔event conversion
  • Source estimate types (TASK 14): VectorSourceEstimate, VolSourceEstimate, MixedSourceEstimate; label operations (extract_label_time_course with 5 modes, stc↔label, grow/split labels)
  • Decoding (TASK 15): CSP/SPoC/SSD/SlidingEstimator/GeneralizingEstimator/ReceptiveField; I/O export (EDF/BrainVision), AEC/PEC connectivity, HTML Report, linear regression, RM-ANOVA, bootstrap CI, fixed-length epochs
  • Standard montages (TASK 6): 10-20/10-10/10-05 electrode montage support
  • Maxwell movement compensation (TASK 6): Full Maxwell filter movement compensation
  • Multitaper TFR (TASK 7): Multitaper time-frequency representation
  • Picard ICA (TASK 7): Picard-based ICA decomposition algorithm
  • Surface Laplacian (TASK 7): Current source density via surface Laplacian
  • Auto ICA classification (TASK 7): Automated ICA component labeling
  • Depth priors & resolution metrics (TASK 8): PSF/CTF resolution metrics for source estimation
  • Full SourceMorph (TASK 8): Complete cross-subject cortical morphing
  • TRAP-MUSIC (TASK 8): TRAP-MUSIC source localization algorithm
  • TF-MxNE (TASK 6): Time-frequency mixed-norm estimates
  • Simulation API (TASK 9): simulate_raw/_stc/_evoked, addNoise/Ecg/Eog/Chpi in new sim library
  • MNE Inspect plugins: Electrode visualization plugin (inspect_electrodes), MRI slices plugin (inspect_mri_slices)
  • MNE Analyze Studio skills: MNA-aware agent skills extension (mna_skills)
  • MNE Analyze cortical surface plugin: STC load with fThresh/fMid/fMax colormap, time slider, play/pause
  • Real-time CMNE: CMNE method added to MNE Scan minimum norm settings (combo: MNE/dSPM/sLORETA/CMNE) with model-checkpoint picker
  • Polhemus FASTRAK driver: Serial backend with auto-detection, pivot calibration, optical path calibration engine, Kabsch SVD registration with vertex disambiguation
  • NIfTI reader: NIfTI-1 single-file reader (.nii/.nii.gz) for MRI slice plugins
  • Documentation screenshot tool: mne_doc_shots manifest-driven screenshot generation for manual pages
  • macOS .dmg generation: create-dmg integration with codesign and notarization hooks
  • API provenance registry: doc/api_registry.json with 90+ entries; tools/validate_api_registry.py CI validation

MNE Inspect

  • Multimodal visualization — Integrated MultimodalScene with electrode, MRI slice, brain surface, STC overlay, dipole, connectivity, and sensor renderables
  • Electrode plugin — sEEG depth electrodes and ECoG grid/strip visualization with instanced contact spheres
  • MRI slices plugin — 3D orthogonal MRI slice rendering with per-viewport visibility, intensity windowing controls, voxel-to-world transforms, and session persistence
  • Video overlay — Tracker-driven video texture projection on head surface with billboard fallback and vertex-displaced depth relief
  • Probe tracking — Catheter probe visualization (station 3) with tip-only rotating crosshair, glow aura, and ethereal instrument shader

MNE Scan

  • Real-time CMNE — Cascaded MNE method added to minimum norm plugin with MNE/dSPM/sLORETA/CMNE selector and ONNX model-checkpoint picker
  • Quickstart documentation — Hardware setup, deployment, and quickstart manual pages

MNE Align

  • New coregistration application — 7-step wizard: Load FIFF digitizer points → Verify → Acquire fiducials → Acquire head shape → Acquire EEG cap → ICP alignment → Save -trans.fif
  • Polhemus FASTRAK integration — Live serial acquisition with auto-detect, pivot calibration, hemisphere configuration
  • 3D visualization — Interactive BrainView with signal-driven sync, session persistence, marker rendering
  • Session persistence — Full save/restore of coregistration state including acquired points and nested group prefixes
  • Optical calibration — Constrained solver with known tracker-to-objective distance, sample persistence, direct objective center capture

