Hi Arthur, I've been working on the mismatched names and, after comparing authors on articles in Wikidata with the Crossref metadata and checking the number of statements on the author item, I deprecated about 1000 ORCID claims. Some of these items are likely duplicates but with the ORCID deprecated it will be safe to merge them later. The check against Crossref has also revealed some cases where the author order from Pubmed is wrong, which doesn't make things any easier!
The only item I've knowingly sorted out from the bad merges is Q41173397 but there may also have been some overlap with the mismatched names. It should be possible to automate the checking for Krbot edits using the Revisions API.