fix: locale-sensitive header normalization

[jbeckwith-oai]

Summary

Fixes #1928.

This updates buildHeaders to normalize header names to lowercase ASCII before passing them into Headers.delete() and Headers.append().

Root cause

The original issue reports OpenAI-Organization and OpenAI-Project becoming invalid header names like openaı-organization under a Turkish locale. Plain JavaScript String.prototype.toLowerCase() is not locale-sensitive in Node, so changing only the internal comparison key is not enough. The issue workaround points at the real failure mode: a bundled/runtime Headers implementation may normalize the mixed-case header name with Turkish locale rules when we call delete() or append() with the original name.

By handing Headers the SDK's already-normalized ASCII header name, we avoid depending on the runtime's header-name casing behavior.

Validation

• pnpm exec jest tests/buildHeaders.test.ts --runInBand
• env LANG=tr_TR.UTF-8 LC_ALL=tr_TR.UTF-8 pnpm exec jest tests/buildHeaders.test.ts --runInBand
• pnpm exec jest tests/index.test.ts tests/lib/azure.test.ts tests/log.test.ts tests/lib/workload-identity.test.ts tests/lib/bedrock.test.ts tests/api-resources/webhooks.test.ts --runInBand
• pnpm exec jest tests/buildHeaders.test.ts tests/api-resources/chat/completions/completions.test.ts --runInBand
• pnpm exec jest tests/api-resources/beta/threads/threads.test.ts tests/api-resources/videos.test.ts --runInBand
• pnpm exec prettier --check src/internal/headers.ts tests/buildHeaders.test.ts
• pnpm exec eslint src/internal/headers.ts tests/buildHeaders.test.ts
• pnpm exec tsc --noEmit
• pnpm build

[HAYDEN-OAI]

looks good overall. thanks for the careful root-cause analysis and targeted coverage!

[HAYDEN-OAI]

nice fix, and the regression coverage for locale-sensitive Headers implementations is convincing. one edge case worth following up on: toLowerCase() performs full unicode case conversion, so an invalid name like cooKie becomes the valid cookie before Headers can reject it. validating the original name against the ascii http-token grammar before lowercasing would preserve the existing rejection behavior without reintroducing the turkish locale issue.