Skigen Integration

  • Delegated algorithms: StandardScaler, MinMaxScaler, LOF, LedoitWolf, OAS, EmpiricalCovariance, FactorAnalysis, CSP, SPoC, SSD
  • Policy: Never reimplement sklearn algorithms in mne-cpp; implement in skigen first
  • Init scripts: Updated init.sh/init.bat to clone skigen automatically
  • CI: API registry validation workflow

CI/CD

  • Fixed polyglot .bat SPDX headers causing bash syntax errors on Linux/macOS — REM Copyright (c) replaced with :;# comment style valid in both shells
  • Fixed mne_doc_shots link errors on Linux with code coverage (__gcov_exit unresolved) — added --coverage compile/link flags
  • Fixed mne_doc_shots MSVC RuntimeLibrary mismatch on Windows (MD_DynamicRelease vs MT_StaticRelease) — set MSVC_RUNTIME_LIBRARY to match project libraries
  • Added SPDX license-header CI gate (license-headers.yml)
  • Added API registry validation CI gate (api-registry-validate.yml)
  • Added Doxygen API docs CI pipeline (api-docs.yml)
  • Added screenshot regression CI (screenshot-regression.yml)
  • Fixed CI deploy auth: replaced expired GIT_CREDENTIALS with GH_PAT
  • Fixed Linux Qt builds: added XKB/XCB/EGL/GLES dev library dependencies
  • Fixed Windows CI: use vswhere for VS toolchain discovery
  • Fixed macOS CI: clear stale pip cache before setup-python
  • Fixed 15 pipeline annotations in staging/codeql workflows
  • Removed broken auto-fix-build workflow
  • Consolidated license-headers workflow into staging pipeline
  • Qt 6.11.2 toolchain across all CI workflows

Documentation

  • SPDX license headers: 1013 files migrated to unified SPDX+Doxygen format with chronological author order and rename history
  • Doxygen API docs pipeline: tools/doxy2mdx/doxy2mdx.py (1394 lines) generating 355 MDX pages with Mermaid inheritance diagrams, XML-backed cross-references, and per-library indexes
  • MNE Align manual (mne-align.mdx): 7-step wizard documentation
  • MNE Scan manual (mne-scan-hardware.mdx, mne-scan-quickstart.mdx, mne-scan-deployment.mdx): 18 plugins enumerated
  • MNE Analyze Studio manual (mne-analyze-studio.mdx): Agent-based workflow documentation
  • MNE Inspect multimodal manual (mne-inspect-multimodal.mdx): Multimodal visualization guide
  • Build guide: Updated for init-script workflow, Skigen dependency, and script paths
  • API sidebar: Restructured — Namespaces/Classes/Files/Hierarchy/Members nested under Overview category
  • Thin-plate-spline warp example (Closes #475)

Refactoring

  • Renamed conn library to connectivity to match mne-python naming
  • Removed redundant null checks before delete (Closes #857)
  • Modernized header format: unified v3 SPDX+Doxygen with @author/@since/@date tags
  • Renamed FastTrak → Fastrak (correct product name)

Bug Fixes

  • Fixed QWARN deprecation warnings in test_mne_io_coverage (replaced with qWarning())
  • Fixed FIFF coordinate transformations: retain all transforms read from a file
  • Fixed high-DPI rendering on fractional-scale monitors
  • Fixed HPI digitizer sphere visibility in disp3D
  • Fixed MNE Scan record action label/tooltip toggle
  • Fixed stale static-analysis findings
  • Fixed init.sh index error (#994)
  • Fixed segfault when opening FIF files saved from MNE Python (gantry angle type mismatch, #993)
  • Fixed cortical_surface plugin DLL resolution on Windows
  • Fixed connectivity registry header paths after library rename
  • Fixed WASM SerialPort, Doxygen MathJax, and deploy auth
  • Fixed WASM file dialog guard for MRI loading (onLoadMri)
  • Fixed Polhemus serial port disconnect, vertex-based frame flip, calibration flow
  • Fixed session restore: group existence check for nested prefix, repopulate AcquiredPoints
  • Fixed API docs: escaped operator[] headings causing broken Markdown links

