iteration	commit	status	primary_metric_name	primary_metric_value	description
0	e7c0ed95a333	baseline	weighted_median_pair_plus_kspace_seconds	94.717	baseline
1	0157b0571dbb	rejected	weighted_median_pair_plus_kspace_seconds	94.369	Cache PPPM TIP4P virtual-site reconstruction and H-index lookup once per local oxygen per compute, reusing it across particle_map, make_rho, fieldforce, per-atom, and slab correction [Correctness passed, but 94.369 is only a 0.37% improvement over 94.717 and misses the benchmark contract's 2% minimum improvement threshold.]
2	2bbd33a175f2	correctness_failure	weighted_median_pair_plus_kspace_seconds	82.844	Backport the serial OPT hot loop into base lj/cut/tip4p/long so the benchmarked pair kernel uses branch-specialized eval paths and local f[i] accumulation. [Hard reject: despite a 12.54% speedup, candidate 2bbd33a175f failed TIP4P correctness with large force mismatches in both cases and broad thermo divergence in test-32r-long, so the baseline incumbent remains.; case test-32r-long etotal_series relative_delta exceeds threshold (8.92069e-05 > 1e-06); case test-32r-long pe_series relative_delta exceeds threshold (0.00218177 > 1e-06); case test-32r-long ke_series relative_delta exceeds threshold (0.00408985 > 1e-06); case test-32r-long temp_series relative_delta exceeds threshold (0.00408985 > 1e-06); case test-32r-long press_series relative_delta exceeds threshold (0.200262 > 1e-05); case test-32r-long forces_xyz abs_delta exceeds threshold (0.107617 > 1e-08); case test-32r-long forces_xyz rms_delta exceeds threshold (0.0276885 > 1e-08); case test-64r-short forces_xyz abs_delta exceeds threshold (1.44e-05 > 1e-08); case test-64r-short forces_xyz rms_delta exceeds threshold (1.44068e-06 > 1e-08)]
3	3a45586c8afa	accepted	weighted_median_pair_plus_kspace_seconds	81.885	Specialize `pair_lj_cut_tip4p_long` with compile-time energy/virial/table dispatch, a shared TIP4P site-cache helper, and separate oxygen/non-oxygen `i` hot paths while preserving the original force writes and `ev_tally_tip4p()` semantics. [Correctness passed with no guardrail errors, and the candidate improves the primary metric from 94.717 to 81.885, a 13.55% gain that clears the 2% acceptance bar.]
4	847eb8e5dd4b	rejected	weighted_median_pair_plus_kspace_seconds	81.985	Cache local PPPM TIP4P M-site reconstruction and 1D rho stencils in particle_map() so make_rho(), fieldforce, per-atom, and slab correction reuse the cached state instead of recomputing find_M()/compute_rho1d() [Correctness passed, but 81.985 is 0.12% slower than incumbent 81.885 and misses the 2% improvement bar.]
5	42d987d04a9c	rejected	weighted_median_pair_plus_kspace_seconds	81.298	Skip the extended TIP4P Coulomb precheck for exact non-oxygen/non-oxygen pairs in `pair_lj_cut_tip4p_long`, while keeping the incumbent oxygen/M-site path unchanged for any oxygen-containing interaction. [Correctness passed, but 81.298 is only a 0.72% improvement over incumbent 81.885 and misses the benchmark contract's 2% minimum improvement bar.]
6	89aa52c5901e	correctness_failure	weighted_median_pair_plus_kspace_seconds	79.639	Specialize `pair_lj_cut_tip4p_long` for the benchmark’s global-only tally path, fast-path exact non-water/non-water Coulomb pairs, and locally accumulate non-water `f[i]` updates while preserving the incumbent TIP4P oxygen/M-site force distribution [Hard reject: candidate 89aa52c5901e is 2.74% faster than incumbent 3a45586c8afa but fails benchmark correctness, with broad test-32r-long thermo divergence and forces_xyz mismatches in both cases.; case test-32r-long etotal_series relative_delta exceeds threshold (8.42997e-05 > 1e-06); case test-32r-long pe_series relative_delta exceeds threshold (0.00183969 > 1e-06); case test-32r-long ke_series relative_delta exceeds threshold (0.00345568 > 1e-06); case test-32r-long temp_series relative_delta exceeds threshold (0.00345568 > 1e-06); case test-32r-long press_series relative_delta exceeds threshold (0.222002 > 1e-05); case test-32r-long forces_xyz abs_delta exceeds threshold (0.105529 > 1e-08); case test-32r-long forces_xyz rms_delta exceeds threshold (0.0282935 > 1e-08); case test-64r-short forces_xyz abs_delta exceeds threshold (8.3e-06 > 1e-08); case test-64r-short forces_xyz rms_delta exceeds threshold (1.10787e-06 > 1e-08)]
7	c37498ebee61	rejected	weighted_median_pair_plus_kspace_seconds	81.445	Keep the exact H-H Coulomb cutoff fast path in `pair_lj_cut_tip4p_long`, and reuse the first oxygen `find_M()` result in `PPPMTIP4P::fieldforce_ik()` / `fieldforce_ad()` while hoisting constant alpha/slab factors out of the inner oxygen force-distribution branch [Correctness passed, but candidate c37498ebee61 improves the incumbent by only 0.54% (81.885 to 81.445), missing the 2% acceptance bar.]
