I can confirm, based on my personal experience, that coherent light (parallel rays) is not going to result in even illumination(unless you are using a telecentric lens I suppose, but that’s just an aside) and that you need to focus the condensing optic’s projection of the (point) light source into the optical center of the taking lens.
The condenser optics (focal length) should match the focal length of the camera lens, and its size (diameter) should be sufficient to ‘cover’ the negative; in an enlarger using undiffused illumination (‘collimation’) this would be by design.
The Focal length of the collimating lens must be >= the focal length of the taking lens and the the diameter must be such, that the entire FOV of the taking lens is covered by the cone of light converging in the taking lens. Greater focal length on the condensing optics is not an issue and will only result in slight loss of light over the entire frame which is a not issue compared to the amount of efficiency the system gains through a collimated illumination. I will add, that the light source, to my knowledge, should be no smaller than the aperture on the taking lens. This is not an issue as you can stop down without losing light if everything is setup correctly
For anyone stumbling upon this thread I will add, that the amount of retouching require to clean up the added dust and scratches(especially on aging film) is such, that I would personally advocate against a collimated light source. I recently built a system that has both and can switch between the light sources without moving the film or the camera, resulting in captures in perfect registration, with all aspects equal except for exposure time and illumination scheme. The difference cannot be overstated. I like having both, as it allows me to use large portions of the diffuse capture to reconstruct damages on the collimated capture, but building and calibrating such a system was so much effort, that I wouldn’t say is justified by the results. Here are some examples (~5900 DPI, using the Minolta 5400 lens onto a monochromatic area sensor, RGB+IR light source):
Especially when the emulsion has begun to crack with age, this effect can be quite severe:
