Just to be sure: I did a similar test on a test page with a test text, and everything worked as expected : edit conflict detected and I couldn't save my modification.

Then, I tried the same test with the same text as the problem I reported : no edit conflict detected, and intermediate version ignored :

• Starting point
• Edit through the wikitext editor to reflect Bloubéry edit
• Edit through WPCleaner to reflect my edit : intermediate version silently ignored

I'm not seeing the problem here. It looks like it correctly automatically merged your bot edit with the intervening edit.

If it was an auto-merge, shouldn't the "Liens externes" section not have been added back? Correctly auto-merging the intermediate link removal change should have resulted in the bot edit just being the category/defaultsort change.

Is it possible the bot did get an edit conflict but re-submitted the same text back, causing the link to come back?

Have you tested the same thing except using the wikitext editor in a browser for all steps instead of wpcleaner?

I can confirm that the didn't get an edit conflict, so didn't try to submit the same text again.

The first editor was surprised to see the text he deleted being added again, and I fully understand his surprise, as I got the same feeling.

For me, it's an edge case of the conflict resolution algorithm. I hope it can be tweaked to prevent this.
The problem is probably that it didn't see that a section deleted in the first edit was kept in the second edit (for me, I wasn't moving the section, but the portal and categories)

In T117312#1774638, @aaron wrote:

If it was an auto-merge, shouldn't the "Liens externes" section not have been added back? Correctly auto-merging the intermediate link removal change should have resulted in the bot edit just being the category/defaultsort change.

The bot edit was apparently also moving the "Liens externes" text above the categories (or, equivalently, moving the categories to after the "Liens externes" text, but the diff-merge algorithm seems to have picked it up as the former) and deleting the blank line between the header and the single line of content. The "remove" part of the move was identical to the intervening edit and so merged fine, while the "add it elsewhere" part was to a different part of the article and so didn't prevent a merge either.

It would be difficult for the merge algorithm to know the intention of the intervening edit was to remove the text completely and to know the bot edit would be seen as adding it back, slightly differently and in a different place.

Right, that makes sense.

One could play around with the algorithms and how much they tend to count similar text moved around as possibly the same or totally new, though that's complex and has it's own tradeoffs in terms of thinking edits conflict that don't.