shortanchor compression is lossy

When the Sphinx search module encodes the search index, it uses a compression scheme for anchor names. Unfortunately this compression scheme:

https://github.com/sphinx-doc/sphinx/blob/c7e7cd0ebf4ed5ced75eb0511b34b837d487ab88/sphinx/search/__init__.py#L326

Unfortunately this compression scheme is lossy for a few reasons:

1. The compression is based on comparing the HTML-escaped fullname to the anchor name. However, the fullname is never stored in the search index, only the dispname. The JavaScript search implementation therefore has to decode the anchor name based on the dispname, rather than the fullname, and if they differ then this would lead to the wrong result. For some domains, like Python, the dispname is always equal to the fullname, but for other domains, like std, that is not always the case.
2. The compression uses the special values of "" and "-" to encode compressed values. However, that means that it cannot encode an actual anchor equal to "" or "-". For example, I have some Sphinx customizations that use a separate page for each object, and therefore have an empty string as the anchor name.

Start at sphinx/search/init.py around the linked compression code, then trace the JavaScript search decoding that consumes the index. Reproduce cases where dispname differs from fullname and where the anchor is empty or "-", and verify that the revised encoding and decoding preserve each anchor.