8	7d859a5521df	rejected	weighted_median_pair_plus_kspace_seconds	85.841	Specialize `pair_lj_cut_tip4p_long` Coulomb handling by interaction class (`O-O`, `O-H`, `H-O`, and exact non-oxygen/non-oxygen) via a templated helper, preserving incumbent TIP4P force distribution and `ev_tally_tip4p()` semantics while retaining the safe exact non-water cutoff pruning. [Correctness passed, but candidate 7d859a5521df regressed the incumbent primary metric from 81.885 to 85.841, missing the 2% improvement bar.]
9	0356a6e5d35a	rejected	weighted_median_pair_plus_kspace_seconds	80.674	Tighten `pair_lj_cut_tip4p_long` Coulomb prechecks only for exact `H-H` and actual-current-step `O(i)-H(j)` reach tests, while keeping all oxygen-`j` paths plus the incumbent TIP4P force/tally write order unchanged. [Correctness passed, but 80.674 is only a 1.48% improvement over incumbent 81.885 and misses the 2% acceptance bar.]
10	25819367f233	rejected	weighted_median_pair_plus_kspace_seconds	80.914	Tighten exact oxygen-i/non-oxygen-j pair Coulomb prechecks in `pair_lj_cut_tip4p_long` and reuse the first PPPM TIP4P oxygen M-site mapping plus hoisted alpha/slab constants in `fieldforce_ik()` and `fieldforce_ad()` [Correctness passed, but 80.914 is only a 1.19% improvement over incumbent 81.885, below the 2% acceptance bar.]
11	2a0b902a8b5a	accepted	weighted_median_pair_plus_kspace_seconds	79.88	Tighten exact oxygen-to-nonoxygen Coulomb prechecks and add a dedicated water-H `i` fast path in `pair_lj_cut_tip4p_long` that skips impossible LJ work while preserving the incumbent TIP4P force/tally order. [Correctness passed and 79.88 beats incumbent 81.885 by 2.45%, clearing the 2% acceptance bar.]
12	4f3910700624	rejected	weighted_median_pair_plus_kspace_seconds	82.549	Prune impossible hydrogen-neighbor LJ work and reuse cached oxygen M-sites to tighten reverse nonoxygen-to-oxygen TIP4P Coulomb prechecks in pair_lj_cut_tip4p_long [Correctness passed, but candidate 4f3910700624 regressed weighted_median_pair_plus_kspace_seconds from 79.88 to 82.549 versus incumbent 2a0b902a8b5a.]
13	c467f241b716	rejected	weighted_median_pair_plus_kspace_seconds	78.33	Skip dead water-H LJ work in `pair_lj_cut_tip4p_long` and cache PPPM TIP4P oxygen `find_M()` / H-index results across `particle_map()`, `make_rho()`, and `fieldforce_ik()`. [Correctness passed, but 78.33 improves incumbent 79.88 by only 1.94%, missing the benchmark contract’s 2% acceptance bar.]
14	b11c640621e4	rejected	weighted_median_pair_plus_kspace_seconds	78.535	Skip dead hydrogen-neighbor LJ work in `pair_lj_cut_tip4p_long` and replace the heavier PPPM full oxygen-site cache with lightweight per-step cached H indices plus on-demand M-site reconstruction in `particle_map()`, `make_rho()`, and the TIP4P fieldforce paths. [Correctness passed, but candidate b11c640621e48ee223768704bd744c8e8cb29952 improves the incumbent by only 1.68% (79.88 to 78.535), below the benchmark contract’s 2% acceptance bar.]
