Tune Qdrant 1M recall and latency profile

[CaspianG]

Goal: improve the service-backed 1M memory profile so the benchmark is useful for production readers.

Current checked-in baseline:

• Qdrant service, 1M memories
• recall@10: 0.506
• average query latency: 45.81 ms

Target:

• recall@10 >= 0.95 on the same generated benchmark profile
• p95/p99 latency reported, not only average latency
• reproducible command documented in docs/BENCHMARK_REPORT.md
• results committed as JSON under benchmarks/

Likely work:

• tune Qdrant collection/index parameters;
• compare vector normalization and scoring assumptions;
• tune candidate windows and reranking;
• add a clearer large-scale benchmark profile for CI/manual runs.

[CaspianG]

Progress landed in v2.2.3.

New checked-in artifacts:

• benchmarks/production_load_qdrant_100k_tuned_results.json
  • 100k vectors, Qdrant service
  • recall@10 1.000
  • avg 10.28 ms
  • p99 21.26 ms
• benchmarks/production_load_qdrant_1m_tuned_results.json
  • 1M vectors, Qdrant service with tuned search
  • recall@10 0.984
  • avg 116.80 ms
  • p99 209.28 ms
• benchmarks/production_load_qdrant_1m_ef_sweep_results.json
  • best recall point: hnsw_ef=2048
  • recall@10 0.977
  • avg 64.76 ms
  • p99 103.77 ms

Also added:

• p99/max latency reporting in the benchmark runner;
• Qdrant tuning env vars: WAVEMIND_QDRANT_HNSW_EF, WAVEMIND_QDRANT_EXACT, timeout controls;
• node/zone-loss and quorum reporting in cluster planning;
• RedisHotMemoryCache for multi-worker hot query caching;
• wavemind maintenance for scheduled TTL purge, consolidation, and index repair.

Status: 100k is now strong on the tested machine. 1M recall is strong, but p99 still needs one more tuning pass before claiming a stable sub-100 ms SLO.