You have mentioned the filmic highlight reconstruction several times. However, it is not really reconstruction:
- it does not work with raw data
- has no idea what pixels are clipped on the sensor
- instead of hard (sensor-level) clipping, it works with the white level set in filmic
- its purpose is not really to recover detail, but to smoothen the transition to white.
Darktable has recently received a rather powerful highlight recovery method, but even that cannot deal with the situation where large areas have all channels clipped (unless there are smooth gradients it can propagate into the blown area – but there will be no pixel-level detail, only smooth propagation of lightness and a colour. Check out this thread for (early) examples from Iain Fergusson, who came up with the methods, which were then later made available in darktable as the inpaint opposed and segmentation based methods.
In this photo, it would have had to re-synthesise texture, which it cannot do.