fix(testing): match shell echoes the shell actually produces

[chaliy]

What changed

The fuzz leak detector now recognizes a shell echo of user input regardless of how the shell transformed that input, and regardless of whether the diagnostic fits on one line. Test-harness and fuzz-target only — no shell behavior changes.

Why

glob_fuzz went red again on main (run 219, input \n \0{ code:<(])\n) — the second failure of this class in two days:

[glob_fuzz] stderr leaks banned shape ` { code:` (after stripping shell echoes):
---raw stderr---
bash: { code:<(])
: command not found

#2319 fixed one diagnostic's wording. That was necessary but treated a symptom. Run 219 exposed the two harness gaps underneath both failures:

1. The pre-filter checked raw bytes; the shell transforms them. glob_fuzz skips inputs containing a UNIVERSAL_BANNED shape, but NUL is dropped during word expansion — so \0{ code: and </r\0ustc/ slip past a literal contains and reappear in stderr as { code: and /rustc/. Both crashes hinge on exactly this. Targets now call input_echo_would_trip, which checks the input as the shell will render it back.

2. The echo filter was line-based; the input slot can contain newlines. The command name here ends in \n, so bash's one-line bash: %s: command not found template renders across two lines and neither half matched. Matching is now span-based: an unclosed bash: / ls: cannot access line extends to the first later line ending in a template suffix.

Quote and backslash removal can synthesize banned shapes the same way NUL stripping does. That's why the detector no longer leans on the pre-filter alone — the two are independent defenses, and fixing only the pre-filter would leave the same class open.

Before / After

Both crash inputs replayed through the real fuzz target, rebuilt on this branch:

$ cargo +nightly fuzz run glob_fuzz fuzz/artifacts/glob_fuzz/crash-2ac9af1dcb0bb66e7849b347828a8236e12fdda3
Executed crash-2ac9af1dcb0bb66e7849b347828a8236e12fdda3 in 1 ms
$ cargo +nightly fuzz run glob_fuzz fuzz/artifacts/glob_fuzz/crash-fae53719665e9c77d2e1a1b7d4da7e43902525d0
Executed crash-fae53719665e9c77d2e1a1b7d4da7e43902525d0 in 0 ms

On main, the first of those aborts with libFuzzer: deadly signal.

Risk

• Low — no library code changes; src/testing.rs is #[doc(hidden)] test-only.

Span matching widens what gets stripped, so the guard is that it strips as little as possible: the span closes on the earliest candidate line, so a leak printed after a complete diagnostic survives (strip_span_stops_at_the_first_closing_line), and an unclosed bash: line swallows nothing (strip_keeps_unclosed_bash_prefix_span). A leak cannot be swallowed inside a span, because each shell diagnostic is formatted and written as one string with nothing interleaved.

The pre-existing strip_keeps_* tests — which exist precisely to catch over-stripping — all still pass unmodified.

Verified: 16/16 testing:: unit tests, 6/6 glob_fuzz scaffold tests, workspace lib/bins/tests, doc tests, realfs, failpoints, proptest_security, clippy -D warnings, cargo fmt, check_okf.py, check_doc_links.py, plus a long glob_fuzz session against the rebuilt target (CI reached its failure at ~121k executions).

ssh_supabase_connects fails locally (sandbox blocks outbound port 22) — environmental, green on main in CI.

Checklist

• Tests added or updated
• Backward compatibility considered

---

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	39400b5	Commit Preview URL Branch Preview URL	Aug 21 2026, 10:53 AM