Found by running the regression tests.
Huh, where? The GL tests pass for me with master...
d3d10core, but it's not unique to that. Note that the issue won't happen when using core contexts, due to the check for WINED3D_GL_LEGACY_CONTEXT in shader_glsl_update_legacy_states().
The intended solution here (which may not be the best one, I don't know) is that any draw that uses a stateblock should have set these constants. This is done by invalidating stateblock->changed.ffp_ps_constants when the primary stateblock is created [wined3d_stateblock_invalidate_initial_states()], so that those constants are pushed when wined3d_device_apply_stateblock() is called even if no states were ever set by the application.
The trouble with that approach is that glsl_fragment_pipe_alpha_test_func() also gets called for draws that don't use stateblocks. Specifically: shader_glsl_apply_draw_state() -> shader_glsl_update_legacy_states() -> glsl_fragment_pipe_alpha_test_func(), because of WINED3D_SHADER_TYPE_PIXEL in "shader_update_mask".