自适应基数树索引服务

Primary  : pri\x01<repo>\x01<rel_path>\x01<line5>:<col3>:<kind>\x01<name>
Inverted : sym\x01<kind>\x01<name>\x01<repo>\x01<rel_path>:<line>

cargo build --release                    # binary: target/release/blazing_art_mcp
cargo test                               # 13 tests (7 unit + 6 integration)
cargo clippy --all-targets -- -D warnings

# Default Criterion suite (point lookup, prefix scan, bulk insert,
# scaling sweep at 10k/100k/1M, real-data tier 1 = own src/):
cargo bench --bench art_benchmarks

# Add the 5M scaling tier (requires ~2.5 GB peak RSS):
BLAZING_ART_BENCH_SCALE=full cargo bench --bench art_benchmarks

# Real-data tiers 2 + 3 (require local repos):
CPYTHON_LIB_PATH=$HOME/cpython/Lib       cargo bench --bench art_benchmarks
TYPESCRIPT_SRC_PATH=$HOME/typescript/src cargo bench --bench art_benchmarks

# Per-backend isolated RSS measurement (run all three for the table in BENCHMARKS.md):
BLAZING_ART_RSS_TARGET=art      cargo bench --bench memory_rss -- --nocapture
BLAZING_ART_RSS_TARGET=btreemap cargo bench --bench memory_rss -- --nocapture
BLAZING_ART_RSS_TARGET=hashmap  cargo bench --bench memory_rss -- --nocapture

./target/release/blazing_art_mcp \
  --entities data/entities.json \
  --events data/events.json

./target/release/blazing_art_mcp --http 127.0.0.1:4242
# Then visit http://127.0.0.1:4242/ for the demo UI,
# or POST JSON-RPC to http://127.0.0.1:4242/mcp.
# /health returns {"status":"ok"}.

./test-mcp.sh                            # exercises all four legacy tools over stdio
python3 /tmp/blazing_art_demo.py         # exercises ingestRepo / findSymbols / deleteRepo over HTTP
                                         # (requires server running with --http 127.0.0.1:4242)

{
  "mcpServers": {
    "blazing-art": {
      "command": "/abs/path/to/blazing-art-mcp/target/release/blazing_art_mcp",
      "args": [
        "--entities", "/abs/path/to/blazing-art-mcp/data/entities.json",
        "--events",   "/abs/path/to/blazing-art-mcp/data/events.json"
      ]
    }
  }
}

{
  "mcpServers": {
    "blazing-art": {
      "transport": "http",
      "url": "http://127.0.0.1:4242/mcp"
    }
  }
}

kiro-cli mcp add \
    --name blazing-art \
    --scope workspace \
    --command /abs/path/to/blazing-art-mcp/target/release/blazing_art_mcp \
    --args "--entities,/abs/path/to/data/entities.json,--events,/abs/path/to/data/events.json"

# Optional but recommended for slow first-ingest:
kiro-cli settings mcp.noInteractiveTimeout 60000

[mcp_servers.blazing_art]
command = "/abs/path/to/blazing-art-mcp/target/release/blazing_art_mcp"
args = ["--entities", "/abs/path/to/data/entities.json",
        "--events",   "/abs/path/to/data/events.json"]

async function lookup(name) {
  const r = await fetch("http://127.0.0.1:4242/mcp", {
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify({
      jsonrpc: "2.0", id: 1, method: "tools/call",
      params: { name: "lookupEntity", arguments: { name } },
    }),
  });
  return r.json();
}

# Layer 1 — unit + e2e + property/fuzz (fast; runs on every push).
cargo test --all-targets
#   ↳ 30 tests:
#      15 unit (memory + ingest)
#       8 e2e (sample-repo + multilang fixtures)
#       3 proptest (oracle vs BTreeMap, key round-trip, concurrent chaos)
#       4 handcurated_schema (gold-set sanity)

# Layer 2 — gold-set retrieval eval (runs on every push; gates merges).
cargo run --release --bin eval_goldset
bash scripts/check_eval_threshold.sh
#   ↳ Recall@1 / @5 / @20, MRR, p50/p99, per-category breakdowns;
#     fails if any metric below floors in eval/threshold.json.

# Layer 3 — live coding-agent harnesses (manual; cost real LLM API calls).
bash tests/harness/run_claude_code.sh    # skips if `claude` not installed
bash tests/harness/run_kiro_cli.sh       # skips if `kiro-cli` not installed
#   ↳ Drives each agent through 10 scripted prompts under
#     treatment (MCP enabled) vs control (no MCP); regex assertions
#     in tests/harness/asserts.txt require treatment to match the
#     correct answer AND control to MISS it.

src/
├── lib.rs                       # module declarations
├── main.rs                      # CLI + transport dispatch
├── memory.rs                    # ART-backed Memory (entities, events, symbols) + 7 unit tests
├── protocol.rs                  # JSON-RPC 2.0 dispatch (transport-agnostic)
├── ingest.rs                    # tree-sitter walker + key encoder
└── transport/
    ├── mod.rs
    ├── stdio.rs                 # newline-delimited JSON-RPC over stdin/stdout
    └── http.rs                  # axum: POST /mcp, GET /mcp (405), GET /health, GET /
benches/
├── art_benchmarks.rs            # Criterion: point/prefix/insert + scaling sweep + real-data
├── memory_rss.rs                # standalone (harness=false), macOS-only via getrusage
└── shared/keygen.rs             # synthetic key generator + index builders, shared by both
tests/
├── ingest_e2e.rs                # 6 integration tests against the fixture
└── fixtures/sample-repo/src/    # 3-file Rust fixture (10 known declarations)
static/index.html                # vanilla-JS web demo (no framework)
data/                            # 2,758 entities + 896 events (AI/ML papers from Qdrant)
examples/                        # smaller sample dataset
BENCHMARKS.md                    # publication-ready numbers + methodology
CLAUDE.md                        # docs-vs-reality table for future sessions