15	48e378f90ae8	rejected	weighted_median_pair_plus_kspace_seconds	79.58	Add an IK-only PPPM TIP4P oxygen cache that stores per-step H indices, lower-left grid deltas, and rho1d stencil weights in `particle_map()` and reuses them in `make_rho()` and `fieldforce_ik()` while leaving pair, AD, per-atom, and slab paths unchanged. [Correct and slightly faster (79.58 vs 79.88), but the 0.38% gain misses the benchmark contract’s 2% acceptance bar.]
16	ab55ea3e8508	accepted	weighted_median_pair_plus_kspace_seconds	78.12	Restore safe hydrogen-neighbor LJ pruning in `pair_lj_cut_tip4p_long` and reuse cached TIP4P oxygen M/H state across PPPM `particle_map()`/`make_rho()`/`fieldforce_ik()` with `memset` density zeroing and linear IK brick indexing [Correctness passed and 78.12 beats incumbent 79.88 by 2.20%, clearing the 2% acceptance bar.]
17	ad7811dc14f3	rejected	weighted_median_pair_plus_kspace_seconds	77.512	Generation-stamped TIP4P pair/PPPM caches plus a common non-special pair fast path and cross-step PPPM oxygen-H mapping reuse between reneighbor steps [Correctness passed, but 77.512 improves incumbent 78.12 by only 0.78%, below the benchmark contract’s 2% acceptance bar.]
18	b87604e6465e	rejected	weighted_median_pair_plus_kspace_seconds	77.755	Pair-only TIP4P optimization: replace the per-step `hneigh[][2]` reset with generation-stamped cached M-site bookkeeping and add a common non-special-neighbor fast path in `pair_lj_cut_tip4p_long`, leaving PPPM unchanged. [Correctness passed, but 77.755 improves incumbent 78.12 by only 0.47%, below the 2% acceptance bar.]
19	1478ab81846a	correctness_failure	weighted_median_pair_plus_kspace_seconds	78.069	Buffer hydrogen-i Coulomb force accumulation in pair_lj_cut_tip4p_long and flush once after the neighbor loop while preserving the incumbent oxygen and TIP4P redistribution write order [Hard reject: candidate 1478ab81846a is only 0.07% faster than incumbent ab55ea3e8508 but fails authoritative field-tolerance correctness in both cases.; case test-32r-long etotal_series relative_delta exceeds threshold (8.42997e-05 > 1e-06); case test-32r-long pe_series relative_delta exceeds threshold (0.00183969 > 1e-06); case test-32r-long ke_series relative_delta exceeds threshold (0.00345568 > 1e-06); case test-32r-long temp_series relative_delta exceeds threshold (0.00345568 > 1e-06); case test-32r-long press_series relative_delta exceeds threshold (0.222002 > 1e-05); case test-32r-long forces_xyz abs_delta exceeds threshold (0.105529 > 1e-08); case test-32r-long forces_xyz rms_delta exceeds threshold (0.0282935 > 1e-08); case test-64r-short forces_xyz abs_delta exceeds threshold (8.3e-06 > 1e-08); case test-64r-short forces_xyz rms_delta exceeds threshold (1.10787e-06 > 1e-08)]
20	57da41d1ad04	rejected	weighted_median_pair_plus_kspace_seconds	78.045	Cache all-charge PPPM/TIP4P IK rho stencils and base brick indices in make_rho() and reuse them in fieldforce_ik() while retaining the incumbent oxygen M/H cache [Correctness passed, but 78.045 is only a 0.10% improvement over incumbent 78.12 and misses the benchmark contract’s 2% acceptance bar.]
21	2599e1f9f4c8	accepted	weighted_median_pair_plus_kspace_seconds	76.248	Use contiguous alias-friendly x/f/part2grid views in pair_lj_cut_tip4p_long and TIP4P PPPM particle_map/make_rho/fieldforce_ik without changing force, tally, or cache semantics [Correctness passed and candidate 2599e1f9f4c8 reduces weighted_median_pair_plus_kspace_seconds from 78.12 to 76.248, a 2.40% improvement over incumbent ab55ea3e8508 that clears the 2% acceptance bar.]
