Now that I've worked on a larger set of images, there are many things I can say...
First, it turned out that starting with just a partial import of the Lightroom catalog was a good idea: in this way I could find a serious problem that would have appeared in my keyword workflow, it actually did some damage, but limited and repairable from a backup. Then I found a workaround.
The problem is that apparently LR and C1 can co-operate in respect to keywords, in the sense that C1 can be instructed to auto-sync metadata (it writes XMP as soon as you change something in it; LR doesn't work in a fully automated way, but it can nevertheless do the same thing).
In practice, there's a problem if you use hierarchical keywords. If e.g. you have a > b > c > d and tag a photo with 'd', LR considers that a single keyword (in XMP you see it e.g. as 'a|b|c|d'). When C1 imports it, it keeps the hierarchy, but also tags the photo with all the path elements; in other words, you have five keywords: a > b > c > d, but also a, b, c, d. This is quite a mess: some hierarchies I use (taxonomies for animals and plants, geotags) have keywords with the same name, but different places in hierarchy. With my initial setup, the "flattened" keywords made their way back to Ligthroom and created some damage. After recovering from a backup, I just disabled the feature of C1 that automatically writes keywords, and everything now is ok. This means that C1 cannot be used in any way with keywords - a minor annoyance for those related to IQ features, such as colour tags and ratings, that requires looking at the post-processed image in C1, then switch to LR, find that image in the browser and tag it. But in the end it's not hard as it might seem.
As a way for extra safety, I used git (a Source Control Management tool) to track all my XMP files, so I can eventually detect unwanted changes in single files and quickly revert them. Git will stay only for a while, just the time I learn to master C1 to avoid operator errors.
Given that, I proceeded to import my whole collection of photos (about 40.000) in the C1 catalog. I'm post-processing the photos since past October in C1, and keep the old ones in LR. I'm occasionally "promoting" some old photo managed by LR to C1, in particular for those that didn't turn out in their best shape, because perhaps C1 can do a better job (in many cases, it does).
Having the whole catalog imported I can say that, yes, C1 is much faster in the general usage - after some time, longer than one hour, it seems to slow down and restarting it has a good effect. But in some specific case it happened in LR too (in a worse way). Exporting images (seven sizes for each photo, for publishing on my website) is _way_ faster; and the good thing is that I can do that in a single command, while in LR I had to launch seven different tasks. What C1 misses is the "publish folder" feature of LR, that kept track of which images were edited after the latest export, so I need to track them manually (too bad C1 also misses an extra simple feature, that is the capability of filtering collections by editing date: I could at least see the photos edited in the past hour or past day).
Given that, after the first quick work with a few hours in Bellinzona, I went three days in Haute-Provence and returned with 3x900=2700 photos to post-process. 3x means that I always shoot in bursts, to maximise chances of getting an image without motion blur. Furthermore, on the higher mountains there were a strong wind, that often was shaking me while shooting (and more often moving the branches of trees), thus many subjects have produced up to 9-15 shots. This just to say that one of the first steps in my workflow is picking the sharpest image in a small set. In LR I did it with the compare editor, that allows to see a pair of images even at 1:1 and panning the zoomed window in sync, to search and evaluate the parts with the critical details. Working through a dozen of such images means to have 12 pair comparisons, keeping at each pass the best one. Sometimes with some extra job later, because some shots are of the same sharpness and I want to keep all of them for then picking on some other criterium (e.g. if there were moving parts, such as animals, keeping their best placement).
In C1 everything is simpler - and faster! I can bring together up to 11 photos and zooming/panning in sync all of them. This means that I can immediately see the blurred ones and trash them, and then zoom back to a whole view and pick the best of the remainders. Something in the zoom in/out commands seems to require some more strokes than in LR, but it can be I'm just missing some keyboard accelerator. In any case, yesterday I processed a first batch of 700 images (down to 150) in a few hours, and it was much faster than I'd have done in LR. BTW, for the pixel peeping to be quick in LR, you need to have 1:1 previews (otherwise you have to wait every time quite a few seconds for seeing the pixels). 1:1 previews are quite slow to be generated (and I have a two years old MacBook Pro, 16GB and everything on the internal SSD). Fortunately they can be generated in batch, and I just do something else during the process. On the other hand, C1 doesn't have 1:1 previews, because it doesn't need them: it's much faster in getting to 1:1 on the screen.
For what concerns colours, I can confirm that I like much more the C1 preset profiles for my cameras than those in LR - it seems I can work with them instead of using generated profiles out of the ColorChecker. This is particularly true for the blue in the sky, that I absolutely didn't liked in LR and was better, but still something required some hue fix, with my custom profiles.
I just applied a default preset of -33 to contrast, because it seems that the common taste is about more saturated and deeper black than I like.
I have still some doubts with yellow rendering - in some cases foliage still turns out more brownish that it should be, but at this point I have to set up a test also to verify it's not a personal thing of my brain. In any case, selective colour correction in C1 is so powerful that I think it could be possible to fix this problem with a single channel patch, embedded in a custom style, rather than requiring a completely different profile.