iteration	commit	status	primary_metric_name	primary_metric_value	description
0	e7c0ed95a333	baseline	weighted_median_bond_seconds	0.32627	baseline
1	911f06a50100	accepted	weighted_median_bond_seconds	0.28156	Hoist `evflag`/`eflag`/`newton_bond` branches out of the serial `bond_class2` and `angle_harmonic` hot loops with OMP-style specialized `eval` helpers, and use flat `_noalias` typed views for coordinates, forces, and topology lists. [Correctness passed and the candidate cut weighted_median_bond_seconds from 0.32627 to 0.28156 (-13.70%), so it should replace the incumbent.]
2	25f839bdfc42	correctness_failure	weighted_median_bond_seconds	0.28546	Reuse per-type coefficients and reciprocals, evaluate the class2 polynomial in a lower-overhead form, reuse force components, and only materialize `ev_tally` force arrays in `bond_class2` and `angle_harmonic` specialized hot-loop instantiations. [Hard correctness failure in both benchmark cases, and the candidate was also 1.39% slower than incumbent on the primary metric.; case test-32r-long etotal relative_delta exceeds threshold (9.3752e-05 > 1e-08); case test-32r-long pe relative_delta exceeds threshold (0.0022341 > 1e-08); case test-32r-long ke relative_delta exceeds threshold (0.0041836 > 1e-08); case test-32r-long temp relative_delta exceeds threshold (0.0041836 > 1e-08); case test-32r-long press abs_delta exceeds threshold (1.38024e+08 > 1e-06); case test-32r-long force_components abs_delta exceeds threshold (0.106059 > 1e-08); case test-32r-long force_components rms_delta exceeds threshold (0.0295306 > 1e-08); case test-64r-short etotal relative_delta exceeds threshold (7.58474e-08 > 1e-08); case test-64r-short pe relative_delta exceeds threshold (4.16662e-08 > 1e-08); case test-64r-short ke relative_delta exceeds threshold (5.10999e-08 > 1e-08); case test-64r-short temp relative_delta exceeds threshold (4.97398e-08 > 1e-08); case test-64r-short press abs_delta exceeds threshold (2880 > 1e-06); case test-64r-short force_components abs_delta exceeds threshold (1.24e-05 > 1e-08); case test-64r-short force_components rms_delta exceeds threshold (1.19468e-06 > 1e-08)]
3	4a79c588ea08	accepted	weighted_median_bond_seconds	0.27301	Add a `bond_class2` single-bond-type fast path that dispatches to a dedicated `eval_one_type` helper and hoists constant coefficients out of the inner loop. [Correctness passed and the single-bond-type `bond_class2` fast path improved weighted_median_bond_seconds from 0.28156 to 0.27301 (-3.04% vs incumbent).]
4	e5c1ddf090fa	rejected	weighted_median_bond_seconds	0.27052	Add an `angle_harmonic` single-angle-type fast path that dispatches to `eval_one_type` and hoists `k[1]`/`theta0[1]` out of the inner loop. [Correctness passed, but the candidate improved weighted_median_bond_seconds by only 0.91% versus the incumbent, below the 2% acceptance threshold.]
5	224da0125669	rejected	weighted_median_bond_seconds	0.27041	Cache per-bond force components in bond_class2 and restore the single-angle-type fast path in angle_harmonic. [Correctness passed, but weighted_median_bond_seconds improved only 0.95% versus the incumbent, below the 2% acceptance threshold.]
6	e639b3949f97	rejected	weighted_median_bond_seconds	0.27144	Accumulate consecutive same-`i1` bonds in the single-type `bond_class2` fast path to reuse the leading atom's coordinates and force accumulator across each run. [Correctness passed, but same-`i1` bond-run accumulation improved weighted_median_bond_seconds by only 0.58% versus the incumbent, below the 2% acceptance threshold.]
7	47d88a349494	rejected	weighted_median_bond_seconds	0.27624	Restore the `angle_harmonic` single-angle-type fast path and add an explicit same-`i1` two-bond block in the single-type `bond_class2` kernel to reuse the leading atom state while preserving per-bond math and tally order. [Correctness passed, but iteration 7 regressed the primary metric to 0.27624, which is 1.18% slower than the incumbent 0.27301.]
8	e3c2ae9ee96a	rejected	weighted_median_bond_seconds	0.27111	Restore `angle_harmonic` single-angle-type dispatch and tighten one-type bonded hot paths by only materializing angle `ev_tally` arrays when `EVFLAG` is active and only computing `bond_class2` energy-only intermediates in `EFLAG` instantiations. [Correctness passed, but the candidate improved weighted_median_bond_seconds by only 0.70% versus incumbent 4a79c588ea08, below the 2% acceptance threshold.]
9	6fbda0abc94b	rejected	weighted_median_bond_seconds	0.26892	Add a dedicated two-way unrolled `bond_class2` one-type hot path for the dominant no-`evflag`, Newton-on case while preserving exact per-bond math and update order. [Correctness passed, but the two-way unrolled `bond_class2` hot path improved `weighted_median_bond_seconds` by only 1.50% versus the incumbent, below the 2% acceptance threshold.]