[2.2.1] - 2026-04-28

Bug Fixes

  • mne_inspect — STC scrolling broken on desktop (regression introduced in 2.2.0). Scrubbing the STC time-slider froze playback, produced flicker / "stretched-triangle" geometry artefacts, and could crash the application after a few seconds.
    • Root cause: commit b58c18a8c2 ("Energy/computational savings…") switched the per-surface vertex buffer from QRhiBuffer::Immutable to QRhiBuffer::Dynamic to enable in-place colour updates. Dynamic is documented for small, frequently-updated payloads (UBOs, a few hundred bytes); using it for a multi-MB brain-surface VBO triggers two failure modes on Metal/Vulkan:
      1. Each in-flight frame slot keeps its own physical buffer. Skipping the per-frame updateDynamicBuffer() (because m_gpu->dirty was false) caused the GPU to sample the wrong slot — visible as stuck frames and wildly out-of-range vertex positions.
      2. Allocating a 4-5 MB Dynamic buffer × in-flight slots × every visible surface multiplied GPU memory pressure by roughly 3×, leading to allocation failures during interactive scrubbing.
    • The WebAssembly build was unaffected because the WASM merged-surface renderer uses Immutable buffers re-uploaded every frame.
    • Fix: revert just the desktop VBO type to QRhiBuffer::Immutable (re-created when vertex data is dirty). All other improvements from b58c18a8c2 (demand-driven render, cached vertex count, IBO upload gating, removal of redundant updateBuffers in renderSurface) are kept.
    • Files: src/libraries/disp3D/renderable/brainsurface.cpp.

[2.2.0] - 2026-04-26

Highlights

  • MNX as MNE Scan project storage — Pipeline configurations now load and save as .mna (JSON) or .mnx (CBOR) files via MnaIO; legacy XML pipeline format removed. Default auto-save path is default.mna.
  • Unified File-menu naming across applications — mne_scan, mne_inspect, and mne_browse now use the same "Open Project… / Save Project… / Export Project as MNX…" pattern.
  • MNE-C CLI parity (82 tools) — every command-line tool from MNE-C has a mne-cpp counterpart; final 4 ports landed in this release (mne_make_movie, mne_convert_lspcov, mne_convert_ncov, mne_dacq_annotator).
  • Lossless MNA / MNX round-trip — every MNA struct preserves unknown JSON/CBOR keys via an extras field, so newer schema fields survive a round-trip through older software.
  • Performance as a co-equal optimization goal — periodic code-optimization prompt and matching optimization-requirements.md doc cover both idiomatic-code refactors and quantitative perf wins (≥ 5 % wall-clock).

