ShallowFBCSPNet split_first_layer=False is broken by unconditional CombinedConv init/remapping #1211

Summary

ShallowFBCSPNet(split_first_layer=False) still constructs the historical unsplit first layer as self.conv_time, but the post-CombinedConv initialization and legacy checkpoint mapping assume self.conv_time_spat exists unconditionally.

This makes the documented non-split path fail during construction, and legacy conv_time.* checkpoint keys are remapped to a module that does not exist.

Regression

In Braindecode v0.7.0, the unsplit path initialized self.conv_time directly:

init.xavier_uniform_(self.conv_time.weight, gain=1)
if self.split_first_layer or (not self.batch_norm):
    init.constant_(self.conv_time.bias, 0)

The current implementation creates conv_time_spat = CombinedConv(...) only when split_first_layer=True, but later does:

init.xavier_uniform_(self.conv_time_spat.conv_time.weight, gain=1)

regardless of the mode. The current compatibility mapping similarly rewrites conv_time.* -> conv_time_spat.conv_time.* even for the unsplit model.

Expected behavior

• split_first_layer=True: keep the current CombinedConv initialization and legacy key migration.
• split_first_layer=False: initialize the direct conv_time layer and leave legacy conv_time.* keys on that module.
• classifier key migration remains unchanged in both modes.

Proposed regression coverage

1. construct and run ShallowFBCSPNet(..., split_first_layer=False);
2. verify the unsplit model owns conv_time and not conv_time_spat;
3. load a legacy-style state dict where conv_time.* remains unsplit while conv_classifier.* migrates to final_layer.conv_classifier.*.

I have a focused patch prepared that only branches the CombinedConv-specific initialization/remapping and leaves the default split-first-layer behavior unchanged.