22	22ab6a7a53b5	rejected	weighted_median_pair_plus_kspace_seconds	76.144	Reuse the pair style's current-step TIP4P O/H/M cache inside PPPM with a timestep guard and orthogonal-box fallback to the incumbent PPPM find_M cache [Correctness passed, but 76.144 is only a 0.14% improvement over incumbent 2599e1f9f4c8’s 76.248, well below the contract’s 2% acceptance bar.]
23	bd8ee0840992	rejected	weighted_median_pair_plus_kspace_seconds	75.536	Merge the correctness-clean pair-only generation-stamped TIP4P site cache and common non-special-neighbor fast path into the current alias-view `pair_lj_cut_tip4p_long` incumbent [Correctness passed, but 75.536 improves incumbent 76.248 by only 0.93%, below the contract’s 2% acceptance bar.]
24	a4391bb48f95	rejected	weighted_median_pair_plus_kspace_seconds	79.984	Specialize TIP4P PPPM order-5 make_rho()/fieldforce_ik() stencil kernels with a generic fallback [Correctness passed, but the order-5 PPPM stencil specialization regressed the incumbent by 4.90%, far short of the 2% improvement bar.]
25	757c16f42b32	rejected	weighted_median_pair_plus_kspace_seconds	75.36	Pure-water `lj/cut/tip4p/long` pair fast path plus generation-stamped TIP4P M-site reuse and common `sb == 0` handling in `pair_lj_cut_tip4p_long` [Correctness passed, but 75.36 improves incumbent 76.248 by only 1.16%, below the 2% acceptance bar.]
26	4cbc0fa85603	accepted	weighted_median_pair_plus_kspace_seconds	72.919	Contiguous generation-stamped TIP4P pair/PPPM caches plus a pure-water lj/cut/tip4p/long pair fast path that keeps exact oxygen-site cutoffs and incumbent force/tally ordering [Correctness passed and the candidate cuts weighted_median_pair_plus_kspace_seconds from 76.248 to 72.919, a 4.37% improvement over incumbent 2599e1f9f4c8 that clears the 2% acceptance bar.]
27	5840dac0d9f1	accepted	weighted_median_pair_plus_kspace_seconds	70.984	Inline TIP4P pair/PPPM cache-hit fast paths while keeping cache-miss reconstruction in dedicated out-of-line slow helpers [Correctness passed and 70.984 beats incumbent 72.919 by 2.65%, clearing the 2% acceptance bar.]
28	aaa3d5fa36f1	rejected	weighted_median_pair_plus_kspace_seconds	71.085	Streamline PPPMTIP4P by removing the unconditional particle_map() allreduce and specializing the oxygen/non-oxygen plus z-force hot paths in particle_map(), make_rho(), and fieldforce_ik() without changing atom traversal or TIP4P force distribution [Correctness passed, but candidate aaa3d5fa36f1 regressed the incumbent primary metric from 70.984 to 71.085 and missed the 2% improvement bar.]
29	b9d56dcc532a	rejected	weighted_median_pair_plus_kspace_seconds	70.525	Use typed alias-friendly cached TIP4P M-site views in pair_lj_cut_tip4p_long and pppm_tip4p while keeping the incumbent cache-miss reconstruction and exact force/tally ordering unchanged [Correctness passed, but candidate b9d56dcc532a improves the incumbent by only 0.65% (70.984 to 70.525), below the 2% acceptance bar.]
30	b028e7ec4bbc	correctness_failure	weighted_median_pair_plus_kspace_seconds	71.716	Cache per-step TIP4P O->M deltas/reach in pair_lj_cut_tip4p_long and reuse them to form exact oxygen-path Coulomb separations plus tighter cached-oxygen prechecks without changing force redistribution or tally order [Hard correctness failure in both benchmark cases, and the candidate also regressed the incumbent by 1.03%.; case test-32r-long etotal_series relative_delta exceeds threshold (8.74878e-05 > 1e-06); case test-32r-long pe_series relative_delta exceeds threshold (0.00235329 > 1e-06); case test-32r-long ke_series relative_delta exceeds threshold (0.00440903 > 1e-06); case test-32r-long temp_series relative_delta exceeds threshold (0.00440903 > 1e-06); case test-32r-long press_series relative_delta exceeds threshold (0.253437 > 1e-05); case test-32r-long forces_xyz abs_delta exceeds threshold (0.118184 > 1e-08); case test-32r-long forces_xyz rms_delta exceeds threshold (0.0291073 > 1e-08); case test-64r-short etotal_series relative_delta exceeds threshold (1.18848e-06 > 1e-06); case test-64r-short pe_series relative_delta exceeds threshold (3.56251e-05 > 1e-06); case test-64r-short ke_series relative_delta exceeds threshold (6.64653e-05 > 1e-06); case test-64r-short temp_series relative_delta exceeds threshold (6.64676e-05 > 1e-06); case test-64r-short press_series relative_delta exceeds threshold (0.00278554 > 1e-05); case test-64r-short forces_xyz abs_delta exceeds threshold (0.0186149 > 1e-08); case test-64r-short forces_xyz rms_delta exceeds threshold (0.000930271 > 1e-08)]
31	9eb54352a2e0	rejected	weighted_median_pair_plus_kspace_seconds	71.282	Precompute map-resolvable TIP4P oxygen M-sites before the pair loop and read the prepared cache directly, with the incumbent lazy path retained for unresolved ghost oxygens [Correctness passed, but 71.282 regressed incumbent 70.984 by 0.42% and missed the 2% acceptance bar.]