10	cf4ba8e7204d	accepted	weighted_median_bond_seconds	0.26621	Combine the correctness-safe `bond_class2` one-type two-way unrolled hot kernel with the correctness-safe `angle_harmonic` one-type dispatch and constant hoisting on top of incumbent `4a79c588ea08`. [Correctness passed and weighted_median_bond_seconds improved from 0.27301 to 0.26621 (+2.49% vs incumbent), clearing the 2% promotion threshold.]
11	672cbe901cae	accepted	weighted_median_bond_seconds	0.24606	Add a dedicated `angle_harmonic` one-type no-`evflag`, Newton-on hot path with scalar force temporaries and two-way exact unrolling. [Correctness passed with no hard rejects, and the dedicated `angle_harmonic` hot path improved `weighted_median_bond_seconds` from 0.26621 to 0.24606 (+7.57% vs incumbent), clearing the 2% promotion threshold.]
12	0826ffd39444	accepted	weighted_median_bond_seconds	0.23436	Cache one-type bond_class2 bond geometry and reuse paired bond data in angle_harmonic's dominant no-evflag, Newton-on hot path, with direct fallback when local ordering does not match. [Correctness passed and weighted_median_bond_seconds improved from 0.24606 to 0.23436 (+4.75% vs incumbent), clearing the 2% promotion threshold.]
13	30584dbe2ab0	accepted	weighted_median_bond_seconds	0.22529	Replace the one-type bond-to-angle reuse with an angle-oriented cache built in `bond_class2`, then consume it from a dedicated branch-free, two-way-unrolled `angle_harmonic` hot path with direct fallback to the incumbent local-geometry path when bond/angle pairing does not align. [Correctness passed and weighted_median_bond_seconds improved 3.87% versus incumbent, clearing the 2% promotion threshold.]
14	0890ed81e853	rejected	weighted_median_bond_seconds	0.26506	Replace the one-type bond-to-angle geometry cache with a prepared per-angle force cache built in `bond_class2`, then consume it as a direct force-apply hot path in `angle_harmonic` while preserving the validated structural guard and direct fallback. [Correctness passed, but moving full angle-force preparation into bond_class2 regressed weighted_median_bond_seconds by 17.65% versus incumbent 30584dbe2ab0.]
15	fdcd68522a83	correctness_failure	weighted_median_bond_seconds	0.21063	Compact the angle hot cache to bond deltas plus inverse bond lengths and consume them in the cached `angle_harmonic` hot path to remove cached-path divisions. [Hard correctness failure in both benchmark cases despite a nominal 6.51% speedup, so incumbent 30584dbe2ab0 remains.; case test-32r-long etotal relative_delta exceeds threshold (7.92579e-05 > 1e-08); case test-32r-long pe relative_delta exceeds threshold (0.00213102 > 1e-08); case test-32r-long ke relative_delta exceeds threshold (0.00400269 > 1e-08); case test-32r-long temp relative_delta exceeds threshold (0.00400269 > 1e-08); case test-32r-long press abs_delta exceeds threshold (1.39153e+08 > 1e-06); case test-32r-long force_components abs_delta exceeds threshold (0.115462 > 1e-08); case test-32r-long force_components rms_delta exceeds threshold (0.0280042 > 1e-08); case test-64r-short etotal relative_delta exceeds threshold (6.84866e-08 > 1e-08); case test-64r-short pe relative_delta exceeds threshold (3.14967e-08 > 1e-08); case test-64r-short ke relative_delta exceeds threshold (6.84812e-08 > 1e-08); case test-64r-short temp relative_delta exceeds threshold (6.63765e-08 > 1e-08); case test-64r-short press abs_delta exceeds threshold (2360 > 1e-06); case test-64r-short force_components abs_delta exceeds threshold (1.22e-05 > 1e-08); case test-64r-short force_components rms_delta exceeds threshold (1.13192e-06 > 1e-08)]
16	1f058192fe33	rejected	weighted_median_bond_seconds	0.23245	Compact the angle hot cache to bond deltas plus exact bond lengths, recomputing cached-path rsq values in angle_harmonic after dropping cached rsq fields. [Correctness recovered, but dropping cached rsq fields and recomputing them in angle_harmonic regressed weighted_median_bond_seconds by 3.18% versus the incumbent.]
17	738ffb9e3a30	rejected	weighted_median_bond_seconds	0.22321	Precompute the one-type bond/angle pairing layout once per topology build and reuse it in bond_class2's hot angle-cache fill to remove repeated per-step topology validation while preserving the incumbent cache payload and angle_harmonic consumer math [Correctness passed, but weighted_median_bond_seconds improved only 0.92% versus incumbent 30584dbe2ab0, below the 2% promotion threshold.]
