feat: capture per-trial LLM token usage via TrajectoryProxy

[EYH0602]

Summary

Wire up the existing TrajectoryProxy (benchflow/trajectories/proxy.py) so every trial automatically captures per-trial LLM token usage (input/output tokens) from API responses. Today, acp_trajectory.jsonl records ACP session events (tool calls, messages, thoughts) but no API-level usage data.

Motivation

• Cost tracking: users need to know how many tokens each trial consumed to estimate experiment costs.
• Infrastructure already exists: TrajectoryProxy handles both streaming SSE and regular responses, reconstructs usage for both Anthropic and OpenAI formats, and exposes total_input_tokens / total_output_tokens properties on the Trajectory model.
• Downstream consumers expect it: analysis scripts already read agent_result.n_input_tokens / n_output_tokens from result.json — they're just never populated.

Proposed Implementation

1. Start the proxy in Trial.connect() / Trial.connect_as()

Before the agent launches, spin up a TrajectoryProxy targeting the real LLM endpoint:

from benchflow.trajectories.proxy import TrajectoryProxy

proxy = TrajectoryProxy(target=real_api_base, session_id=self._trial_name)
await proxy.start()

2. Route agent LLM traffic through the proxy

Inject the proxy URL into the agent's environment so all LLM calls go through it:

agent_env["OPENAI_BASE_URL"] = proxy.base_url     # for OpenAI-compatible agents
agent_env["ANTHROPIC_BASE_URL"] = proxy.base_url   # for Anthropic agents

3. Stop the proxy and collect usage in Trial.disconnect()

await proxy.stop()
traj = proxy.trajectory
self._n_input_tokens += traj.total_input_tokens
self._n_output_tokens += traj.total_output_tokens

4. Write token counts into result.json

Add n_input_tokens and n_output_tokens to the result dict in SDK._build_result().

5. (Optional) Save raw LLM trajectory

Write llm_trajectory.jsonl alongside acp_trajectory.jsonl for detailed per-exchange analysis:

trajectory/
├── acp_trajectory.jsonl      # ACP session events (existing)
└── llm_trajectory.jsonl      # Raw LLM API exchanges with usage (new)

Key Properties

• Zero adapter changes — all agents benefit automatically since the proxy sits between the agent and the LLM API.
• Provider-agnostic — both OpenAI and Anthropic streaming/non-streaming formats are already handled by _reconstruct_response().
• Backward-compatible — existing result.json gains two new optional fields; no schema break.

Acceptance Criteria

• result.json includes n_input_tokens and n_output_tokens for every completed trial
• Token counts match the sum of usage fields from actual API responses
• Works for both OpenAI (gpt-*) and Anthropic (claude-*) models
• Proxy adds < 50ms overhead per API call
• summary.json aggregates total tokens across all trials in a job