New Features

  • MNA graph-execution library (mne_mna): New library for the .mna analysis node-graph format with YAML-based operator definitions, topological execution, and MnaOpRegistry for built-in and user-registered operators
  • MNX binary container — CBOR-based binary equivalent of .mna (MNX1 magic header), with embedded-data support and identical key structure for lossless format conversion
  • MNA Scan integration — MnaGraph m_pPipelineGraph mirrors the live plugin scene in PluginSceneManager; MnaDataKind ↔ ConnectorDataType mapping shipped in mna_scan_types; PluginGui::loadConfig/saveConfig route through MnaIO::read/MnaIO::write
  • Stream-mode MnaGraphExecutor — MnaGraphExecutor::startStream() / stopStream() with StreamContext for real-time acquisition pipelines
  • Machine-learning library (mne_ml): New library with MlTrainer for sklearn-compatible model training via embedded Python, supporting SVM, Random Forest, Logistic Regression, and custom estimators
  • STS covariance estimators: Added Ledoit-Wolf and Oracle Approximating Shrinkage (OAS) estimators to the sts library with cross-validation support against scikit-learn
  • Source-space cluster permutation tests — StatsCluster::oneSamplePermutationTest(), fTestPermutationTest(), tfce() (threshold-free cluster enhancement); StatsAdjacency::fromSourceSpaceTemporal() for spatio-temporal adjacency
  • Spectral connectivity expansion — GrangerCausality ("GC"), DirectedTransferFunction ("DTF"), PartialDirectedCoherence ("PDC"); shared MVAR fitting via MvarModel
  • CMNE sparse inverse: Added Cascaded MNE (InvCMNE) solver to the inverse library for spatially sparse source estimation
  • DSP Infomax ICA: Extended the dsp library with an Infomax-based ICA decomposition algorithm
  • BIDS coordinate system: Added BidsCoordinateSystem class for round-trip parsing and writing of coordinatesystem.json, and BidsElectrode::toFiffDigPoints() for converting BIDS electrodes to FIFF digitizer point sets
  • FreeSurfer atlas lookup: Added FsAtlasLookup for Desikan-Killiany / Destrieux atlas region querying
  • MRI slicer: Added MriSlicer class for extracting orthogonal slices (Axial/Sagittal/Coronal) from NIfTI volumes
  • 3D electrode visualization: ElectrodeObject now includes QRhi GPU buffer management (updateBuffers, vertexBuffer, indexBuffer, instanceBuffer) following the BrainSurface pattern; new electrode.vert / electrode.frag shaders with instanced contact spheres and Blinn-Phong lighting
  • 3D MRI slice visualization: New SliceObject renderable for textured MRI volume slices with voxel-to-world transforms and intensity windowing; new slice.vert / slice.frag shaders
  • FiffAnnotations integration: AnnotationModel refactored to use FiffAnnotations as backing store instead of internal QVector<AnnotationEntry>, with fiffAnnotations() / setFiffAnnotations() public accessors
  • Python cross-validation tests: Added GUARD_PYTHON / GUARD_PYTHON_PACKAGE macros ensuring Python-dependent tests fail in CI (MNE_REQUIRE_PYTHON=true) instead of silently skipping
  • PythonTestHelper extensions: Added writeMatrix(), readMatrix(), and evalMatrixViaFile() convenience methods for Eigen↔NumPy round-trip validation
  • MNA Scan plugin registry: Added scan-plugins.json schema and MnaScanTypes class for declaring MNE Scan plugins as MNA operator schemas
  • W-file I/O: Added read_w and write_w methods to FiffStream for FreeSurfer .w format surface data
  • Cortical surface mapping: New MNECorticalMap class with icoDownsample() for icosahedral downsampling, smoothOperator() for nearest-neighbor surface smoothing, and computeMorphMaps() for cross-subject cortical morphing
  • Scalp surface generation: Added makeScalpSurfaces() to MNEBemSurface for creating decimated scalp meshes at multiple resolutions
  • Processing history: Added copyProcessingHistory() to FiffStream for transferring processing provenance between FIFF files

MNE Inspect

  • UI overhaul — Native QMenuBar (File / View / Tools / Help) with macOS system-menu integration, Recent Projects submenu (max 10), Ctrl+O / Ctrl+Q / Space shortcuts; closable + movable Controls and Loaded Files dock widgets with custom flat title bars; status bar with progress; saveState() / restoreState() persistence
  • Loaded Files panel — live QTreeWidget with Name / Type / Path columns and context menu (Remove, Copy Path, Show in Finder); double-click STC entry activates it in the source-estimate combo
  • Progressive-disclosure controls — load buttons removed from sidebar; menu actions are the single entry point; control groups disabled until matching data loads; data unload via Loaded Files context menu
  • WASM compatibility — menu bar renders inline; QFileDialog guarded for WASM; recent-files menu disabled on WASM

MNE Scan

  • File menu uses Project terminology — &New Project (Ctrl+N), &Open Project… (Ctrl+O, accepts *.mna *.mnx), &Save Project… (Ctrl+S, defaults .mna), &Export Project as MNX… (Ctrl+Shift+E, always binary CBOR)
  • Auto-save default path — default.xml → default.mna (auto-loaded on startup if present)
  • Foreign-data preservation — round-trip saves preserve unknown fields via m_loadedMnaProject
  • Modernised pipeline canvas — rounded-rect nodes with gradient fills, drop shadows, type labels (SENSOR / ALGORITHM), centred name, hollow-input / filled-output ports, Bezier connections with tangent arrowheads, dot-grid background, Tailwind-inspired palette, selection glow

MNE Browse

  • Project reset on standalone open — opening a raw file without a project resets m_mnxProject to prevent stale-project contamination across all 5 file-open entry points (desktop menu, WASM picker, recent files, drag-drop, CLI)

WASM Progressive Web App

  • PWA manifest + service worker — scripts/wasm/manifest.json, scripts/wasm/sw.js, 192×192 / 512×512 icons; offline caching with versioned cache keys; combined COOP/COEP header injection
  • No server required — deploy bundle to any static host; service worker injects required headers for SharedArrayBuffer

