ruvector 0.2.40

fix(security): eliminate Rust and npm dependency advisories (#739)

Refresh all committed Rust locks, eliminate actionable RustSec findings, make the npm graph reproducible and audit-clean, retire vulnerable optional backends, harden RuVocal production dependencies, and repair the affected publishable packages.

Closes #736.

fix(rvf): preserve one-based and sparse IDs in COW branches (#741)

Size COW membership filters by the highest vector ID, bound dense membership allocation, add hostile-capacity regressions, and stage corrected RVF runtime/native/SDK releases.

fix: harden graph and postgres data integrity (#738)

Fix graph replication serialization, PostgreSQL HNSW concurrent-build safety and ef_search propagation, and SONA dimension/statistics correctness. Harden benchmark and CI execution, including cancellation of superseded runs.

Closes #727.
Closes #728.
Closes #729.
Closes #732.

fix: harden graph and postgres data integrity (#738)

Fix graph replication serialization, PostgreSQL HNSW concurrent-build safety and ef_search propagation, and SONA dimension/statistics correctness. Harden benchmark and CI execution, including cancellation of superseded runs.

Closes #727.
Closes #728.
Closes #729.
Closes #732.

chore(release): bump ruvector-core workspace to 2.3.0, ruvector-exten…

…sions to 0.1.2 (#685)

Version bumps to publish the recently-merged Lattice embeddings work:
- Workspace (ruvector-core + 25 sibling crates): 2.2.3 -> 2.3.0 (minor --
  new opt-in `lattice-embeddings` feature is additive, no default-build
  changes). crates.io already has 2.2.3 published; this unblocks a new
  ruvector-core release.
- ruvector-extensions (npm): 0.1.0 -> 0.1.2. Local package.json was
  stale -- the registry already had 0.1.0 and 0.1.1 published
  out-of-band, so 0.1.2 is the next available version.

Note: neither package currently has a working automated publish
pipeline (release.yml, the documented Rust release pipeline, fails
with a workflow-file startup_failure; ruvector-extensions has never
had CI publish coverage at all) -- publishing this round by hand with
the equivalent gates run locally (full test suite incl. the
lattice-embeddings feature, npm test suite via tsx since `npm test`'s
plain glob silently only runs one of five test files, dry-run for both
registries) in lieu of CI.

chore(release): bump ruvector-core workspace to 2.3.0, ruvector-exten…

…sions to 0.1.2 (#685)

Version bumps to publish the recently-merged Lattice embeddings work:
- Workspace (ruvector-core + 25 sibling crates): 2.2.3 -> 2.3.0 (minor --
  new opt-in `lattice-embeddings` feature is additive, no default-build
  changes). crates.io already has 2.2.3 published; this unblocks a new
  ruvector-core release.
- ruvector-extensions (npm): 0.1.0 -> 0.1.2. Local package.json was
  stale -- the registry already had 0.1.0 and 0.1.1 published
  out-of-band, so 0.1.2 is the next available version.

Note: neither package currently has a working automated publish
pipeline (release.yml, the documented Rust release pipeline, fails
with a workflow-file startup_failure; ruvector-extensions has never
had CI publish coverage at all) -- publishing this round by hand with
the equivalent gates run locally (full test suite incl. the
lattice-embeddings feature, npm test suite via tsx since `npm test`'s
plain glob silently only runs one of five test files, dry-run for both
registries) in lieu of CI.

chore(release): @ruvector/rvf 0.2.3, ruvector 0.2.34 (#641 fixes)

Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_017sXWL4ox5bhC86FYwJpmyK

ruvector 0.2.33 — MCP tool-count accuracy + optimizer module + contra…

…ct-test skip fix

ruvllm 2.3.0 — GPU optimization sweep (ADR-258)

New model: RDT (Recurrent-Depth Transformer) + OpenMythos recurrent-depth model
in Rust/Candle with Adaptive Computation Time halting.

Key optimizations:
- Vectorized ACT halting: 4–21× GPU prefill speedup (RTX 5080)
- candle 0.9 + cudarc 0.19: CUDA 13.0 native (RTX 5080/SM 12.0)
- KV cache pre-allocation via scatter_set: O(N²)→O(N) decode bandwidth
- On-device argmax: 128 KB → 4 bytes per greedy decode step
- GPU top-k sort for temperature sampling: 128 KB → ~320 bytes per step
- Fused CUDA ACT kernel (--features fused-act) via cudarc nvrtc
- Zero-copy tensor pointer path: candle public API, no patching needed
- True per-token streaming (generate_stream_sampled callback)
- RDT: generate_sampled + generate_stream_sampled with GPU top-k sampling

Benchmarks (RTX 5080 BF16):
  Prefill seq=512: ~91 ms CPU → 3.54 ms CUDA (21× speedup)
  Decode prompt32/gen16: 73.4 ms CPU → 44.3 ms CUDA (-40%)