bugfix(client/rust): isolate MAT_VIS_HF_BASE per-test (#241)

[gerchowl]

Summary

Closes #241.

The Rust client suite mixed httpmock-based unit tests (which set MAT_VIS_HF_BASE to a 127.0.0.1 mock server) with @live tests that need to hit real HF. Even with cargo test -- --test-threads=1 the env var leaked across tests in the same process, so live tests 404'd on http://127.0.0.1:<port>/v2026.04.2/release-manifest.json instead of huggingface.co — the underlying bug behind #241 (and the SIGABRT symptom that originally got reported).

Choice: Option C — in-tree EnvGuard RAII helper

Considered three options from the issue:

• A — serial_test: doesn't restore env, just serializes execution; new dev dep.
• B — scopeguard::defer!: correct, but requires a new dev dep.
• C — in-tree EnvGuard: ~20 LOC, zero new deps, typed handle that's hard to forget. Picked C.

EnvGuard::set("MAT_VIS_HF_BASE", url) stores the prior value, sets the mock URL, and Drop restores (or removes if previously unset). The existing ENV_LOCK: Mutex<()> is kept so the suite stays race-safe under the default parallel cargo test runner — std::env::set_var is unsafe in the 2024 edition for that reason. Drop order: each test declares _lock before _env, so reverse-decl order drops _env (env restoration) while _lock is still held.

The hand-rolled env restore in e2e_fetch_color_png and resolved_tag_falls_back_to_default is also folded into the same helper / lock pattern — the e2e test now becomes panic-safe (was previously best-effort with manual unsafe blocks).

#253 rollback

Rolled back the .dagger/src/mat_vis_ci/main.py::test_client_rust dual-invocation workaround from #253. With proper test-side isolation, the single-invocation form is correct again — and avoids the second apt-get install + cargo test boot.

Before / after

Before (against bugfix/241-rust-live-isolation head, single cargo invocation): live tests after mock tests fail with 404 on http://127.0.0.1:<port>/v2026.04.2/release-manifest.json — env var leaked.

After (this branch, single cargo invocation):

$ MAT_VIS_LIVE_TESTS=1 MAT_VIS_LIVE_TAG=v2026.04.2 cargo test -- --test-threads=1
...
test result: ok. 21 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 12.95s

All 21 tests (including 4 live_*) pass cleanly in a single process — no SIGABRT, no 127.0.0.1 leakage in live test output.

Test plan

• cargo build --tests clean
• cargo test -- --test-threads=1 (no live) — 21 passed
• MAT_VIS_LIVE_TESTS=1 MAT_VIS_LIVE_TAG=v2026.04.2 cargo test -- --test-threads=1 — 21 passed including 4 live tests
• CI green on the rolled-back single-invocation test_client_rust