v3.25.0

## ruflo 3.25.0 — Anti-overfitting proofs (+ an inactive Lattice embedder seam)

### Anti-overfitting (#2580) — real, shipped
- **Frozen public human-labeled eval set** (`.claude/eval/human-relevance-frozen-v1.json`) — hash-pinned, tamper-evident; the red/blue anchor the flywheel must never regress.
- **Per-generation human-relevance deltas** in every receipt (`deltas.humanRelevance`) — so "self-retrieval up, human relevance flat → OVERFITTING" is **visible** in `flywheel status`, not hidden.
- **Clean-room replay acceptance test** (`scripts/replay-generation.mjs`) — replay a promoted generation from its receipt alone: identical hashes, re-run `accept/v1+sig`, **offline** (network trapped). Wired into CI.

### ⚠️ Correction — the "Lattice WASM embedder tier" (#2581) is INACTIVE
The earlier notes for this release overstated it. **There is no `@ruvector/lattice-wasm` package** (npm 404), and no "Lattice" embedder package exists in the ruvector ecosystem. What shipped is a **fail-closed, optional adapter seam** that dynamically imports an (env-configurable) package name and **degrades to ruvector-ONNX → hash when it is absent — which it always is today.** It is therefore **dormant / no-op** and causes no regression, but it is NOT a working multi-model embedder. The models referenced (bge / qwen3-0.6b / paraphrase-miniLM) exist only in a Rust ONNX example, not a publishable package. A follow-up will either remove the seam or wire it to a real embedder.

**Net for users:** upgrade for the anti-overfitting proofs; embeddings behave exactly as in 3.24.0 (ruvector ONNX where available).

Additive · backwards-compatible · fail-closed. `npx ruflo@latest`.