Motivation

• Prevent duplicate expensive initializers and races when multiple threads concurrently call Factory::get_or_init for the same (domain, label, spec, variant) identity.
• Preserve existing fast-path cache behavior and reentrant semantics while ensuring only the first initializer runs for a given identity.

Description

• Add a std-gated InitLocks (backed by Mutex<BTreeSet<ArtifactId>> and a Condvar) and an InitGuard type, and wire an init_locks field into Inner to coordinate per-artifact initialization.
• Acquire an initialization guard in Factory::get_or_init (init_locks.acquire(&id)) and re-check the cache after acquiring the guard so waiters return the already-cached value instead of duplicating work.
• Add a concurrency regression test concurrent_same_identity_runs_initializer_once demonstrating eight simultaneous callers run the initializer exactly once and receive the same cached Arc.
• Include small formatting/lint tidy-ups to satisfy the project linting scripts.

Testing

• Ran cargo test -p uselesskey-core concurrent_same_identity_runs_initializer_once --lib and it passed.
• Ran cargo test -p uselesskey-core --lib and the full unit test suite passed (129 passed; 0 failed).
• Ran cargo check -p uselesskey-core --no-default-features and it completed successfully.
• Ran cargo clippy -p uselesskey-core --all-targets -- -D warnings and the crate passed clippy checks; also ran the workspace lint fixes (cargo xtask lint-fix) to apply formatting when cargo xtask lint-fix --check initially flagged a formatting diff.

Codex Task