18	bed99cfa6486	rejected	weighted_median_bond_seconds	0.22343	Prepare and reuse a prevalidated one-type bond-angle orientation layout in `bond_class2` hot cache fill while keeping the incumbent angle cache payload and `angle_harmonic` math unchanged [Correctness passed, but `weighted_median_bond_seconds` improved only 0.83% versus incumbent `30584dbe2ab0`, below the 2% promotion threshold.]
19	acd350270fd3	rejected	weighted_median_bond_seconds	0.22126	Compact the cached angle hot entry to store the exact bond-length product `r12 = r1*r2` instead of separate `r1`/`r2`, preserving cached `del*` and `rsq` semantics while reusing that exact product in the cached `angle_harmonic` hot path. [Correctness passed, but weighted_median_bond_seconds improved only 1.79% versus incumbent 30584dbe2ab0, below the 2% promotion threshold.]
20	c0a62a11cfaf	rejected	weighted_median_bond_seconds	0.22182	Combine prevalidated bond-angle orientation layout reuse in `bond_class2` with exact `r12` hot-cache compaction consumed by `angle_harmonic`. [Correctness passed, but the candidate improved weighted_median_bond_seconds by only 1.54% versus the incumbent, below the 2% promotion threshold.]
21	e4103f7dab04	correctness_failure	weighted_median_bond_seconds	0.21864	Store exact `r12 = r1*r2` in the hot angle cache and reuse a shared local `inv_r12` in the cached `angle_harmonic` hot path to remove the duplicate division by the same bond-length product. [Hard correctness failure in both benchmark cases from shared `inv_r12` reuse despite a nominal 2.95% speedup, so incumbent `30584dbe2ab0` remains.; case test-32r-long etotal relative_delta exceeds threshold (7.66767e-05 > 1e-08); case test-32r-long pe relative_delta exceeds threshold (0.00184228 > 1e-08); case test-32r-long ke relative_delta exceeds threshold (0.00346259 > 1e-08); case test-32r-long temp relative_delta exceeds threshold (0.00346259 > 1e-08); case test-32r-long press abs_delta exceeds threshold (1.29238e+08 > 1e-06); case test-32r-long force_components abs_delta exceeds threshold (0.100389 > 1e-08); case test-32r-long force_components rms_delta exceeds threshold (0.0278753 > 1e-08); case test-64r-short etotal relative_delta exceeds threshold (7.21825e-08 > 1e-08); case test-64r-short pe relative_delta exceeds threshold (6.86454e-08 > 1e-08); case test-64r-short ke relative_delta exceeds threshold (1.44472e-07 > 1e-08); case test-64r-short temp relative_delta exceeds threshold (1.46005e-07 > 1e-08); case test-64r-short press abs_delta exceeds threshold (1130 > 1e-06); case test-64r-short force_components abs_delta exceeds threshold (1e-05 > 1e-08); case test-64r-short force_components rms_delta exceeds threshold (1.15835e-06 > 1e-08)]
22	efcb4c8f585c	rejected	weighted_median_bond_seconds	0.22213	Compact the one-type hot angle cache to exact `r12 = r1*r2` and add prepared hot-angle index/layout metadata so `bond_class2` reuses stable topology preparation and cached `angle_harmonic` stops rereading `neighbor->anglelist` while preserving the incumbent arithmetic order. [Correctness passed, but the candidate’s 0.22213 weighted_median_bond_seconds is only a 1.40% improvement over incumbent 30584dbe2ab0, below the 2% promotion threshold.]
23	567a79cfb769	rejected	weighted_median_bond_seconds	0.22195	Widen the cached `angle_harmonic` hot consumer to four-way exact unrolling while preserving the incumbent cache payload, cached-path arithmetic order, guard, and direct fallback. [Correctness passed, but the four-way cached `angle_harmonic` unroll improved `weighted_median_bond_seconds` by only 1.48% versus incumbent, below the 2% promotion threshold.]
24	2c69753dfc48	rejected	weighted_median_bond_seconds	0.22158	Combine exact `r12` hot-angle cache compaction in `bond_class2` with a four-way cached `angle_harmonic` consumer while preserving the incumbent guard and cached-path division semantics [Correctness passed, but 0.22158 weighted_median_bond_seconds is only a 1.65% improvement over incumbent 0.22529, below the 2% promotion threshold.]
25	bf07f28c4118	accepted	weighted_median_bond_seconds	0.21875	Compact the hot angle cache to exact r12 payloads and widen the one-type cached bond_class2 producer to four-way exact unrolling while preserving the incumbent angle_harmonic arithmetic order and fallback guard [Correctness passed and weighted_median_bond_seconds improved from 0.22529 to 0.21875, a 2.90% gain over the incumbent that clears the 2% promotion threshold.]
26	a0800c8480fc	rejected	weighted_median_bond_seconds	0.22115	Restore prepared hot-angle layout/index reuse in bond_class2 and a four-way cached angle_harmonic consumer while keeping the incumbent exact r12 cache payload and four-way bond producer [Correctness passed, but weighted_median_bond_seconds regressed from 0.21875 to 0.22115 versus incumbent bf07f28c4118, so the candidate should not replace the incumbent.]