[ICLabel] Complete viewprops diagnostics parity

[suraj-ranganath]

Add eeg_icalabelstat with EEGLAB-style threshold summaries, explicit standalone ICLabel network limits, and focused ICLabel/viewprops tests. Document the component review workflow and mark the Phase 6 parity rows implemented while preserving native EEGPrep diagnostic displays.

Limitations called out intentionally:

• Standalone Python ICLabel ships only the bundled default netICL.mat; lite and beta remain explicit MATLAB/Octave-engine paths unless their artifacts are packaged later.
• pop_iclabel intentionally has no dialog Help button because the ICLabel plugin inputdlg3 call does not pass a help callback; packaged pophelp text remains available.
• eeg_icalabelstat keeps EEGLAB console-style print output by default, with verbose=False and stream= available for non-console callers.

Closes #163

[claude]

Claude finished @suraj-ranganath's task in 3m 34s —— View job

---

Claude PR Review — #167 ICLabel viewprops diagnostics parity

• Read the diff against base
• Review eeg_icalabelstat.py (new function)
• Review iclabel.py network-limit changes
• Review pop_icflag.py / pop_iclabel.py changes
• Check EEGLAB parity
• Review tests
• Post review feedback

Overall this is a clean, well-tested addition. The standalone network-limit guard, the torch-import deferral, and the eeg_icalabelstat summary all read well and are covered by focused tests. I found one likely regression and a couple of smaller notes.

---

🔴 Likely regression: removing help_text hides the pop_iclabel Help button

pop_iclabel.py drops help_text="pophelp('pop_iclabel')" from the dialog spec. The Qt renderer gates the Help button on help_text being truthy:

# src/eegprep/functions/guifunc/qt.py:296
if spec.help_text and spec.show_help_button:
    help_button = QtWidgets.QPushButton(spec.help_label)
    ...

With help_text now None, the Help button no longer renders in the ICLabel dialog. Note the original value was not broken — pophelp._normalise_function_name already unwraps pophelp('pop_iclabel') → pop_iclabel (qt.py:1050 also falls back to spec.function_name), so help content resolved correctly before. The net effect of this change is simply that the button disappears, which is a parity loss vs. EEGLAB's dialog.

Two ways to resolve, depending on intent:

• Keep the button (recommended): restore a truthy help_text (e.g. help_text="pophelp('pop_iclabel')", matching every other pop_* dialog in the repo), or
• if the goal was to rely on function_name resolution, change the gate at qt.py:296 to (spec.help_text or spec.function_name) and spec.show_help_button so all dialogs benefit — but that's a broader change worth its own discussion.

Either way, please add a GUI test asserting the Help button is present for pop_iclabel, since nothing currently covers it.

Fix this →

---

🟡 Parity question: is eeg_icalabelstat an EEGLAB function or EEGPrep-native?

The ICLabel plugin isn't vendored under src/eegprep/eeglab/, so I couldn't diff against a MATLAB oracle. The name follows the eeg_* convention and the PR describes "EEGLAB-style threshold summaries," but I'm not aware of an eeg_icalabelstat.m in the upstream ICLabel plugin. If this is an EEGPrep-native diagnostic (which AGENTS.md permits), it'd be worth a one-line note in the module docstring/help saying so, rather than implying a 1:1 port — that keeps the parity matrix honest. If it does mirror an upstream function, please confirm the printed format matches (label text, capitalization, ordering) so it's a true port.

Minor on the format: the printed label uses IClabel class "..." — odd capitalization vs. the canonical ICLabel. Worth double-checking that's deliberate.

🟡 print() in a library function

eeg_icalabelstat._print_summary uses print(..., file=stream). AGENTS.md prefers logging over print outside scripts. This is borderline-acceptable as EEGLAB console-parity output, and the stream/verbose params make it controllable — just flagging for consistency. No change required if the console-mimicking behavior is intentional.

✅ Things that look good

• iclabel._normalize_algorithm correctly validates default/lite/beta and the NotImplementedError for non-default networks under engine=None fires before the torch import — good ordering, well tested (test_python_iclabel_rejects_unbundled_alternate_networks_before_runtime_dependencies).
• Deferring the torch import so MATLAB/Octave engines no longer require torch is a nice correctness fix.
• _threshold_vector scalar/per-class handling and the [0,1] validation are solid; class-name fallback to the 7 standard classes is covered.
• Calling eeg_icalabelstat in the pop_icflag GUI path is correctly guarded with not isinstance(EEG, list) and _require_iclabel.

---

I did not run the test suite or pre-commit in this review. Worth running uv run pytest tests/test_iclabel_statistics.py tests/test_gui_pop_iclabel.py plus a GUI smoke check of the Help button before merge.

[suraj-ranganath]

🤖 Phase 6 ICLabel/viewprops visual parity evidence

iclabel_pop_prop_extended_dashboard_side_by_side.png	pop_icflag_dialog_side_by_side.png

pop_iclabel_dialog_side_by_side.png	pop_viewprops_dialog_side_by_side.png