Model behavioral state diversity for specialist frontier selection

[chaoz23]

Product outcome

Choose frontier states by mechanically meaningful future-behavior diversity rather than path text, raw variable uniqueness, or geographic-style distance that Ink stories do not possess.

Parent: #97. Related: #53 shared frontiers, #88 authored outcomes, #98 memory bounds, #105 marginal evidence.

Candidate signatures

Keep separate, inspectable dimensions:

• current authored knot/container;
• offered choice-set identity;
• relevant variable snapshot/transition rarity;
• satisfied/unsatisfied gate predicates;
• loop/hub phase;
• turn/random sensitivity;
• reached goals/assertion-boundary proximity;
• observed successor signatures where available.

Do not collapse dimensions into one opaque semantic score. Similarity is a scheduling hypothesis, not proof of future equivalence.

Acceptance criteria

• Versioned deterministic signatures are bounded and privacy-safe.
• Exact state deduplication remains separate from approximate behavioral grouping.
• Frontier selection preserves representatives across dimensions and protected long-tail probes.
• Adversarial fixtures include same variables/different futures, different histories/same future, decoy variable churn, duplicate choice labels, and a rare gate transition.
• Compaction cannot erase all representatives of a protected family.
• Reports explain which diversity dimension retained or evicted a state.
• Matched Build the search-strategy promotion benchmark harness #56 tests report critical evidence, authored coverage, terminal variants, memory, and time separately.
• No claim of semantic equivalence or default promotion without broad evidence.