1) Document the behavior of "typeset -n existing_var" (via Jun T. comment)
2) Prohibit "typeset -nm pattern" because, well, it's insane. Add test.
3) Improve doc for ${(!)ref} including ${{t!)ref} (Jun T.)
4) Fix doc for how-to unset of a named ref (Jun T.)
5) Allow "typeset +r -n ref" and "typeset +r +n ref" (Jun T.)
6) Fix "typeset -r -n ref=param" to create readonly references
7) Avoid accidental removal of PM_UNSET flag (Jun T.) and update test
8) Fix "typeset -gn ref=value" and add a test for it
9) Add tests for read-only reference behavior
10) Fix infinite recursion when resolving scope of an unset local
named reference, add test.
This solves the following problems in the _prefix completer:
- The old code had logic for dealing with compstate[unambiguous] that
was unnecessary. It works fine without it.
- Because of this logic, if a widget set compstate[insert]=1 after
calling _main_complete, an `x` was left after the completion on the
command line.
- If the same widget also set `compstate[to_end]=`, then instead, the
last character of the inserted completion would be treated as an
autoremovable suffix, with the actual suffix being inserted to the
line as a normal character.
- After inserting a completion, the cursor would move to the end of the
entire current word on the command, not the end of word that was
inserted. This is not what you want with _prefix, since you are trying
to complete a word _before_ the one on the command line, after which
you usually want to insert a separator, such as a space or slash,
before the next word.
When completing inside ~[...] (_with_ the trailing `]` present), the
following bugs occured:
- Subscript completion was skipped entirely when there were one or more
slashes ('/') in the subscript, which is incorrect, since slashes are
allowed there.
- Instead of going through _complete, $_comps[-subscript-] was called
immediately, causing _setup to be skipped.
- If succesful, _main_complete was exited right after, causing
menu-style, comppostfuncs and other essential completion features to
be skipped.
This is a special case where TRAPEXIT is unset within a TRAPEXIT
as it should never run in a nested context, so just save the
function structure temporarily on the heap.
A bug with zmodload when unloading/reloading a static module caused the
state of the shell options to change during K01 test. Worked around it.
Also changed warnnestedvar messages to look more like other such.
This affects CSI sequences that aren't explicitly bound but arrive
within the usual KEYTIMEOUT time limits. A single undefined-key widget
is run instead of unintended bindings for Escape and other characters in
the sequence.