The ResetInventoryCache function creates a race condition. After resetting globalInventoryCache to a new instance, the old sync.Once in the previous instance is lost. If another goroutine was in the middle of calling init() on the old instance, it would complete initialization on the old instance while new callers would get the new uninitialized instance. Additionally, resetting doesn't prevent the old sync.Once from keeping the "already called" state.

While the comment correctly warns this is test-only and not thread-safe, a safer approach would be to reset the fields inside the existing CachedInventory struct rather than replacing the entire global variable. However, sync.Once cannot be reset by design.

Consider adding a mutex-based solution for tests or document more clearly that tests must ensure serial execution when calling ResetInventoryCache.

Tests in this file intentionally do not use t.Parallel() because they mutate global state via ResetInventoryCache(). This is correct, but it should be documented with a comment at the top of the test file to explain why these tests must run serially, especially since most other test files in this package use t.Parallel().

The IsCacheInitialized check is unreliable. It checks if len(globalInventoryCache.tools) > 0, but this has two problems:

1. If AllTools(t) returns an empty slice (which shouldn't happen in practice but is theoretically possible), IsCacheInitialized would return false even after successful initialization
2. After ResetInventoryCache() is called, the check looks at the new instance which will have empty tools, making it work correctly. However, the comment at line 106 says "We can't directly check sync.Once state" which suggests the intent is to check if the Once has fired, not just if tools are populated.

A more reliable approach would be to add an explicit boolean flag that's set to true inside the once.Do callback, which would accurately reflect whether initialization has occurred regardless of whether AllTools returns an empty slice.

The documentation is contradictory. Line 76 states "The cache must be initialized via InitInventoryCache before calling this function" but line 77 immediately says "If the cache is not initialized, this will initialize it with NullTranslationHelper."

This creates confusion about whether initialization is required or optional. Consider rephrasing to clarify the behavior, such as: "The cache should typically be initialized via InitInventoryCache before calling this function. If not already initialized, it will automatically initialize with NullTranslationHelper as a fallback."

Comment uses inconsistent terminology. The comment says "NewInventoryBuilder" but the actual function is called "CachedInventoryBuilder". This creates confusion about which function is being referred to.