Resources created with count are tracked in state by positional index, as aws_instance.web[0], [1], and so on. Removing an entry from the middle of the list shifts every subsequent element down one index. Terraform sees the resource at each shifted index as having changed attributes and plans destroy-and-recreate for all of them.
The plan output makes this visible but easy to skim past, because the diff for each shifted resource looks like an ordinary in-place-impossible change rather than a warning.
Use for_each with a map or a set of strings instead. Keys are stable identifiers rather than positions, so removing one entry affects only that resource. for_each requires values known at plan time, which is why a set derived from another resource's output sometimes forces count; in that case pin the ordering or restructure so the keys come from configuration. To recover from an already-shifted state without destroying anything, use moved blocks, available since Terraform 1.1, which record the rename in configuration rather than requiring terraform state mv commands run by hand.