Command-Line Tools

  • Final MNE-C CLI parity (4 new): mne_make_movie (STC frames → PNG via FsSurface orthographic projection), mne_convert_lspcov (LISP S-expression covariance → FIFF FiffCov), mne_convert_ncov (ASCII ncov → FIFF), mne_dacq_annotator (event annotation replacing legacy Motif GUI)
  • Inverse tools (7 new): mne_compute_cmne, mne_inverse_pipeline, mne_label_ssp, mne_average_estimates, mne_process_stc, mne_make_uniform_stc, mne_map_data
  • Preprocessing tools (7 new): mne_add_triggers, mne_change_baselines, mne_change_nave, mne_copy_processing_history, mne_fix_stim14, mne_make_derivations, mne_toggle_skips
  • Info tools (6 new): mne_check_eeg_locations, mne_evoked_data_summary, mne_list_coil_def, mne_list_proj, mne_sensor_locations, mne_show_mna
  • Forward tools (3 new): mne_fit_sphere_to_surf, mne_make_scalp_surfaces, mne_transform_points
  • Conversion tools (3 new): mne_convert_ctf_markers, mne_epochs2mat, mne_mna_bids_converter
  • Total CLI inventory: 82 tools — full MNE-C feature parity

CI/CD

  • Upgraded codecov/codecov-action from v5 to v6 to resolve Node.js 20 deprecation warnings
  • Added MNE_REQUIRE_PYTHON: 'true' environment variable to all platform test steps
  • Added scikit-learn to Python test dependencies
  • Updated lodash from 4.17.23 to 4.18.1 (Dependabot security fix)

Bug Fixes

  • Fixed potential silent test skips — Python-dependent tests now respect MNE_REQUIRE_PYTHON enforcement
  • Fixed deprecated qWarn calls in test_mne_source_analysis (replaced with qWarning())
  • Fixed unused [[nodiscard]] return values in test_tool_inverse_computation

Documentation

  • Added 26 new command-line tool documentation pages
  • Comprehensive MNA / MNX format manual page (mna-format.mdx, ~692 lines) — file types, BIDS-like data hierarchy, computational graph, execution modes, parameter system, op registry, verification & provenance, MNX byte layout, schema versioning, complete annotated JSON example
  • Rewrote MNA API reference page (api-mna.mdx) with full class inventory (63 → 572 lines)
  • Added ML library API reference page (api-ml.mdx)
  • Added Statistics library API reference page (api-sts.mdx)
  • Updated tools overview with all 26 new tools and new "Inspection and Information" section
  • Updated overview and API pages with Mna, Ml, Sts library entries
  • Expanded glossary with 14 new terms (CMNE, DICS, LCMV, MNA, MxNE, ONNX, etc.) and 3 file format entries (.mna, .mnx, .onnx)
  • Fixed orphaned pages (mna-format, api-mna) by adding to sidebar navigation
  • Added zero-tolerance warning policy to development conventions
  • Added mne_analyze feature-parity gap analysis (Section 14 of gap-analysis.md); concrete plugin-level tasks moved to optimization-requirements.md § 15
  • Added reusable code-optimization Copilot prompt (.github/prompts/code-optimization.prompt.md) and matching optimization-requirements.md backlog with explicit performance scope

[2.1.0] - 2026-04-10

New Features

  • MNE Browse overhaul

    • QRhi-based 2D channel rendering for Metal/Vulkan/D3D12 performance
    • ICA browser with component visualization, rejection, and original-data reset
    • Epoch review workflow with evoked computation and histogram view
    • Annotation workflow with FIF interoperability and auto-merge
    • Virtual channel derivations with weighted reference workflows
    • Covariance matrix heatmap visualization and whitening controls (W key)
    • Projector dialog (J/Shift+J) and SSP toggle
    • 2D layout view with mouse-wheel zoom, auto-scale, and channel scaling spinboxes
    • Dark mode, drag-and-drop file open, recent files list
    • GPU clipping/z-score shaders, fullscreen (F11), zen mode (F)
    • Linear detrending (D key cycles None/DC/Linear)
    • GFP overlay on butterfly view
    • Interactive ruler tool with amplitude snap and semi-transparent overlay
    • Double-click bad-channel marking, click-to-toggle bad channels
    • Stimulus lane above time ruler with event markers
    • Per-description event/annotation filtering and visibility toggles
    • Channel sort by type, epoch grid lines (G), overview bar toggle (O), scroll speed control
    • Source-estimate export workflow
    • Session filtering workflow
    • Settings persistence across sessions via QSettings
    • Event file saving to FIF and ASCII formats
    • WASM loading screen and optimized WebAssembly build

