fix(eval): emit llm_trajectory.jsonl for streaming claude-agent-acp rollouts

[Yiminnn]

Summary

claude-agent-acp did not write trajectory/llm_trajectory.jsonl for successful rollouts (0/50), while pi did (44/44) and claude-acp's error path did (11/17). This breaks trajectory parity and makes successful claude rollouts unusable for any LLM-trajectory consumer (export / RL training).

Root cause

The per-rollout LiteLLM proxy captures the LLM trajectory from its litellm_settings.callbacks success/failure events. BenchFlow's vllm/ provider maps the upstream to openai/<model>. For the anthropic /v1/messages route against an openai/-prefixed upstream, LiteLLM 1.89 bridges through the OpenAI Responses API (/responses), and its Responses streaming adapter never invokes the success callback — so streaming /v1/messages success produces zero callback records → empty server.trajectory.exchanges → _write_llm_trajectory early-returns and writes nothing.

• claude-agent-acp (Claude Code) streams /v1/messages → hits the gap (0/50 on success).
• pi calls /v1/chat/completions directly → streaming success logs fine (44/44).
• claude-acp errors go through post_call_failure_hook (failures do log) → that's the 11/17 on the error path.

Fix

Set litellm.use_chat_completions_url_for_anthropic_messages via the proxy's litellm_settings, so the anthropic /v1/messages bridge uses /chat/completions — the universally-supported endpoint whose streaming path logs success correctly. Native anthropic/ providers don't use the Responses bridge (unaffected); the proxy is BenchFlow-owned and per-rollout, so there is no external blast radius.

Verification

Reproduced with a live-proxy harness (the real proxy started via the runtime + a mock OpenAI-compatible upstream), driving the exact openai/ route the vllm/ provider produces:

case (openai/ route = #833 config)	upstream hit	success rec	exchanges	file
messages-stream (claude-acp) — before	/responses	0	0	✗
messages-stream (claude-acp) — after	/chat/completions	1	1	✓
chat-stream (pi) — after	/chat/completions	1	1	✓ (unchanged)

A regression test (test_proxy_config_forces_chat_completions_for_anthropic_messages) asserts the flag is present in the generated proxy config. tests/test_litellm_config.py + tests/test_litellm_hardening.py pass (57 passed).

Closes #833

[bingran-you]

Users Simulation automation review (2026-06-27T12:21Z): ready for this simulation scope.

I verified the PR through cost-light CLI/SDK and focused code paths rather than a live expensive model rollout. The LiteLLM proxy flag is wired in the canonical config builder, focused runtime/hardening checks passed, and the trajectory validator still fails closed when trajectory/llm_trajectory.jsonl is missing.

Commands/evidence:

• uv run pytest tests/test_litellm_config.py -q
• uv run pytest tests/test_litellm_runtime.py -q
• uv run pytest tests/test_litellm_hardening.py -q
• uv run ruff check src/benchflow/providers/litellm_config.py tests/test_litellm_config.py
• uv run ty check src/benchflow/providers/litellm_config.py
• CLI/SDK dry paths: benchflow --help, benchflow eval run --help, benchflow eval list --help, benchflow skills --help, benchflow hub check --help, benchflow sandbox --help, Python SDK import smoke.
• validate_run_artifacts.py passed on a healthy synthetic rollout and failed closed after removing llm_trajectory.jsonl.

Residual risk: the regression test is static config coverage, not a real streamed claude-agent-acp provider canary, so this does not prove live Claude streaming end to end.