Prepare larger MS MARCO scale tier

[SonAIengine]

Summary

• add tag-aware node lookup for Memory/SQLite/Composite backends and use it for QueryAnchor category loading
• add side-by-side MS MARCO large-shard options (--large-output-suffix, --msmarco-path) so 1M and 5M manifests can coexist
• document the 1M reuse improvement and local 5M MS MARCO shard

Results

• 1M MS MARCO persistent reuse: 50-query search improved from 9.1s to 7.5s with unchanged MRR@10 0.479 and Hit@10 31/50
• First QueryAnchor category load on the 1M DB: about 0.218s vs previous roughly 1.7-2.0s
• Created local gitignored 5M shard: tests/benchmark/data/msmarco_passage_5m.json + .corpus.jsonl (5,000,000 rows)

Tests

• uv run ruff check src/synaptic/extensions/query_anchor.py src/synaptic/backends/memory.py src/synaptic/backends/sqlite.py src/synaptic/backends/composite.py examples/ablation/download_benchmarks.py examples/ablation/run_tier1_benchmarks.py tests/test_query_anchor.py tests/test_backend_memory.py tests/test_backend_sqlite.py tests/test_download_benchmarks.py tests/test_tier1_benchmarks.py
• uv run pytest tests/test_download_benchmarks.py tests/test_tier1_benchmarks.py tests/test_query_anchor.py tests/test_backend_memory.py::TestMemoryBackendNodes::test_list_nodes_by_tag_filters_kind_and_limit tests/test_backend_sqlite.py::TestListNodesByTag -q
• uv run python examples/ablation/run_tier1_benchmarks.py --only msmarco --subset 50 --use-sqlite-graph --sqlite-db-path tests/benchmark/data/msmarco_1m.db --reuse-sqlite-db --corpus-limit 1000000 --max-search-sec 30 --min-hit-rate-at-10 0.0 --min-mrr 0.0