feat(metrics): populate token usage, add latency percentiles, record …

[sandeepyadav1478]

…max_workers

The result schema declares GenerationData.prompt_tokens/completion_tokens, but the model is imported by two runners and never constructed, so every published result carries no token data. Both SDKs already report exact usage on each response; llm_client discarded it at the return statement. No new dependency and no local re-tokenization is needed.

• llm_client: accumulate SDK-reported usage per eval item. Accounting is held in a ContextVar rather than on the client because one LLMClient is shared by max_workers concurrent items; asyncio copies the context at Task creation, so items cannot bill tokens to each other. Handles both OpenAI (prompt/completion_tokens) and Anthropic (input/output_tokens) field names.
• locomo: record answerer tokens per cutoff, separately from whole-item spend. Context-per-query and cost-to-run-the-benchmark are different quantities and folding the judge into the first would misstate it.
• metrics: compute_cost_metrics() reports p50/p95 latency, not just a mean. Retrieval latency is right-skewed, and a mean is not comparable to a published p50. It reads search_latency_ms from both locations the benchmarks use (locomo and beam write it at item top level, longmemeval nests it under "retrieval"), so it also works on already-published results.
• metadata: record max_workers. Latency was measured under concurrency that the result files never stated, which left every published figure uninterpretable.

selfcheck.py covers provider field names, concurrent isolation, untracked calls, percentiles and back-compat with results that predate this change. The repo has no test framework, so it is assert-based and runs on plain python.