Patch release shipping the darwin-core systems Darwin loop (PR #2481) and a critical tarball-bloat fix.

Darwin core-systems wins (PR #2481)

| Dim | Δ | Detail |
|-----|---|--------|
| skill-distillation | 0.4286 → 1.0 (+133% rel) | ADR-155 SKILL-DISCO bench: AND→OR predicate then promote-all. 3/7 → 7/7 successful traces promoted. Exceeds the SKILL-DISCO arXiv 2026 +22% target on this synthetic bench. |
| causal-graph hot path | 115ms → 3ms (-97%) | Two optimizations in v3/@claude-flow/cli/src/mcp-tools/agentdb-tools.ts: (a) EXISTS-probe early-exit on missing seed; (b) module-level dynamic-import cache for graph-query/pathfinder handlers. All 21 smoke tests still pass. |
| reasoning-bank | +0.0024 | scifact bm25 0.0→0.05 tie-breaker. Cumulative across this + prior BEIR Darwin loop: 0.5694 → 0.628 = +10.3% rel. |
| benchmark infra | new --only flag | `node scripts/benchmark-intelligence.mjs --only=` enables per-dimension measurement (saves multi-minute wall when only one dim needs scoring). |
| 2 baselines confirmed at-target | — | sona-adapt 0.0041 < 0.005 (WASM-opaque); moe-gate afterConfidence 0.88 > 0.85 (WASM-opaque). |

Tarball-bloat fix (critical)

The 3.14.4 publish initially failed with 413 Payload Too Large — the tarball had bloated from 77MB to 276MB / 46k files because the root `files` field swept in `.claude/worktrees/**` (3.3GB of Darwin workflow worktrees from this session).

Added explicit excludes in `package.json` files-pattern:

• `!.claude/worktrees/**`
• `!.claude/projects/**`
• `!.claude/sessions/**`
• `!.claude/scheduled_tasks.lock`

Result: 2.3 MB / 1096 files — slimmer than the 3.14.1 baseline (77MB / 14440 files). Published tarball now contains only what's actually needed (bin/, dist/, .claude-plugin/, .claude/agents/, .claude/commands/, etc.).

Install

```bash
npx [email protected]

or

npx @claude-flow/[email protected]
```

All three packages — @claude-flow/cli, claude-flow, ruflo — at 3.14.4 with consistent latest / alpha / v3alpha dist-tags.

Honest dimensions

• skill-distillation hit the synthetic bench ceiling (1.0/1.0) — real ALFWorld/WebArena eval would be the next validation
• HNSW recall@10 = 0.89 at N=5k vs documented 0.99 — flagged for separate investigation (NAPI doesn't honor efConstruction beyond cap)
• reasoning-bank's BEIR bench doesn't exercise the actual ReasoningBank subsystem — proper bench needed for next leap