executor: a backquoted body inside double quotes unescapes " (#743)

[berrym]

Closes #743.

The rule is context-dependent

POSIX 2.6.3 gives the backquoted form an escape set that depends on where it appears: a backslash is special before $, a backtick or another backslash — and before " as well when the backquoted form is inside double quotes.

#732 implemented the three-character set. That is correct unquoted, where lush and every peer already agreed, but incomplete inside double quotes:

$ lush -c 'echo "pre`printf \"a b\"`post"'
pre"apost                       # bash, dash: prea bpost

It was never truncation

The issue (which I filed) described this as the word being truncated. It is not. The \" survived into the body, so the sub-parse saw printf \"a b\" and passed printf two arguments — "a and b" — of which printf rendered only the first. Unescaping the quote makes it a real quote delimiter, so the body passes one argument. That is the entire difference.

The context is carried, not guessed

expand_quoted_string_prov already has an in_double_quotes parameter for exactly this kind of question, and its three callers already pass it correctly (true for a double-quoted word, false for a subscript key). It is threaded to lush_backtick_unescape through a new expand_command_substitution_ex; the old name remains as the unquoted-context wrapper, so no other call site changes.

$( ) deliberately does not take the rule

Its body is a fresh command, so \" keeps its backslash even inside double quotes:

$ lush -c 'echo "pre$(printf \"a b\")post"'
pre"apost                       # bash, dash: identical

lush already matched both peers here. The test pins it so the two forms cannot be conflated later.

Verification

• 5 new checks in tests/integration/test_nested_backticks.c: the reported case, the same without surrounding literals, a single-word body unaffected either way, the unquoted form that must keep its backslash, and the $( ) form that must not take the rule.
• 3 of the file's 34 checks fail against the parent build, with a control proving the binaries differ.
• Full suite 195/195. ASan 65/65 and the backtick test 1/1 under ASan — both run with the repaired asan-check, whose worktree sync had been silently stale earlier today.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!