See branch daemonize-and-analyze-individual-files. This is incomplete, and makes assumptions about Phan which may be untrue.

EDIT: POC is done

It stops at the parse phase. When it gets a request, it updates the parsed files if they were changed on disk, removes old classes/functions/constants, and adds the new classes/functions/constants from new files.
Note that the daemon will use up a lot of memory in the background (Unlike the earlier sqlite3 work, this doesn't save state to disk). (As much memory as the parse phase uses)

pitfalls

• Undoing duplicate class/method/function declarations may have bugs, especially relating to the alternate fqsens (appending numbers to make them unique)
• symlinks
• Rate limiting, memory usage
• Memoizing stale data
• inheritance
• Files modified while being analyzed (e.g. in IDEs)

Further Enhancements

• use inotify PECL if working, and installed (Faster than checking set of whole files, only see changed files). Add optional dependency.
• Refactor this? (done)
• Add ability to reload parts of the config file, e.g. add new folders to analyze (Reload completely with a different process?)
• (Farther off) API to infer the type of the variable/property/method return value under the cursor of a saved file, somehow. use that for completions.

See the closed issue 22 in upstream

• Hopefully can resurrect vim plugin, with or without vim's new async support.