32	8bfdf92c40a5	rejected	weighted_median_pair_plus_kspace_seconds	70.124	Defer cached TIP4P hydrogen-index loads until exact oxygen interactions pass the final Coulomb cutoff in pair_lj_cut_tip4p_long, and use an xM-only cached M-site accessor in pppm_tip4p::particle_map() [Correctness passed, but 70.124 improves incumbent 70.984 by only 1.21%, below the 2% acceptance bar.]
33	ff67ca3c0752	rejected	weighted_median_pair_plus_kspace_seconds	72.294	Defer TIP4P pair hydrogen-index loads until exact oxygen-cutoff hits, add a common nonspecial Coulomb fast path in `pair_lj_cut_tip4p_long`, and use an xM-only PPPM cache accessor in `particle_map()` [Correctness passed, but candidate ff67ca3c0752 regressed the incumbent primary metric from 70.984 to 72.294, so it misses the 2% improvement bar.]
34	625e91025a03	rejected	weighted_median_pair_plus_kspace_seconds	70.877	Split TIP4P pair/PPPM caches into compact M-site coordinate arrays plus separate H-index/stamp state, deferring pair oxygen-neighbor H-index loads until exact Coulomb-cutoff hits [Correctness passed, but 70.877 improves incumbent 70.984 by only 0.15%, far below the 2% acceptance threshold.]
35	44b554c8a168	rejected	weighted_median_pair_plus_kspace_seconds	70.511	Defer pair-side TIP4P hydrogen-index and Coulomb-special loads until exact oxygen-path survivors, plus an xM-only PPPM particle_map cache accessor [Correctness passed, but candidate 44b554c8a168 improves the incumbent by only 0.67% (70.984 to 70.511), below the 2% acceptance bar, so 5840dac0d9f1 remains incumbent.]
36	d6aa774723cb	rejected	weighted_median_pair_plus_kspace_seconds	70.737	Use the lazy TIP4P pair/PPPM cache-access shape with an xM-only PPPM particle_map accessor, and hoist exact-survivor pair charge and coordinate math without changing force, tally, or write order [Correctness passed, but candidate d6aa774723cb improves the incumbent by only 0.35% (70.984 to 70.737), well below the 2% acceptance bar.]
37	e04c1aedddd3	rejected	weighted_median_pair_plus_kspace_seconds	70.108	Compose typed alias-friendly TIP4P cached M-site coordinates with lazy pair-side hydrogen-index use and an xM-only PPPM particle_map accessor, preserving incumbent cutoff and TIP4P force/tally ordering [Correctness passed, but 70.108 improves incumbent 70.984 by only 1.23%, below the contract’s 2% acceptance bar.]
38	095bee6d96dc	rejected	weighted_median_pair_plus_kspace_seconds	71.763	Lazy TIP4P xM-only cache access plus pair hydrogen-index deferral and PPPM stencil-base reuse [Correctness passed, but 71.763 regressed incumbent 70.984 by 1.10%, so this candidate cannot replace the incumbent.]
39	7c4d21219695	rejected	weighted_median_pair_plus_kspace_seconds	71.327	Split TIP4P pair/PPPM cache hits into xM-only versus hydrogen-index reads, and memoize the pair-side oxygen reach scalar alongside cached xM updates while preserving exact surviving M-site geometry [Correctness passed, but candidate 7c4d21219695 regressed the incumbent primary metric from 70.984 to 71.327, so it cannot replace 5840dac0d9f1.]