  • MNE Analyze Studio

    • Agent-based workflow engine with grounded LLM safety
    • Multi-provider planner profiles (OpenAI Responses API)
    • Reusable inspect 3D surface views
    • Extension-based view architecture with hosted view tooling
    • VS Code chat UX integration and .mna project format

  • BIDS library (mne_bids): New library for reading and writing BIDS-formatted datasets with testframe

  • Beamformers: Added beamformer classes to the inv library with testframes

  • DSP library enhancements

    • WelchPsd: Welch power spectral density estimation
    • MorletTfr: Morlet wavelet time-frequency representation
    • xDAWN: xDAWN spatial filtering for ERP denoising
    • KMeans: Cosine and correlation distance normalization implemented
    • Real-time covariance: Channel picks applied (MEG/EEG only)

  • Logging: Renamed ApplicationLogger to MNELogger with MNE_LOG_LEVEL and MNE_LOG_MODE environment variables

CI/CD

  • Qt 6.11 support with auto-detection in init.sh
  • Build Qt Installer Framework 4.11.0 toolchains from source
  • Offline self-contained installers (.run, .dmg, .exe)
  • Nightly WASM builds deployed alongside stable releases
  • Reusable Qt toolchain assets consumed by CI workflows
  • Eigen downloaded at configure time (no longer bundled)

Bug Fixes

  • Fixed Metal/QRhi rendering on macOS: native window, infinite repaint loop
  • Fixed butterfly whitening scale and circular channel arrangement
  • Fixed 2D layout channel filtering, scale persistence, and item sizing
  • Fixed crosshair repaint cascade, split MEG grad/mag scale labels
  • Fixed DSP SSS and bad-channel regressions
  • Fixed DSP pi constants on Windows, params defaults on Clang/GCC
  • Fixed Windows linker error for nested Params struct exports
  • Fixed macOS test flakiness and WASM build errors
  • Fixed mne_anonymize unique_ptr usage
  • Fixed ruler amplitude precision (1 → 3 decimal places)
  • Fixed stim chip overlap, event line z-order, and resize resampling
  • Fixed scroll pan boundary and smooth inertial scroll

Documentation

  • Updated website to reflect renamed/split libraries
  • Updated build guide paths (tools/ → scripts/)
  • Fixed release download page for installers and WebAssembly
  • Updated Zenodo DOI to 19139238

[2.0.0] - 2026-03-20

Versioning

  • MNE-CPP libraries: Version 2.0.0
  • CLI tools (51 ported MNE-C tools, mne_dipole_fit, mne_inspect): Derive version from the MNE-CPP project version (2.0.0) via CMake compile definition
  • MNE Scan: Version 1.0.0 (first stable release as a standalone application)
  • MNE Analyze: Version 1.0.0 (first stable release as a standalone application)
  • IConnector plugin interface: Bumped to mne_rt_server/2.0

Breaking Changes

  • Library renames: connectivity → conn, communication → com, rtprocessing → dsp, inverse → inv, disp3D_rhi → disp3D
  • Class renames: Inverse library classes now use Inv prefix (InvRapMusic, InvMinimumNorm, InvSourceEstimate, InvHpiFit, …)
  • Moved classes: InvInverseOperator, InvMeasData, InvMeasDataSet moved from inv to mne library (as MNEInverseOperator, MNEMeasData, MNEMeasDataSet); backward-compatible forwarding headers provided
  • Moved classes: FwdForwardSolution moved to MNEForwardSolution in the mne library; forwarding header in fwd/ for backward compatibility
  • Removed: InvCorSourceEstimate class (superseded by unified InvSourceEstimate)
  • Split: MNEMath split into Linalg and Numerics in the math library
  • Moved: events library moved into mne_analyze application (no longer a standalone library)
  • Qt requirement: Qt 6.10.0 minimum (Qt 5 no longer supported)
  • Eigen upgrade: Eigen 5.0.1

