fix(linter): SC2242 counts case/loop nesting instead of flagging it (closes #331)

[noahgift]

Two defects in rash/src/linter/rules/sc2242.rs, one shape — a detector reporting a result it did not measure.

Reproducer (14 lines; bash -n accepts it, and it runs)

#!/usr/bin/env bash
set -euo pipefail
while IFS= read -r id; do
  case "$id" in '') continue ;; esac
  for a in 1 2; do
    for b in 3 4; do
      printf '%s%s\n' "$a" "$b"
    done
  done
  if [ -z "$id" ]; then
    continue                    # <- reported as "outside a loop"
  fi
  printf '%s\n' "$id"
done < <(printf 'x\n')

✗ 11:1-13 [error] SC2242: Can only break/continue from loops. Use 'exit' to exit case or function

Root cause

check() tracks context with in_case / in_loop / in_function as booleans, not depth counters:

1. An inner done closes the outer loop. for/for set in_loop = true (already true), then each
   of the two done lines sets in_loop = false — so after the nested loops close, the still-open
   while is invisible.
2. A one-line case … esac never closes. is_case_end is line.trim_start() == "esac", which is
   false for a case … ;; esac written on one line, so in_case stays true for the rest of the file.

Either alone is survivable; together they make in_case && !in_loop true inside a loop, and every later
break/continue is flagged. Found while linting a done_when script in paiml/infra (PMAT-534): the
rule reported 2 errors on a script bash -n accepts and that runs correctly.

Fix

fix/sc2242-nesting-depth: keyword-TOKEN depth counters (;, (, ), &, | and backtick are
separators, so a break ;; esac clause yields break and esac as their own tokens); openers count
before the check and closers after, so a one-line case whose clause breaks is still judged inside its own
case while a one-line for … done balances. Still a line/keyword heuristic, not a parser — a keyword
inside a quoted string or a heredoc is still miscounted, and the code says so — but it is no longer wrong
for code anyone writes.

Six tests added, including the reproducer above, donefile/casework (a word containing a keyword is
not the keyword) and a case-in-loop-in-case nesting. Discrimination proven: reverting check() to
the boolean behaviour turns exactly the two new nesting tests RED
(a_continue_after_nested_loops_close_is_not_in_the_case, a_one_line_case_closes_on_its_own_line) and
leaves the other 14 green. The rule still fires on the real defect: a break in a case outside any loop
is still one error, whether the case is written on one line or three.

---

Verification

• cargo test -p bashrs --lib linter:: — 6067 passed, 0 failed
• cargo clippy -p bashrs --lib -- -D warnings — clean
• cargo fmt --check — clean
• The built binary (not the claim about it) gives 0 errors on the script that started this, and the 14-line reproducer lints clean.
• Discrimination proven: reverting check() to the boolean behaviour turns exactly the two new nesting tests red — a_continue_after_nested_loops_close_is_not_in_the_case and a_one_line_case_closes_on_its_own_line — and leaves the other 14 green.

Found while landing paiml/infra PMAT-534, where these were 2 of the 6 bashrs lint errors on a done_when script. The other 4 were real findings and were fixed there.

🤖 Generated with Claude Code