Chromosome-unique elite pools

SolutionPool retains at most one entry per coordinate-equal chromosome. This rule is applied inside the existing insertion lock, including manager-hosted pools. No proxy API, public return shape, dependency or configuration flag was added.

Admission and repeated observations

• Different chromosomes with tied objectives remain eligible. This is not objective-value deduplication, unlike the audited C++ update rule.
• Equal or worse observations of a stored chromosome are ignored.
• A strictly better observation replaces that chromosome's stored cost, source and owned keys, then restores ordinary ascending-cost ordering. It does not evict a different elite just to retain another copy of the same vector.
• Novel chromosomes retain the existing capacity/worst-cost rule and stable ties.
• Validation still happens before duplicate detection, including for invalid repeated observations. Inserted and returned arrays remain detached.

Equality uses the original finite coordinate values, without float packing, rounding, tolerance or objective-based keys. Adjacent representable values, different large integers and different lengths remain distinct; numerically equal integers/ floats and signed zeros compare equal. Original scalar representation is retained.

This is not a fitness cache. Decoder/objective callbacks still execute and count normally. If a stochastic callback reports a better cost at the same chromosome, that improved observation remains usable; the pool keeps its best observation, not an average or uncertainty model. Callers remain responsible for their objective semantics and stochastic state.

IPR and custom pools

Stored uniqueness prevents ordinary local/managed IPR from sampling two copies of one chromosome. IPR also checks the converted binary64 parent arrays before relinking: if they are equal, it uses the existing bounded sleep/backoff instead. That covers custom providers returning clones and distinct large integers collapsing to the same effective search coordinates. Custom pools still need two usable parents to support search; they are not silently rewritten or endlessly reseeded.

The existing startup rule still attempts at most one extra seed. If both random draws are identical and only one elite survives, IPR backs off within its ordinary stop/cancellation checks. No unbounded “retry until distinct” loop is introduced.

Validation and work evidence

Local tests exercise original failing clone admission, valid tied vectors, improved/ worse observations, detached inputs, exact numerical identity, invalid duplicates and concurrent writers. Manager tests exercise authoritative RPC admission and pair retrieval. IPR tests cover custom effective clones and repeated seed draws; its prefilled NumPy initialization order remains unchanged, while later pair-draw traces reflect the corrected population occupancy.

benchmarks/references/ipr-unique-elites-reference.jsonl repeats the bounded 36-case diagnostic from the historical diagnosis, at 1,024 completed evaluations per case, with all initialization charged. All 36 cases have unique retained vectors and zero identical-parent selections. Production traces match the isolated exact-vector admission control in those deterministic workloads. The older references/ipr-diagnosis-reference.jsonl remains the identified pre-fix baseline.

The two-seed scores remain unchanged in that matrix; ten-seed scores change in mixed directions relative to old production admission. No general quality, feasibility or end-to-end speed improvement is established. A subsequent unchanged-block correction now reduces the controlled one-differing-block kernel from 55 to two callbacks while retaining virtual path and deterministic tie behavior. Stochastic callback trajectories can change.

Migration and scope

This is a behavioral change for every algorithm sharing SolutionPool, not only IPR. Occupancy counts, guides, sampled ranks, pool-aware refinement and seeded/ timed trajectories can change. Query/serialization/proxy schemas remain unchanged. Pin 0c3a573 to reproduce the preceding duplicate-admitting policy; no legacy flag. The running Knapsack comparison pinned to 7ee93ca is not changed or reinterpreted.

The RNG-isolation replay tool deliberately supplies the current pool to both algorithm implementations. Such replay checks compare RNG ownership within the same pool context; they do not reproduce a historical application's old admission policy. Historical replay/performance files retain their original source scope.