fix(eval): persist plan-execute token usage for cost/tokens reporting

[AmirF194]

Description

PlanExecuteRunner meters real token usage via _TokenMeter and writes accurate totals onto the OTel span (gen_ai.usage.input_tokens/output_tokens), but that data never reaches the persisted trajectory. persist_trajectory() only receives trajectory=list[StepResult], and StepResult has no token field, so _from_plan_execute() in src/evaluation/metrics.py always builds OpsMetrics with tokens_in=tokens_out=0, and _estimate_cost(model, 0, 0) short-circuits to None. Every offline evaluation report shows zero tokens and no cost for the plan-execute agent specifically; every SDK-based runner reports correctly because their Trajectory dataclass carries per-turn token fields.

Fix Details

Thread the meter's totals through as two new optional keyword arguments on persist_trajectory() (tokens_in, tokens_out), written onto the record only when given, so the six other runners' calls and their persisted shape are unchanged. PlanExecuteRunner.run() now passes self._meter.input_tokens/output_tokens. metrics_from_trajectory() reads them back off the record (getattr(..., None) or 0, so older persisted files without these keys still parse) and passes them into _from_plan_execute(), which now populates OpsMetrics.tokens_in/tokens_out and the cost estimate instead of hardcoding zeros.

Impact on Benchmarking

• Baseline change: tokens_in, tokens_out and est_cost_usd for plan-execute scenario results move from always 0/None to the run's real usage. aggregate_ops()'s existing sums (tokens_in_total, tokens_out_total, est_cost_usd_total) are the only readers of these fields elsewhere in the codebase, and they already treat a None cost as excluded, so no other code path assumed the old zeros.

Related Issues

• Fixes: plan-execute runner's metered token usage never reaches persisted trajectories, so evaluation reports 0 tokens / no cost #499

Verification Steps

• New regression tests (test_orchestrator_persists_token_usage_alongside_trajectory, test_persist_includes_run_level_tokens_when_given, test_plan_execute_list_trajectory_reads_run_level_tokens) fail against unmodified main and pass on this branch; ran in a clean python:3.12-slim container with uv sync, __file__-checked against /repo/src/....
• Full suite for the three touched modules (src/agent/tests/, src/evaluation/tests/, src/observability/tests/): 187 passed, 8 pre-existing failures unrelated to this change (missing google.protobuf in test_file_exporter.py, an unrelated scorer issue in test_static_json_scorer.py), same 8 failing identically on unmodified main.
• coverage run/report confirms every changed line in the three touched files is covered.
• Not checked: no live AGENT_TRAJECTORY_DIR/evaluation-CLI run against a real LLM backend; the repro is the unit/integration test suite above, not a manual end-to-end run.

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules. (repo has no Ruff config or dependency; nothing to run)
• I have signed off my commits (DCO).