Convoluted hook testing infra

[lorenzwalthert]

Is your feature request related to a problem? Please describe.
Convoluted testing infra

Describe the solution you'd like

In general, I think run_test is very convoluted, since std_err (as documented in run_test + me reading the source code) can be

• a string to signal we expect a certain error message (asserted with matching std err against the expected error message in case).
• NA to signals we don't expect the file passed to the hook test to change (asserted with comparing file contents before and after).
• NULL (the default) as a shortcut to avoid typing out expect_success since usually when you expect something in std err, you probably expect the hook to fail. The signature for the reader's convenience:

  run_test(
    std_err = NULL,
    expect_success = is.null(std_err),
    ...
  )

I think to untangle this, we first have to fix run_test_impl() and reduce cognitive overload / multiple meaning on the arguments std_err and std_out :

• use std_err and std_out only for message matching
• leverage read_only argument recently added to decide if file content should be the same or not (instead of relying std_err = NA).
• This might be enough keep the std_err = NULL shortcut without being too confusing.

Describe alternatives you've considered

Not sleep anymore at night since testing infra is messed up 😸

Additional context

Came up in https://github.com/lorenzwalthert/precommit/pull/551/files/b8f72d26d158d86eee8aad887db05ea06f622399#r1541574615