fix(path-cli): unify test env lock to stop cmd_pathbase racing the cache (red main)

[akesling]

Symptom

main is red. The CI run for the #112 merge (run 28177878573) failed one test:

cmd_resume::tests::resolve_input_url_uses_cache_on_hit_without_refetching
  panicked at crates/path-cli/src/cmd_resume.rs:873:
  download of alex/pathstash/fe94b6f9-... failed (500 Internal Server Error): boom

That test seeds the cache and points downloads at a 500 mock to prove resolve_input never hits the network on a cache hit. It reached the network — a cache miss.

Root cause — two locks, one global resource

The failure is a test-isolation race, not a #112 regression (#112 changed only production upload code; it surfaced a latent flake by chance — the parent run passed by luck, and the test passes in isolation).

The process environment was guarded by two independent mutexes:

• cmd_pathbase's EnvGuard → a private ENV_LOCK
• cmd_resume / cmd_cache / cmd_export / config → config::TEST_ENV_LOCK

Both mutate/read TOOLPATH_CONFIG_DIR. Under parallel cargo test, an EnvGuard::set("TOOLPATH_CONFIG_DIR", dirB) can run while the cache-hit test holds TEST_ENV_LOCK with dirA, so config_dir() reads dirB → the seeded cache file isn't there → download → 500. (std::env::set_var/var_os are also not thread-safe, so concurrent access across the two locks is a data race on the environ regardless.) The EnvGuard SAFETY comment claiming "no other tests mutate or read TOOLPATH_CONFIG_DIR" was false.

Fix

EnvGuard now acquires the shared config::TEST_ENV_LOCK; the private ENV_LOCK is deleted. There is now exactly one env lock in the crate, so every env-touching test is mutually exclusive by construction. (The PATH/HOME resume guards already run only inside TEST_ENV_LOCK-holding tests, so they were never part of the race.)

Test-only change; no production code.

Verification

• 14 full path-cli --lib runs green, including 6 at RUST_TEST_THREADS=32
• cargo clippy -p path-cli --all-targets -- -D warnings clean
• Couldn't reproduce the original failure locally (needs CI's parallelism); the fix is a structural single-lock guarantee rather than a timing tweak.

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[github-actions]

🔍 Preview deployed: https://b19868e7.toolpath.pages.dev