crates-2026-09-26

**Plain-language summary:** https://gist.github.com/ruvnet/dd8a2246f104f69bd7d755637c5d77cd

## crates.io: 120 of 122 crates published
- Every publishable library crate is now on crates.io at `main`'s version (the remaining 2 are example apps). #1020 adds version requirements to 60 path dependencies. #1028 publishes per-crate patches: `ruvector-core` 2.3.1, `ruvector-graph` 2.3.2, `rvf-types` 0.2.2, `rvf-wire` 0.1.1.
- `ruvector-mincut` **2.3.1**: `partition()` and the witness hash are deterministic and non-degenerate (#979, #1021).
- Note: the published `ruvector-core` links the upstream `hnsw_rs`, so the #773 fix in `patches/hnsw_rs` is not in crates.io builds yet.

## pi.ruv.io brain server (deployed: `ruvbrain-00210-zsg`)
- #1024: the graph rebuild runs off-lock and swaps its result in atomically. In a benchmark at production scale (60k nodes), search latency during a rebuild went from a 96 s wait to a 31 ms p99, and the rebuilt graph is bit-identical.
- #1023: `/v1/pipeline/inject*` and `/internal/*` writes are authenticated; UTF-8 truncation panics are fixed; ranked search and the sparsifier no longer hold the graph lock (ADR-349).
- #1022: writes and training are gated until Firestore hydration completes, and `/v1/status` reports `hydrating`. Also includes a rebuildable Dockerfile.
- #1025: the SSE image builds again and sends `BRAIN_SYSTEM_KEY`.
- Cloud Run now runs with min 1 instance and CPU always on, so hydration takes about 1 minute instead of 30–50.

## ESP32 decision firmware (#1018)
- INT8 decisions on ESP32, S2, S3, C2, C3, C6, H2 and P4, with serial OTA and automatic rollback.
- Measured on a physical C6: 21× faster host round trip, and 1000/1000 decisions match the reference.
- Images: `esp32-decision-v0.1.0` and `esp32-decision-v0.2.0`.

## Agent memory (#989)
- Ed25519 `WitnessSigner` for the TARL ledger. Verification requires full coverage, a sealed tail and a signed anchor, and signs a SHA-256 digest (ADR-347).

## OpenJev pipeline (#1029, ADR-008)
- `crates/ruvector-typesafe-train`: fine-tunes the encoder in Rust/candle, transplants the weights into ONNX with `ort` parity, and aborts training on any leakage.
- `crates/ruvector-gpu-runner`: a budget-capped vast.ai runner with auto-destroy.
- `Engine::train_with_calibration` in `ruvector-typesafe-core`, plus the Jev-parity bench (`vs-jev.mjs`, `--model-dir`, `--no-test`).
- **v1 frozen-test result (seed 2):**
  - department: 93.3% vs Jev 85.3% (non-inferior)
  - urgent: 92.7% vs 61.3% (superior)
  - frustration: 86.0% vs 67.3% (superior)
  - p95: 10 ms
  - ECE: 0.0543 (gate is 0.05, fail)
- The pre-registered primary claim is not met, so **the model is not published**. v2 will be judged on fresh held-out data.

🤖 Generated with [claude-flow](https://github.com/ruvnet/claude-flow)