Use test-api-key to prevent login thread leaks

[delner]

Problem

Many of my new tests were failing locally with WebMock errors.

The issue is related to test duration, but it's more specifically about orphan background threads and stub lifecycle:                                                                                                                                        
                                                                                                                                                                                                                                                              
The sequence:                                                                                                                                                                                                                                                
                                                                                                                                                                                                                                                              
1. Test A creates State with api_key: "test-key" (no org_id) → spawns background login thread                                                                                                                                                                
2. Login fails (stub returns 500, or cassette doesn't have login stub) → thread enters exponential backoff retry                                                                                                                                             
3. Test A completes → WebMock.reset! clears stubs (or VCR cassette is ejected)                                                                                                                                                                               
4. Test B starts → registers new stub with different Authorization header (e.g., Bearer <BRAINTRUST_API_KEY>)                                                                                                                                                
5. Orphan thread wakes up, tries login with Authorization: Bearer test-key                                                                                                                                                                                   
6. Stub doesn't match → WebMock::NetConnectNotAllowedError                                                                                                                                                                                                   
                                                                                                                                                                                                                                                              
Why some CI jobs are more affected:                                                                                                                                                                                                                       
                                                                                                                                                                                                                                                                                              
- More tests = more opportunities for:                                                                                                                                                                                                                       
  - Threads to be spawned                                                                                                                                                                                                                                    
  - Threads to be in backoff when stubs change                                                                                                                                                                                                               
  - New tests to register incompatible stubs                                                                                                                                                                                                                 
- Longer test duration increases window for race condition      

Solution

There's a magic test key in the code:                                                                                                                                                                                                                        
                                                                                                                                                                                                                                                               
# In lib/braintrust/api/internal/auth.rb                                                                                                                                                                                                                     
if api_key == "test-api-key"                                                                                                                                                                                                                                 
  Log.debug("Login: using test API key, returning fake auth")                                                                                                                                                                                                
  return AuthResult.new(org_id: "test-org-id", ...)                                                                                                                                                                                                          
end                                                                                                                                                                                                                                                          
                                                                                                                                                                                                                                                              
So api_key: "test-api-key" skips HTTP entirely and returns fake auth.

This also happens to short-circuit the login thread. Since this is the current idiomatic approach, apply it for now. In the future, we should eliminate this "magic" behavior and keep test behavior isolated to the test suite.

[clutchski]

constant?