[Bug]: Provider API-key capture test depends on unrelated 3-second server readiness

[dennisonbertram]

Summary

TestMatrix_ProviderAPIKeyCapture waits for the full harness HTTP server to become healthy within 3 seconds even though its contract is only to verify the API key passed into the provider factory. Under the hosted race suite, parallel startup exceeded that deadline and failed PR #1051 after the provider/server became ready immediately afterward.

Acceptance criteria

• The test synchronizes directly on provider-factory invocation/API-key capture, not HTTP readiness timing.
• The test still proves OPENAI_API_KEY reaches openai.Config.APIKey.
• The harness process is shut down deterministically and cannot leak a goroutine.
• The test passes repeatedly under normal and -race execution.
• The complete regression suite passes in normal, race, and coverage modes.

In scope

• Test synchronization for TestMatrix_ProviderAPIKeyCapture.
• Required regression, plan, impact map, logs, and indexes.

Out of scope

• Production startup sequencing or health endpoint behavior.
• Global timeout increases.

Current-architecture search evidence

• cmd/harnessd/main_test.go:TestMatrix_ProviderAPIKeyCapture captures provider config in the injected provider factory.
• The test then calls awaitHealthy(..., 3*time.Second), coupling an API-key assertion to unrelated server readiness.
• Hosted race run 30583930460 failed at main_test.go:4130, while logs show the same server later listening on the allocated address.

Cross-surface impact

• Harness runtime/API: none; test-only synchronization.
• Cron/callback/continuation: none directly; this removes a false-negative gate blocking their repair chain.
• TUI/native GUI: none.
• Persistence/security: none; the test continues to assert only the injected sentinel key and does not log secrets.
• CI: stabilizes the race suite without weakening product assertions.

Test-first plan

1. Reproduce by running the focused test repeatedly under -race with concurrent package pressure or a deliberately delayed server path if needed.
2. Add a provider-invocation channel and assert the captured key after that signal.
3. Preserve deterministic interrupt and bounded shutdown assertion.
4. Run focused repeated normal/race tests and ./scripts/test-regression.sh.

Rollout / rollback

Test-only change. Roll back the PR if the exact-head hosted normal/race checks or full local gate regress.

Evidence

• PR Keep pinned runs outside completed retention quota #1051 hosted race job: https://github.com/dennisonbertram/go-code/actions/runs/30583930460/job/91010705523