[seflue]

When an org file changes on disk while it is not open in a buffer, Neovim's memory keeps growing.

Such a file stays modified until it is reloaded, because only reload() updates the mtime that is_modified() compares against. Since #1204, a file whose buffer got unloaded ends up in the same state. _get_parser() takes "modified" to mean "the text changed", so every parse(true) on that file builds a new parser and tree from the unchanged self.content.

With notifications enabled, the check runs once a minute and never reloads, since files:load() returns early once the files are loaded. It reaches parse(true) through is_archived() and is_done(). Each new tree has a new root id, which clears the memoize cache, so the next headline parses again. So every check builds about two new trees per headline of such a file. The clock statusline does the same on every update while such a file is clocked in.

Two commits:

• The notification check reloads a file that is not open in a buffer before parsing it. The reload reads the new content and stores the new mtime, so the file is no longer modified and later parses return the existing tree. A file that cannot be read again keeps its old content, and the other files are still checked. Files open in a buffer are left out, because they use the buffer's parser.
• _get_parser() builds a new string parser only when the parser's source is not the current self.content. This covers every caller, not only notifications. A reload with new text gets a new parser, and so does a file whose buffer got unloaded, because that parser's source is a buffer number.

The second commit alone stops the growth, but only the first makes notifications use the new content from disk.

The reload still parses the changed file once, synchronously. Making that async would change every caller of reload(), so I left it out.