fix(expander): split and glob an unquoted ${p:-w} / ${p:+w} / ${p:=w} result

[LESdylan]

What & why

P='a b c'; set -- ${P:-x}; echo $#      # hellish 1, bash 3
set -- ${u:-*.c}                        # hellish: the word *.c, no glob

POSIX (2.6.2, 2.6.5, 2.6.6) says the result of an unquoted parameter expansion is field-split and pathname-expanded, whichever operator produced it. hellish kept the whole result as one field whenever the text of the operator word had no unquoted blank. That rule was wrong in three ways:

• It looked at the word even when the word was not used. So the variable's own value was never split: ${P:-x}, ${P-x}, ${P:+$P}, ${P:?e}, ${U:="a b"}.
• It could not see inside expansions. ${u:-$P} never split, and ${u:-$Q} and ${u:-a*} never globbed.
• A word with a blank was split flat, quotes and all. ${u:-"a b"c d} gave three fields (a / bc / d) where bash gives two (a bc / d), and ${u:-a\ b} gave two.

Root cause. expand_op_token retyped the token to TT_DQWORD (one field, no glob) based on opword_no_split(word text). By the time the result reached the splitter, the quoting of each part of the word was gone.

Change.

• The variable's value, and the new value set by ${p:=w}, stay a plain unquoted expansion, split and globbed exactly like $p.
• A used word of - or + is expanded by pf_op_word_segments (expand_param_opword.c) into segments that keep their quoting:
  • q for '...', "...", $'...', a backslash escape, and a tilde prefix. POSIX 2.6.1 makes ${u:-~} one field even when HOME contains a blank.
  • u for everything else, literal text included.
• The segments are parked behind a SEG_MAGIC marker, the same way ${a[@]} parks its elements. The splitter (emit_op_segments) IFS-splits the u segments into the current field and appends the q segments verbatim as TT_DQENVVAR children, so the glob stage escapes them.
• Empty results are unchanged:
  • an unquoted empty result is still no field, as in git-completion's ${a:+"${a[@]}"} ${d:+--git-dir="$d"};
  • ${x:-""} is still one empty field.
• opword_no_split is gone; nothing else used it.

How I verified it

• Reproduced first. tests/scripts/57_opword_fields.sh is graded against bash --posix 5.3.9. It covers:

  • the variable's own value under each operator;
  • used words: unquoted, quoted, backslash-escaped, $'...', and mixed;
  • globs, command substitution, and $@ / $* / "$@" in the word;
  • empty results next to literal text, :=, and a tilde with a blank HOME;
  • IFS=:, set -f, and the contexts that do not split (quotes, assignment, case, for).

  The binary without this fix diverges on 25 of its 65 lines; this branch matches bash byte for byte.

• Local gates on this commit (ASan debug build):

  • golden tests/tester: 5363/5363;
  • tests/run_scripts.sh against bash --posix: 126/126;
  • verify_alloc.sh: identical output on both heaps;
  • alloc_stress.sh: all clean;
  • tests/pty_suite.sh: 111 ok, 6 skipped, 4 failed. plugin_corpus_test, which loads git-completion and its ${a:+"${a[@]}"} idioms, passes. None of the four failures is this change:
    • prompt_compat_matrix, prompt_drift_matrix and prompt_jobs_badge expect the non-root %/$ prompt and get #, because the container runs as root. They fail the same way on develop, and CI runs them as a normal user.
    • hxp_framework_test hit the 420 s per-file limit on this 4-core container. develop's binary takes 440 s for it on the same machine, and CI's runners finish it inside the limit.

• norminette is OK on every touched file.

Notes / trade-offs

• This is a behaviour change for scripts that relied on the old one-field result of ${p:-a b} unquoted. The new result is what bash, dash and POSIX give; quoting the expansion keeps one field in every shell.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RAmeHfJNm7XjYbMNrQqvkG

---

Generated by Claude Code