Altitude gain in splits is wildly different from elevation graph

Good call.

But do you know for sure that the split table uses GPS+DEM? That sounds rather insane to me. The day before this session I was walking along a narrow path with an almost vertical drop of 30-50 meter a few cm to my right, and a less steep wall to my left. I would not trust DEM to give any useful altitude under those circumstances, because a very small 2D error in my GPS position or in the DEM map would cause my DEM altitude to be wildly wrong. 

Anyway, I enabled elevation corrections. This of course made the elevation curve much more correct. But it also changed the split table a bit. Here are some before and after screenshots:

Before:

After:

Can you enable the "Elev Corrections" on the right below the image of your watch. I'm curious if this only changes the Altitude graph or if it also changes the data in the table

That should make it lock to the ground level, which I guess should fix the problem. You say "The day before this session I was walking along a narrow path with an almost vertical drop of 30-50 meter a few cm to my right, and a less steep wall to my left. I would not trust DEM to give any useful altitude under those circumstances, because a very small 2D error in my GPS position or in the DEM map would cause my DEM altitude to be wildly wrong. " which is right, but in this case, in a walk in a reasonably built up area (according to the map) my guess is that disabled the corrections and locking to the DEM should do it

One thought mate, do you calibrate the altimeter before this?

I know what elevation correction does, and that it can be a workaround. I also agree that DEM would not be a problem on the specific route I linked to.

However, my comment about DEM was towards the general problem. I maintain that using DEM for split elevation gain and loss would be (or is) crazy, because DEM is not reliable enough for that. I would trust the barometric altimeter more for that purpose, because relative elevation change over short time is exactly what a barometric altimeter does well.

Yes, I calibrated the altimeter. As you can see in the two graphs, my starting elevation is the same before and after elevation corrections.

Do you still have the activity on your watch?  If you look at the Elevation Plot there it should offer a proper representation of the recorded elevation.  The scale is hardly detailed enough to do any deep analysis but should at least give you a general idea of the slope profile and indicate high/low measurements.  That probably is what you expect to see in Connect but the back end uses GPS data as a blend and it looks like you might have experienced some cut outs that impact how the graph is displayed.

The altitude plot in the watch has the same error as the plot in Garmin Connect.

I'm stumped then!  Late for work so I'll try and add more detail tomorrow.

Not that it'll make any more sense.

I use UltraTrac a lot and that makes it easy to see how the elevation graph in Garmin Connect varies wildly from the actual splits.  Somehow connect validates that it has GPS coordinates as it renders the graph and so the resulting profile is a series of steps and it ignores the raw altimeter data.  This does NOT represent the actual course profile.

That would look something like this:

The actual split data follows:

The Elevation Plot on my watch would be a representation of the raw data and so certainly would be a visual of the actual climbs/descents.  Connect just 'blends' GPS points so that it flatlines points between GPS silence.  I thought that you might be experiencing a similar behavior since it looks like you have a flatline section (after the slope series you're questioning) but if you're reporting that the plot on the watch also looks wonky then I'm no help and my theory is irrelevant.  I'll have to stew on this a bit more!