Screens and most image file formats can only encode RGB intensities confined within a certain range. For example, images encoded on 8 bits can only contain values from 0 to 255, images on 10 bits from 0 to 1023, and so on… Graphic standards postulate that the maximum of that range, no matter its actual value, will always represent the maximum brightness that the display medium is able to render, usually between 100 and 160 Cd/m² (or nits) depending on the actual standard. We generally call this maximum “100 % display-referred”. The minimum of the range, encoded 0 no matter the bit-depth used, becomes then “0 % display-referred”. 100 % encodes pure white, 0 % encodes pure black.
This is a limitation for image processing applications, because it means that any pixel lying outside of this range will be clipped to the nearest bound, resulting in non-recoverable loss of data (colors and/or textures).
For the longest time, image processing software too was bounded to this limitation for technical reasons, and some still is, but now by design choice. As a result, they would clip RGB intensities at 100 % display-referred between image operations.
darktable uses floating-point arithmetic inside its color pipeline, which means it can handle any RGB value internally, even those outside the display-referred range, as long as it is positive. Only at the very end of the pipeline, before the image is saved to a file or sent to display, are the RGB values clipped if needed.
Pixels that can take values outside of the display range are said to have “unbounded colors”. One could choose to clamp (i.e. confine) those values to the allowed range at every processing step or choose to carry on with them, and clamp them only at the last step in the pipeline. However, it has been found that processing is less prone to artifacts if the unbounded colors are not clamped but treated just like any other color data.
At the end of the pipeline, modules like filmic rgb can help you to remap RGB values to the display-referred range while maximizing the data preservation and avoiding hard clipping, which is usually not visually pleasing.
However, at all times in the pipeline, you must ensure that you do not create negative RGB values. RGB intensities encode light emissions and negative light does not exist. Those modules that rely on a physical understanding of light to process pixels will fail if they encounter a non-physical light emission. For safety, negative RGB values are still clipped whenever they might make the algorithms fail, but the visual result might look degraded. Negative values can be produced when abusing the black level in exposure or the offset in color balance and care should be taken when using these modules.