[models] Add TMSA-Net - #1209

Exact-head validation is green on 3a62b2e6d94c498e9fa92df3767c4e98be33a4ce: all 13 TMSA-Net focused regressions pass, and 12 shared Braindecode integration cases pass on the same head. This complements the pinned reference-parity assay already in the PR body (45/45 state entries, zero logits/input-gradient error). Evidence run: https://github.com/lindicaphxag-tech/lindicaphxag-tech/actions/runs/37136260152. I’ll hold this head for maintainer review.

Current-head validation is now complete on 82b61e569ea9dcf0f90efd3eaf07ce05d1c1b128. Pinned-reference parity maps all 45/45 state entries and remains exact for both logits and input gradients (max_abs = 0.0), while preserving the released 19→16→19 attention bottleneck (head_dim=4, inner_dim=16): https://github.com/lindicaphxag-tech/lindicaphxag-tech/actions/runs/37189419406. The same current head also passes 13 focused TMSA-Net tests, 12 Braindecode integration cases, and all touched-file pre-commit hooks: https://github.com/lindicaphxag-tech/lindicaphxag-tech/actions/runs/37189437290. I’ll hold this head for maintainer review.

Codecov Report

❌ Patch coverage is 98.83721% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 89.20%. Comparing base (6abd308) to head (10444f9).
⚠️ Report is 1 commits behind head on master.

@@            Coverage Diff             @@
##           master    #1209      +/-   ##
==========================================
+ Coverage   89.02%   89.20%   +0.18%     
==========================================
  Files         157      159       +2     
  Lines       19398    19727     +329     
==========================================
+ Hits        17269    17598     +329     
  Misses       2129     2129              

Integration gate (braindecode maintainers)

Target: BCI-IV-2a (BNCI2014_001) within-subject accuracy (session-1 train / session-2 test), 9-subject mean. The paper reports 83.22% on 2a. Gate: within 5% of that (0.791-0.874). The released code's 2a preprocessing is private (newbcicIV2a/ .mat, no script), and the reported number selects on the test set over 2000 epochs with no fixed seed. So we replicate under a declared, labelled substitute protocol: MOABB 4-38 Hz/250 Hz/4 s windows, AdamW lr=2**-12, cosine LR (original's T_max-per-epoch quirk reproduced), batch 16, 2000 epochs, the original's S&R augmentation, and checkpoint = best-validation accuracy on a 20 % hold-out of session 1 (primary), plus a pre-registered last-epoch control. The paper's own best-test rule is shown for context only.

Already verified: numerical parity is exact (logits/input-gradient max-abs diff 0.0, 45/45 state entries, 20303 params), so the port is the same function as Whit3Zhao/TMSA-Net@c60882db. The current head only merges master; tmsanet.py is unchanged. One disclosed training-time note: the port's FeedForwardBlock reuse drops the original's second FFN dropout (eval-neutral).

TMSA-Net (#1209), BCIC-IV-2a, 9 subjects x 5 seeds, port vs original code (matched init, same cached windows):

Best-val, like-for-like on the 45 cells where both codes finished: port 0.6749 vs original 0.6781 (diff -0.0032).

Verdict (5% gate vs paper): PARTIAL: BCIC-IV-2a 9-subject mean (best-val, primary; 45/45 port, 45/45 original cells) port 0.6749 vs paper 0.8322 = -18.9% (outside 5%); like-for-like on the 45 cells where both codes finished: port 0.6749 vs original 0.6781 (diff -0.0032). Documented cause: declared substitute protocol. The paper's 2a preprocessing is a private .mat set (4-38 Hz MOABB windows substituted); the paper selects its number as the max TEST accuracy over 2000 epochs (we select on a 20% validation hold-out of session 1, which also removes 20% of the training data); its seed is random and unpublished.

No action needed from you. If the exact 2a preprocessing script or a fixed seed exists outside the published repo, it would let us tighten the substitute protocol.