I am honestly not even sure if the proposal is technically feasible, due to potential problems modifying the template contents before processing template codes like IF, and possible problems where modifications to the text might be flagged as false positives (that might be impossible to get around) in some of the template security measures.
As for performance, it's not a question solely of template depth. There is actually a hard limit of 15 templates depth before it starts rendering deeper templates as empty. But you can place 100 or 1000 template uses at the same depth (think of the main page text referencing various templates 1000 times -- I have actually seen some customers do this), so you not only need to process the main page text for auto-links, but then each of those 1000 templates. Then think of that happening at each level 15 layers deep in each of the 1000 templates. You can potentially call the autolinker over 1,000,000,000,000,000,000,000,000,000,000,000,000,000,000 ,000 times. Some optimizations would also need to be conceived for templates to make this harder. This is a general problem with templates, and a reason why there are so many security measures in place around them.
But I have debugged some sites where the auto-linker takes upwards of 1 second to run once. It would be a small task for even unintentional wiki edits to cause massive numbers of wiki pages to go down because the autolinker has pushed them to trigger those security measures.
Because of these issues, the proposal could not be just #1, but probably a combination of #1+2. But there would probably need to be a third element in there that limits the number of templates on a given depth that are processed.