[TheGreatAxios]

Summary

• The openai-compatible Chat Completions adapter tolerates explicit null for delta.role and delta.tool_calls on streamed SSE chunks and books them the same as those keys being absent, so vLLM-serializer-relayed endpoints (NEAR AI, Qwen3.8-Flash-Next on vast.ai 2xH200) no longer fail every streamed turn with protocol_mismatch.
• The non-streaming message schema widens symmetrically (same relays shape both request modes from one serializer), and wrong-type values such as role: 42 still fail schema validation as ProtocolMismatchError — only presence-as-null changes meaning.
• The vendored divergence carries a ledger entry (PATCHES.md, providers-ts-openai-null-delta-tolerance) with a matching marker comment; regression coverage in src/provider/chat-completions-null-delta.test.ts proves null-equivalence event sequences end to end through the builtin and wrapper adapters.

Verification

• bun run check passes (lint, typecheck, dead-exports, build, projects-dir guard; sandboxed suite 9188 pass / 0 fail).
• bun run test passes (9188 pass / 0 fail, --randomize --seed 424242).
• bin/vendor-patch-diff shows exactly the intended hunks for providers/openai.ts; scripts/vendor-patch-ledger.test.ts is green.
• Whole-branch review verdict: GO.

Refs CL-3010

[linear-code]

CL-3010

[corbits-triage]

Triage: Pass

No changes needed from you. A maintainer will review this pull request.