Add comprehensive tests, benchmarks, and bug fixes

.github/workflows/release.yml

@@ -88,7 +88,7 @@ jobs:
           sleep 30
           publish llmem-index
           sleep 30
-          publish llmem-cli
+          publish llmem
           sleep 30
           publish llmem-server
 
@@ -150,7 +150,7 @@ jobs:
           if [ "${{ matrix.cross }}" = "true" ]; then
             BUILD_CMD="cross"
           fi
-          $BUILD_CMD build --release --target ${{ matrix.target }} -p llmem-cli
+          $BUILD_CMD build --release --target ${{ matrix.target }} -p llmem
           $BUILD_CMD build --release --target ${{ matrix.target }} -p llmem-server
         shell: bash
 

CONTRIBUTING.md

@@ -31,6 +31,39 @@ Conventional commits via `sr commit`:
 
 Scope by crate: `feat(core):`, `feat(cli):`, `feat(server):`, `docs(spec):`.
 
+## Testing
+
+### Unit and integration tests
+
+Run the full test suite with:
+
+```bash
+cargo test --workspace
+# or
+just test
+```
+
+Tests are colocated with their source in `#[cfg(test)]` modules, except for the CLI which uses a dedicated file at `crates/llmem-cli/tests/integration.rs`.
+
+| Crate | Tests | What is covered |
+|-------|-------|-----------------|
+| `llmem-core` | 21 | Config TOML roundtrip and dot-notation get/set; embedding store binary format and hash-based change detection; inbox capacity eviction and JSON persistence; MEMORY.md index parsing, search, and save; memory file frontmatter parsing and markdown roundtrip; `FileBackend` store/get/remove/list |
+| `llmem-index` | 16 | HNSW insert, remove, save/load, and recall ≥ 90% on 200 vectors; IVF-Flat insert, remove, save/load, and recall ≥ 85% on 200 vectors; cosine similarity, dot product, L2 distance, and normalization; tree-sitter Rust chunking and language-extension mapping |
+| `llmem-quant` | 39 | Lloyd-Max codebook structure and scalar quantize/dequantize at 1–4 bits; bit-packing roundtrip for 1–4 bits including non-byte-aligned counts; `TurboQuantMse` roundtrip MSE, norm preservation, zero vector, and empirical MSE against theoretical bound; `TurboQuantProd` unbiased inner-product property and fast estimate; QJL determinism and unbiased inner-product property; rotation orthogonality, forward/inverse roundtrip, and norm preservation; compressed embedding store save/load for MSE and Prod variants |
+| `llmem-server` | 6 | HTTP handler tests using `tower::ServiceExt::oneshot` (no network): `/health` status and version fields; `/search` empty state, text matching, `top_k` truncation, and level filtering; `/reload` response |
+| `llmem-cli` | 15 | End-to-end CLI tests that spawn the binary as a subprocess: `init`, `memorize` (with type, name, and stdin JSON), `note` (inbox capacity enforcement), `remember` (match and no-match), `reflect`, `consolidate` (dry-run and real), `forget` (success and nonexistent), `ctx switch`/`show`, `config init`/`get`/`set` |
+
+### End-to-end validation
+
+`scripts/validate.sh` runs every CLI command and the RAG server against a clean, isolated `$HOME` directory and reports pass/fail with timing for each operation. It requires release binaries built beforehand:
+
+```bash
+cargo build --release
+bash scripts/validate.sh
+```
+
+The script covers all command groups in order — `init`, `config`, `memorize`, `note`, `remember`, `reflect`, `learn`, `consolidate`, `forget`, `ctx`, and the server endpoints (`/health`, `/search`, `/reload`) — and exits with the number of failures as its exit code.
+
 ## Pull Requests
 
 1. Fork the repository