test(fuzz): skip arithmetic inputs the diagnostic would echo back

[chaliy]

What changed

arithmetic_fuzz pre-filters its input with input_echo_would_trip, the
defense glob_fuzz and cpython_fuzz already use, so inputs whose own text
would come back inside the diagnostic no longer trip the leak detector.

No production code changes — fuzz harness and tests only.

Why

An arithmetic error names the expression and the unparsed rest verbatim, so the
script's own text returns in stderr. Real bash does the same. When that text
happens to contain a UNIVERSAL_BANNED substring, assert_no_leak fires on an
echo rather than a leak — the tokenizer's Tok enum is never formatted
anywhere.

Fuzz run 271 (on current main) found a 20-byte input that does exactly this:

$ bashkit -c 'echo $((:A>>=::TTA:::Tok::::))'
bash: line 1: :A>>=::TTA:::Tok:::::  syntax error: operand expected
(error token is ":A>>=::TTA:::Tok::::")
                              ^^^^^^ the banned shape `Tok::`, supplied by the input

testing.rs already anticipates this class — Tok:: is the example in
strip_real_shell_error_lines's own doc comment — and provides two independent
defenses: a diagnostic that matches a recognized real-shell template gets
stripped, or the target pre-filters at the input layer. Arithmetic diagnostics
are not one of those templates (SHELL_ERROR_SUFFIXES covers
command not found, the errno templates and redirect refusals), and
arithmetic_fuzz had neither defense. That is the whole bug.

Of the seven targets in the nightly fuzz matrix this was the only gap: glob
and cpython already pre-filter, and parser, lexer, analyze and
cpython_http never call the leak check.

Before / After

Before: arithmetic_fuzz run 271 -> deadly signal, exit 77
        "stderr leaks banned shape `Tok::`" on the input's own text
After:  the input is skipped at the input layer; the leak detector stays
        strict for shapes bashkit actually emits

Reproduced locally through the CLI (output above), then covered by two
regressions, both passing with the 9-case scaffold suite:

• run_271_banned_shape_comes_from_the_input — asserts the pre-filter
  recognizes the input, and that the diagnostic is still a faithful,
  bounded echo, so the skip hides a false positive rather than a real leak.
• same_shape_without_the_banned_text_is_still_checked — the same expression
  shape with X:: instead of Tok:: is not filtered and still goes through
  the unfiltered leak check, proving the filter is not blanket-suppressing this
  code path.

Risk

• Low. Test-only; no production code, limits or diagnostics change.
• The pre-filter costs coverage of inputs containing banned substrings, which
  are exactly the ones whose echo is indistinguishable from a leak. The second
  regression pins that the surrounding path is still checked. UNIVERSAL_BANNED
  is untouched, so a shape bashkit really emits still fails.

Checklist

• Tests added or updated
• Backward compatibility considered

https://claude.ai/code/session_01BegucbcUgCEsPZiAzt6GRx

---

Generated by Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	bashkit	7371f36	Commit Preview URL Branch Preview URL	Oct 10 2026, 10:43 AM