Sequence of events:

 radeon_clear_tris() 
   - calls _mesa_Enable(GL_STENCIL_TEST);
     - Changes ctx->Stencil.Enabled
   - draws 
     - ctx->Stencil._Enabled is updated to 1
   - calls _mesa_PopAttrib();
      - Restores ctx.Stencil.Enabled (but not _Enabled)
      - calls _mesa_DrawBuffer()
         - calls radeon_draw_buffer
            - accesses ctx->Stencil._Enabled, which is not up-to-date
            - Enables stenciling on the hardware

So the hardware is left with stenciling on, and subsequent drawing vanishes.

The patch I'll attach fixes the problem, I'm not sure if it's the best way to handle it.