are you saying that Remove doesn't follow symlinks? (effectively RemoveRaw)

@Code0x58 What do you think?

The CI tests confirm the issue for Mac and Linux. I had the same problem with my solaris/illumos port.
Windows doesn't seem to care, but Windows also doesn't really have symlinks.

My immediate thought was that a Remove that fails could do an extra check to see if the requested path to remove is a symlink and do the same resolution and try again, but that should probably be implemented in fsnotify.go across all platforms rather than individually in each os-specific implementation.

Definitely curious to know what @Code0x58 thinks we should do.

Good spot and test!

For now, I think renaming Remove to RemoveRaw and making the Windows implementation consistent is the way to go, as well as adding a naive solution similar to Add (i.e. Remove tries to deference the symlinks). The naive Add and Remove would come with caveats to document - I'd probably recommend not using Add/Remove at all unless you know their caveats work in your use case. I'm not sure it would fix any existing users for users, aside from inconsistency with Windows/other implementations.

With a portable low level interface, and documentation on existing issues, the library would be in a better position for users, and may attract contributions towards documenting or implementing common patterns, and maybe even a generic high level interface. A vague comment on a high level interface is here, but it could probably use a lot of attention to avoid leading users in to problematic cases (with races/consistency). If the retry approach would stop a potential issue, and not introduce any more, then I'd say it's a good thing to do so that the nativity/caveats of Add/Remove is incrementally reduced.

ramble: I think the codebase would need a fair bit of fixing and ramping up in testing to offer a properly dependable cross platform solution for non-trivial watching (trivial such as watching a file in a directory which won't be moved or deleted). I say this following experience trying to fix (in #382) the hidden issues found by linting (usually where errors aren't checked in tests, or test objects are used in goroutines which shouldn't - so may not be working at all). That could be a quite a bit of a task taking a lot of attention, but documenting and testing those safe patterns would be good and attract a community. I imagine a fair few people would like a robust pattern for matching a single config file which may be a file, or a symlink to a file such as when you are making a program which may run in or out of a container, and the config may be a projection mount within a container (e.g. when using Docker, Kubernetes, etc.) - that was what inspired my original PR back in 2019.