New Features

  • 51 CLI tools: Ported all MNE-C command-line tools to modern C++ with QCommandLineParser and ApplicationLogger
    • Conversion: mne_edf2fiff, mne_ctf2fiff, mne_kit2fiff, mne_brain_vision2fiff, and more
    • Forward modeling: mne_forward_solution, mne_setup_forward_model, mne_prepare_bem_model, mne_make_source_space, mne_make_sphere_bem, mne_average_forward_solutions
    • Inverse: mne_compute_raw_inverse, mne_compute_mne, mne_inverse_operator
    • Preprocessing: mne_anonymize, mne_process_raw, mne_compensate_data
    • Surface: mne_surf2bem, mne_check_surface, mne_watershed_bem, mne_flash_bem
    • Info: mne_show_fiff, mne_compare_fif_files, mne_list_source_space, mne_list_bem
    • Server: mne_rt_server
    • Simulation: mne_simulate_data
  • Unified InvSourceEstimate: Single class with composition layers for grid data, focal dipoles, couplings, connectivity, positions, and metadata; includes tokenization support for streaming
  • RHI-based 3D rendering: Migrated from Qt3D to QRhi pipeline with holographic rendering mode, Fresnel effects, dipole overlay, and async STC loading
  • Flexible clustering: Variable cluster sizes in source space computation
  • MNE Scan improvements: Plugin loading progress on splash screen, macOS app bundle icons
  • LSL library modernized: Added lsl_ prefix, comprehensive tests, updated to LSL 1.17.5

Refactoring & Modernization

  • Forward library (fwd): Removed ~700 lines of dead code; replaced printf → qInfo/qWarning; replaced goto → early return/lambdas; replaced C macros (ARSINH, EPS) → constexpr/inline; modernized FwdBemModel members from raw pointers to Eigen types and std::shared_ptr
  • FIFF I/O consolidation: Moved duplicated FIFF reading functions to library classes (FiffStream::read_bad_channels, MNECovMatrix::read, MNERawInfo::find_meas_info)
  • Thread safety: Fixed thread-unsafe static float* in MNEProjOp::project_vector using thread_local
  • Network input validation: Added bounds checking in BabyMEG client byte conversion functions
  • Sparse matrix crash fix: Guarded MNEHemisphere::operator== against segfault on default-constructed sparse matrices
  • Copyright: Updated to 2026 across all library headers
  • Qt6 serialization: Updated QDataStream version from Qt_5_1 to Qt_6_0 in rt_server
  • Centralized CLI versioning: All 51 CLI tools now derive their PROGRAM_VERSION from the CMake project version (MNE_CPP_VERSION) instead of hardcoded strings
  • Format specifier fix: Fixed %d → %lld for qsizetype in BabyMEG client warnings

CI/CD

  • Restructured to Main-Staging-Feature branch model
  • Build matrix: Ubuntu 24.04, macOS 26, Windows 2025 with Qt 6.10.0 and Qt 6.10.2
  • CodeQL and Coverity static analysis (weekly)
  • Code coverage tracking with 35% library threshold
  • CPackIFW-based installer generation (.run, .dmg, .exe)
  • WebAssembly build support

Documentation

  • Docusaurus-based documentation site
  • Organized CLI tools by domain category in sidebar
  • Added MNE handbook reference pages (cookbook, FIFF format, sample dataset, data conversion)
  • Completed Doxygen documentation for InvRapMusic, InvMinimumNorm, InvDipole

Bug Fixes

  • Fixed crash in make_inverse_operator with non-surface-oriented forward solutions
  • Fixed segfault in forward computation
  • Fixed FIFF double I/O, make_dir EOF, write_cov packing, and source_cov kind
  • Fixed MNECTFCompDataSet copy constructor member initialization
  • Fixed coordinate transformation handling in 3D visualization
  • Fixed inflated brain alignment and surface projection tolerances
  • Fixed floating-point comparison tolerances in forward solution tests
  • Fixed sphere simplex test convergence