… tests

5 new scenarios + 3 Go tests for the while_clause shell feature:

- redirect_loop_body_write_blocked: write redirect on loop body rejected
- redirect_loop_body_input_outside_allowed_blocked: input redirect on loop
  body honors AllowedPaths
- cmdsub_forbidden_in_cond_blocked: command substitution in cond honors
  AllowedPaths
- subshell_loop_var_isolation: variables assigned in a subshell-wrapped
  while loop do not leak to the parent
- while_brace_cond_with_exit: `exit N` in a brace-grouped cond exits the
  shell with status N before body runs
- TestUntilFalseRespectsContextCancellation: symmetric twin of the
  existing while-true ctx-cancel test
- TestSubshellInfiniteWhileRespectsContextCancellation: outer ctx cancel
  propagates through subshell into inner infinite while
- TestWhileEmptyCondIsParserError: `while ; do :; done` rejected at parse

…ssions

7 new scenarios documenting inline-command-variable behaviour under hostile
inputs (redirects, denied cmdsub, disallowed commands, IFS scoping, multi-
assignment, invalid identifiers).

The associated FINDING for same-name multi-assignment leak is tracked
privately in the rshell-vuln-hunt repo until a fix lands here.

Adds 3 regression tests:
- -I substituting cmdName remains bound by CommandAllowed post-substitution
- infinite NUL stream in default mode honors the 30s context timeout
- -L value validation mirrors -n (0, negative, > HardMaxArgs rejected)

5 regression tests covering: symlink-out-of-sandbox rejected;
--output-delimiter forwards control bytes within 1MB cap;
1MiB+ line exits 1; -b range overflow at strconv rejected;
-b1-99999999999 processes correctly bounded by line length.

Adds 2 regression tests:
- heredoc post-expansion cap is reachable only as defense-in-depth;
  layered per-var (1 MiB) + per-runner storage (2 MiB) caps fire first
- heredoc writer goroutine honors ctx cancel between 32 KiB chunks

Adds 1 regression test pinning that a symlink whose target lies outside
AllowedPaths is observable in pwd output but reads through the symlink
remain blocked by the sandbox.

…tests

…d-attack tests

…isolation scenario

…tripped + empty-post-strip-lines

…directory + --compress-program rejection

…ectations

The four scenarios were failing CI because they relied on shell features
rshell does not implement, or because their expectations did not match
either rshell or bash behavior:

- shell/while_clause/continue_huge_n: used `$((i+1))` arithmetic expansion
  (not supported) and expected `continue 999999` to break out of the loop.
  Bash actually clamps N to the nesting depth, so `continue 999999` in a
  one-deep loop is equivalent to `continue 1`. Rewritten with an empty/
  non-empty sentinel and corrected expectation.
- shell/field_splitting/ifs_backslash: used `set --` plus `$#`/`$1` which
  rshell does not support. Rewritten to drive a `for x in $s` loop, which
  exercises the same IFS field-splitting behavior.
- shell/case_clause/expansion_does_not_reparse: original script
  `$KW x in y) echo no;; esac` is a parse-time error before $KW ever
  expands. Reduced to `KW=case; $KW`, which still demonstrates that the
  expanded "case" word is treated as a command name, not a keyword.
- shell/function/expansion_does_not_reparse: only the expected stderr
  substring was wrong — rshell emits "unknown command", not bash's
  "command not found". Updated stderr_contains accordingly.

On Windows, the error message contains backslash path separators
(`openat etc\passwd: ...`) and does not match a stderr_contains check
for "/etc/passwd". Match on "passwd" instead, which is sufficient to
distinguish the failed read from any other stderr output and works
on all platforms.