data_analysis/ml: refresh classifier results on Phase 1 (11 subjects)

[AlanYWu]

Summary

Re-trains v1 (band power), v2 (time-freq), v3 (EEGNet) MI pipelines and the blink detector across the current Phase 1 subject pool (sub02–sub12, 11 subjects) and regenerates data_analysis/ml/results.md.

The previous results.md described only sub02–sub05 (4 subjects). This PR brings results in line with the data we now have on Drive.

What changed

• Sample size: 4 → 11 subjects, 77 → 171 MI epochs, 190 → 519 blink epochs.
• MI still hovers near chance — best within-subject is now LDA / temporal_only at 56.7%; best LOSO is RF on v3 EEGNet / all_4ch at 59.4%. The wider pool brings the previous best (61.0% within-subject) down — likely the small-sample number was lucky.
• Blink stays strong: 91.3% within-subject (LDA all_4ch) and 79.8% LOSO LDA — the threshold baseline now narrowly beats LDA on LOSO (80.7%).
• Channel quality table now covers all 11 subjects; auto-channel selector still picks TP9, TP10.

Headline (best-of)

Task	Best within-subject	Best LOSO
MI v1	LDA / temporal_only — 56.7%	SVM-rbf / temporal_only — 57.3%
MI v2	LDA / temporal_only — 55.7%	RF / all_4ch — 56.3%
MI v3	LDA / temporal_only — 54.0%	RF / all_4ch — 59.4%
Blink	LDA / all_4ch — 91.3%	LDA / all_4ch — 79.8% (Threshold 80.7%)

Files

• `data_analysis/ml/refresh_results.py` — regeneration script (v1/v2/v3 + blink, 5-fold within-subject CV + LOSO, renders the markdown)
• `data_analysis/ml/results.md` — refreshed tables + per-subject breakdowns
• `data_analysis/ml/results_data.json` — raw means + per-subject scores

Test plan

• `python data_analysis/ml/refresh_results.py` reproduces both files.
• Reviewer sanity-checks the per-subject MI table — sub04 still tends to be the weakest, sub09–sub11 are noisy due to low MI epoch counts.