fix(providers): stream OpenAI Responses tool calls and reasoning, mark it verified

[zhanghanduo]

Problem

OpenAIResponsesClient.stream() only mapped text / reasoning-summary / usage events, so a streamed Responses-API turn silently lost every function_call (the loop saw no tool call and stopped) and every encrypted_content reasoning item (the next request couldn't continue the model's reasoning state). protocol: responses was listed in UNVERIFIED_STREAM_PROTOCOLS, so the loop never streamed it on its own — but a profile that explicitly set stream_llm_tokens: true (e.g. to survive a gateway's first-byte timeout on long generations) hit this silently and lost tool calls.

Fix

• stream() now forwards function-call argument deltas as they arrive (response.output_item.added / response.function_call_arguments.delta), and on response.completed hands the terminal response's parsed block list (reasoning items with encrypted_content + function calls, in provider order) to the assembler via StreamDelta.reasoning_blocks — the same content chat() already returns. Any call whose events a gateway didn't forward is completed from that terminal response (_unstreamed_call_deltas).
• responses removed from UNVERIFIED_STREAM_PROTOCOLS (now just bedrock), so it streams automatically once an observer or the reasoning-only watchdog asks for deltas — matching anthropic's existing behavior.

Tests

• New tests/test_responses_stream_replay.py: drives the real client through the real loop against a fake SDK, once per mode, over a two-call tool turn. Requires byte-identical follow-up requests and message history between streaming and non-streaming (both with and without per-item SDK events, to cover a gateway that forwards only deltas + the terminal event). Fails without the fix (confirmed via git stash).
• tests/test_stream_llm_tokens_toggle.py updated: responses moved out of the "stays non-streaming, warns" parametrization into its own "streams when an observer wants deltas" test, mirroring the existing anthropic test.
• Full suite: 2338 passed, 2 skipped. ruff check clean.

Live verification

Live-verified against two models on llm-hub's real /v1/responses endpoint via scripts/protocol_smoke/live_responses_stream_smoke.py (new):

• gpt-4o (first key tried; only model that key's channel let through to a native Responses upstream): streaming correctly accumulates one tool call's arguments across 11 deltas, executes it, and produces a second-turn history structurally identical to the non-streaming run.
• gpt-6.1-sol (after a key swap unlocked that channel): same tool-call check, plus a forced reasoning={"effort":"medium","summary":"auto"} run on a tool+reasoning turn — streaming and non-streaming both produce reasoning (with non-empty encrypted_content) → function_call(get_weather, same args) → text, with matching final answers.

Details in the companion repo's temp/2026-10-14_responses-stream-live-smoke.md (not part of this PR; AgentCore itself ships the reusable smoke script).

🤖 Generated with Claude Code