NC: diversity-collapse guard — few-shot sampling + hive + 6th safety invariant

[mercurialsolo]

Problem

Both convergence loops are monocultures by construction:

• retrieve_similar deterministically returns the top-k nearest examples — the same examples win every time.
• hive merger.rs resolves peer-vs-peer by confidence × evidence — high-evidence beliefs dominate and crowd out exploration.

Weng's post flags diversity collapse (bottleneck #4) as especially critical for open-ended, evolutionary loops. This is the one NCT failure mode (canalization / runaway reciprocal causation toward a narrow attractor) that docs/niche-construction.md does not currently address in its Safety Invariants.

Proposal

• Inject diversity into few-shot sampling: sample from the top-N by similarity or dedupe by cluster rather than always returning top-k.
• Give the needs model (direction.rs, NC Phase 0: Sense only — direction.rs + charter.toml (ranked needs backlog, no actions) #359) an exploration term so it doesn't only ever service the loudest friction signal.
• Add a sixth Safety Invariant to docs/niche-construction.md covering diversity preservation.

Why it matters

Prevents the auto-improving loop from collapsing onto a single locally-optimal pattern. Longer-horizon than the near-term wins but architecturally load-bearing before Tier 2/3.

Part of epic #358. Doc change to docs/niche-construction.md.