phase2: multi-task AMICA concat to lift k-factor (issue #33)

[neuromechanist]

Summary

Refines Phase 2 to fix the under-determined ICA on PR #31. ThePresent alone gives k = samples / n_chans² = 1.1; the resulting AMICA components collapse to single electrodes (visual inspection confirmed). Reference pipeline mitigates by concatenating multiple HBN tasks for the ICA training pool while keeping ThePresent as the analysis target.

This PR adds an IcaTasks option to phase2_amica (default = the four passive-viewing movie tasks: DespicableMe, DiaryOfAWimpyKid, FunwithFractals, ThePresent). Lifts k from 1.1 to ≈3.5 on the 100 Hz local data.

Refs epic #1, closes #33. Blocks PR #32 (Phase 3 derivatives) until merged.

What changes

• src/matlab/phase2_amica.m: IcaTasks opt. When multi-task: load all per-task Phase 1 .set files, restrict to channel intersection, pop_mergeset, train AMICA on merged, transplant weights onto Task=ThePresent EEG.
• src/matlab/+hbn/prepare_multitask_amica_input.m (new): per-subject multi-task pre-merge step. Regenerates missing Phase 1 .set per task via phase1_preprocess.
• src/matlab/+hbn/apply_ica_weights.m (new): copies icaweights/icasphere/icachansind/icawinv from training EEG onto target EEG; asserts channel-count match.
• src/matlab/+hbn/write_qa_amica_csv.m: schema gains ica_tasks, ica_samples, k_factor, common_channels.
• tests/matlab/test_phase2_smoke.m: passes IcaTasks="ThePresent" so the fixture-based smoke test stays on the single-task path (fixture has only ThePresent).
• scripts/run_phase2_three_subjects_multitask.m (new): production 3-subject runner.

Why this isn't a brief violation

CLAUDE.md "Never generalize to other HBN movies in this project (ThePresent only)" is about the analysis scope; the contrast, epoching, ERSP, and stats all stay ThePresent-only. The reference pipeline study_handy_scripts.m explicitly uses task_group = ["surroundSupp", "RestingState", "DespicableMe", "ThePresent", "FunwithFractals", "DiaryOfAWimpyKid"] for the ICA training pool and pop_mergeset before runamica17_nsg. We are doing the same, scoped to the four movie tasks for ecological consistency.

Local verification

test_phase2_smoke: OK (118 ICs, 118 dipoles, median RV 0.436)

(Single-task fallback path; the wiring smoke test only exercises one task.)

Test plan

• Single-task smoke green on the fixture.
• CI green on this branch.
• 3-subject multi-task AMICA run (in flight). Expected wall-time per subject: 1-2 h (≈4x ThePresent-only because merged data is ~11 min vs 3.5 min, and AMICA scales near-linearly in samples).
• Verify k_factor in qa_amica.csv lifted from 1.1 to ~3-4.
• eeg-qa-neuroscientist Phase 2 review: topographies look component-like (bilateral / dipolar), not single-electrode; median RV improved.
• Then unblock PR phase3: iclabel classification + non-brain IC flagging #32: re-run Phase 3 on the better weights.

Status

Draft. PR body will be updated with derivatives + QA review once the 3-subject run completes.