[mpecan]

Fixes the four highest-impact findings from a differential session against bash-oracle (about 34k inputs: our corpus, Parable's upstream corpus of gnu-bash/oils/tree-sitter test scripts, plus mutation and grammar fuzzing). Each one either rejected or silently changed common real-world bash, and in each case Parable already matched bash, so these are porting gaps.

The four fixes

1. =~ regex operands (0b25fd9, 1e4f624). The right-hand side of =~ was read with the normal tokenizer, so any unquoted (, | or < was a syntax error: [[ $x =~ ^(a|b)$ ]], [[ $s =~ ([a-z]+)([0-9]+) ]]. A new Lexer::read_cond_regex_word follows bash's PST_REGEXP and Parable's WORD_CTX_REGEX:
   • parens nest, and blanks are allowed inside a group
   • | is a pattern character
   • outside a group, &, ;, <, > end the word (so [[ a =~ a<b ]] is still an error, as in bash)
   • <(…)/>(…) is a process substitution
   • unlike Parable, a ]] inside the word (x]]) stays part of the pattern, as in bash.
2. Quoted here-documents (8a57eac). With <<'EOF', <<"EOF" or <<\EOF, bash keeps \<newline> in the body as-is. The reader always treated it as a line continuation, so a cat <<'EOF' > script.sh body lost every \, and a \ right before the delimiter line hid the delimiter. PendingHereDoc now carries the quoted flag, and backslash processing is skipped for quoted bodies.
3. Reserved words after a compound command (797bac2). Following bash's reserved_word_acceptable, a keyword may come right after fi, done, esac, }, ) or ]]. Before, …; fi done and { a; } done were rejected, and a stray fi fi was accepted with the second fi as a command word.
4. fd prefixes of any int width (1078a7d). is_fd_number only allowed 1–2 digits, so exec 100<>file and 212>log parsed the number as a command word. It now accepts any all-digit value that fits in i32; larger numbers stay words, as in bash. coproc 2>err cmd and x | time 2>1 now detect fd prefixes too.

Follow-ups for regressions exposed by fix 3

Re-running the oracle after fix 3 found invalid inputs that were previously rejected only by accident, through other existing leniencies:

• ae370de: a redirect target is never a reserved word. Otherwise >esac while … parsed a loop.
• 0fa15ca: a coproc simple command collects all of its words. coproc cat * x=y and coproc>f e used to split the leftover words off into a second command.
• 896181e: top-level commands need a separator. (a) b, echo a (x), (a) { b; } and esac case … used to parse as two commands.

Verification

• cargo fmt, clippy -D warnings, cargo test (387 tests) and lint-extra are clean. Each commit was tested on its own.
• Python bindings: tests/parable passes 1638/1638.
• Oracle session, same seeds before and after: 246 mismatches fixed, 0 changed for the worse. 5 inputs newly accepted that bash rejects. All 5 are leniencies that already exist on main (multi-word case patterns like case $x in a b), and an unterminated e[); earlier errors used to hide them. Out of scope here.
• Parable upstream corpus (about 4.5k cases): failures go from 166 to 141, with 0 new failures.

New tests: parser/cond_regex_tests.rs, lexer/heredoc_quoted_tests.rs, parser/reserved_after_compound_tests.rs, parser/fd_prefix_tests.rs, parser/coproc_tests.rs, plus additions to parser/syntax_error_tests.rs. Expected values come from bash-oracle output.

🤖 Generated with Claude Code