IPR fixed-work diagnosis

Historical baseline: the findings below describe 0c3a573, before production chromosome uniqueness. The elite-diversity correction now prevents duplicate admission and skips identical effective parents. The original reference remains unchanged; separate post-fix work records are maintained. Unchanged-block pruning subsequently reduces the direct one-differing-block case from 55 to two evaluations, retaining virtual decisions.

The empty-pool bootstrap enables standalone IPR to start relinking. It does not establish adequate population diversity, efficient path work or solution quality. This diagnostic isolates those questions without another timed quality benchmark.

Reproduce

From the repository root:

uv run --locked python -m benchmarks.algorithms.ipr_diagnosis

Defaults: 32 coordinates, seeds 200–202, sphere/threshold/flat objectives and 1,024 completed objective evaluations per case. The 36-case matrix crosses:

• Ordinary two-seed bootstrap versus ten externally initialized elites.
• Production pool admission versus a diagnostic-only exact-vector duplicate rejection control.

Every external initialization evaluation is charged. The ten-elite arm also pays for production IPR's unconditional initial draw, so its split is 11 seed + 1,013 path evaluations, versus 2 + 1,022 for ordinary bootstrap. First-two-seed fingerprints match across all four arms. A final path may be interrupted by the cap; pairs_started includes an entered but incomplete path. SearchSession still retains the best completed candidate and restores callbacks.

The exact-vector control is isolated and single-threaded; it is not a production pool implementation or API option. It retains different chromosomes with tied costs. Do not substitute objective-value equality for genotype equality.

benchmarks/references/ipr-diagnosis-reference.jsonl records the source revision, solver/probe source SHA256, Python/NumPy versions, one direct path case and all 36 count traces. No wall/CPU timings are collected. The public C++/Python Knapsack benchmark, worktree, cache, environment and output are not accessed by this script.

Findings on this matrix

1. Pair formation works, but occupancy is not diversity

All ordinary two-seed runs perform actual path evaluations, including genuine intermediate vectors. Nevertheless, all three sphere runs end with ten pool entries containing only one distinct vector. Production pool admission accepts exact clones when capacity remains or their score beats the worst entry; IPR can reinsert an unchanged best path/start repeatedly.

Across the three ordinary-bootstrap seeds:

Objective	Distinct final vectors / ten entries	Identical-parent pairs / 48 started
Sphere	1, 1, 1	25 / 48
Threshold	2, 2, 2	13 / 48
Flat	2, 2, 2	24 / 48

The exact-vector control eliminates identical-parent selections in this matrix. It does not improve the best objective in any of the nine two-seed cases. Removing wasted work is not by itself evidence of improved quality.

2. Unchanged blocks are evaluated repeatedly

An independent direct production-kernel case has 20 coordinates and two distinct parents differing in exactly the first two-coordinate block. The relinker makes 55 objective calls for only two distinct vectors: ten start evaluations and 45 guide evaluations. No genuine intermediate vector exists for that pair.

The implementation visits every initially unfixed block, even when its values already equal the guide. It takes num_blocks-1 steps, not a count based on actual parent differences. Therefore shared blocks can also make the full guide reachable early; “always stops before the guide” is not a valid invariant.

3. More starting diversity helps these objectives, not universally

At the same total evaluation cap, ten initial elites improve the final sphere and threshold scores over ordinary bootstrap in all six numerical comparisons here; flat scores remain tied. Production identical-parent selections fall to 3/144 started pairs across those nine larger-population cases, but some pools still lose diversity. The exact-vector control combined with ten seeds gives mixed results relative to the production ten-seed arm: it can finish better, equal or worse.

All 36 traces produce zero path candidates outside the seed coordinate alleles. Internal relinking only copies seed-derived blocks; it cannot invent new coordinate values. A larger seed population changes the available search space, not merely startup overhead. These small separable synthetic objectives are deliberately sensitive to that restriction; they are not representative all-domain evidence.

Reference-source audit

Read-only reference: RKO_Cpp_v2.0 revision 126e35873e50dc308795b58b8c75d37f8de4c857.

• Program/src/MH/IPR.h sets direction=1 every outer iteration. Its external branch exists but is inactive; missing external search is not an established explanation of this comparison.
• Program/src/Method.h::UpdatePoolSolutions rejects already-present objective values, while the historical Python policy admitted exact vector clones (now corrected separately). Copying that policy would also reject distinct tied vectors; it is not an automatic safe fix.
• The C++ driver initializes its population before invoking IPR and perturbs some tied-cost initial entries. The ten-random-elite arm here is not that complete initialization policy or a compiled C++ execution.
• C++ difference flags use coordinate modulo block count while swaps use contiguous blocks, and its block count truncates. Python safely covers partial blocks; do not reproduce reference indexing/edge defects to claim parity.

Subsequent corrections and remaining scope

Exact duplicate admission/identical-parent selection is now addressed by the linked production correction, retaining valid tied vectors and bounded deficient-pair behavior. The linked unchanged-block correction now reuses known current costs for representation-identical moves, retaining virtual steps and deterministic path/tie results. Stochastic callback counts change explicitly. A full-population policy still requires controlled evidence; this diagnosis does not choose one. The full-driver startup contract details the reference pool fill, bounded cost-tie pass, default empty startup and distinct clock boundaries. The opt-in prefill_pool=True phase now supplies shared context with explicit startup accounting; it is not a demonstrated IPR quality/feasibility correction.

The original diagnosis itself changed no production transition/admission policy; the linked subsequent correction is a separately identified behavior change. It does not establish Knapsack feasibility after the bootstrap, end-to-end speedup, full C++ parity or a reason to rewrite the core in Rust. The ongoing pinned benchmark remains a historical baseline and must not be relabeled as current IPR.