Fix syntax error in alias completion when alias contains a single quote

[Saadanjum0]

What

aliases.completion.bash builds a wrapper completion function for each alias by splicing the alias' command and arguments directly into the generated source as unquoted text, e.g.:

prec_word='$alias_cmd $alias_args'
...
COMP_LINE=${COMP_LINE/$alias_name/$alias_cmd $alias_args}

If the aliased command contains a single quote, this produces invalid (or in some cases silently wrong) generated code. For example:

alias foo="echo it's broken"

generates a snippet with an odd number of unescaped quotes, so sourcing it fails with something like:

bash: /tmp/alias_completion-XXXXXXXXXXX: line NN: unexpected EOF while looking for matching `''

which breaks the user's whole shell startup. This matches #2254, where a maintainer-adjacent contributor traced the problem to this same unquoted interpolation and confirmed it's still present on master.

Fix

Run alias_cmd and alias_args through printf %q before embedding them in the generated wrapper, the same way the existing code already does for alias_arg_words. This keeps the generated source valid regardless of what characters the aliased command/arguments contain.

Notes

• Behavior change: COMP_LINE used to be rebuilt by splicing in the raw alias text, so any $VAR in the alias got expanded (e.g. alias g2='cap -x "$HOME"' would show the expanded path in COMP_LINE); after this fix COMP_LINE keeps the literal, unexpanded alias text instead, which is arguably more correct but is a real behavior change worth flagging.
• Known remaining limitation (not a blocker): the while read line loop over alias -p output doesn't use -r, so '\'' sequences from some alias forms still get mangled into '''; completion still loads without crashing, but the resulting words aren't always exactly right. Could be a follow-up.

Testing

Added a regression test (test/completion/aliases.completion.bats) with an alias whose arguments contain a single quote, registering a completion trigger for the aliased command so the test actually exercises the vulnerable code path. Confirmed it fails with a syntax error on the old code (not necessarily the exact EOF message quoted above -- the specific error can vary) and passes with the fix. Ran the full completion and aliases test suites locally (63 tests, all passing).

[seefood]

excellent catch! LGTM.

@akinomyoga want to have a look at well?