The planner and explain()

explain() is your single best window into runtime behavior.

auto plan = collection.explain_query(query, /*k=*/10, filter);
// plan.strategy           — QueryStrategy enum
// plan.candidate_count    — narrowed candidate set size
// plan.metadata_accelerated
// plan.gpu_index
// plan.index_type_name

• Selective filter + accelerated metadata → candidate_count drops by orders of magnitude
• GPU built in → gpu_index = true; the dispatch goes through GpuBackend
• No accelerator, flat index → planner picks brute-force; expect linear-in-N latency

Treat strategy + candidate_count as the two telemetry numbers you actually emit per query.