iteration	commit	status	primary_metric_name	primary_metric_value	description
0	44c83aaae41f	baseline	weighted_median_td_kernel_wall_seconds	128.753286794	baseline
1	b93246c4bf06	accepted	weighted_median_td_kernel_wall_seconds	99.5698850644	Use root-specific Ritz values for LR Davidson residual preconditioning, including vectorized shifts in the shared TD preconditioner [22.7% primary-metric improvement with all correctness guardrails passing and no hard-reject condition.]
2	e159c73ee967	accepted	weighted_median_td_kernel_wall_seconds	58.7037026142	Limit real TDDFT Davidson expansion in `_lr_eig.real_eig` to requested roots to avoid non-target residual/preconditioner work [Primary metric improved 41.0% vs incumbent with correctness passing and no hard-reject condition.]
3	7b1eeb5571f7	rejected	weighted_median_td_kernel_wall_seconds	58.6193054733	Limit symmetric `_lr_eig.eigh` Davidson trial-vector expansion to requested roots while keeping retained Ritz subspace behavior intact [Correctness passed, but the 0.144% primary-metric improvement is below the 2% acceptance threshold.]
4	f46de8e0dbb1	rejected	weighted_median_td_kernel_wall_seconds	59.1827705903	Cap symmetric LR Davidson expansion to requested roots and add a 1e-3 default TD preconditioner level shift [Primary metric regressed 0.816% vs incumbent despite correctness passing, so it fails the 2% acceptance threshold.]
5	ccc6bedc3559	accepted	weighted_median_td_kernel_wall_seconds	45.1467415988	Correct real TDDFT Davidson preconditioning to pass the full lower LR residual block by using `-R_y` in `_lr_eig.real_eig` [Primary metric improved 23.1% vs incumbent with correctness passing, no hard reject, and no forbidden-path edits.]
6	7612d6ba4e26	accepted	weighted_median_td_kernel_wall_seconds	37.4126424799	Cap symmetric `_lr_eig.eigh` Davidson expansion to requested roots to avoid non-target Casida residual/preconditioner work [Primary metric improved 17.1% vs incumbent with correctness passing and no hard-reject or forbidden-path condition.]
7	c1e86bf88a7c	accepted	weighted_median_td_kernel_wall_seconds	31.5845305966	Limit symmetric `_lr_eig.eigh` residual and preconditioner candidate generation to requested target roots [Primary metric improved 15.6% vs incumbent with correctness passing and no hard-reject or forbidden-path condition.]
8	3c057cbab323	rejected	weighted_median_td_kernel_wall_seconds	31.4699873622	Pass occupied-only MO coefficient/occupation arrays into DFT TD response kernel cache setup for restricted TDA/TDDFT and RKS/UKS Casida paths, avoiding zero-occupation virtual columns while preserving matvecs and solver tolerances [Correctness passed, but the 0.363% primary-metric improvement is below the 2% acceptance threshold.]
9	caffdfe9f92b	rejected	weighted_median_td_kernel_wall_seconds	31.5343184173	Add a configurable 0.02 Hartree spectral shift to `_lr_eig.real_eig` correction preconditioning to reduce real TDDFT response batches while leaving Casida `eigh` paths unchanged. [Correctness passed, but the 0.159% primary-metric improvement is below the 2% acceptance threshold.]
10	64da7449ef1c	accepted	weighted_median_td_kernel_wall_seconds	30.0180831933	Add a 0.05 Hartree real_eig correction preconditioner spectral shift to reduce late-cycle B3LYP TDDFT response work [Primary metric improved 4.96% vs incumbent with correctness passing and no hard-reject or forbidden-path condition.]
11	110904bf4af3	rejected	weighted_median_td_kernel_wall_seconds	29.8952809311	Reduce DFT TD response setup/allocation overhead by using occupied-only response cache inputs and cheaper restricted TD/Casida matvec array handling [Correctness passed, but the 0.409% primary-metric improvement is below the 2% acceptance threshold.]
12	701a4b52c9e7	rejected	weighted_median_td_kernel_wall_seconds	29.9924535668	Stable-sort threshold-selected RHF/UHF Koopmans TD initial guesses by increasing diagonal gap before Davidson truncation [Correctness passed, but the 0.085% primary-metric improvement is below the 2% acceptance threshold.]
13	c2b4290991d4	rejected	weighted_median_td_kernel_wall_seconds	36.0526498561	Add configurable LR correction preconditioner shifts in `_lr_eig.py`: a small `+1e-3` shift for symmetric `eigh` corrections and a stronger `+8e-2` shift for `real_eig` corrections to improve Davidson convergence without changing tolerances. [Correctness passed, but the candidate regressed the primary metric by 20.1% vs incumbent and increased Davidson work.]
14	933fdb1739c3	rejected	weighted_median_td_kernel_wall_seconds	29.9177771769	Reduce DFT TD response setup/allocation overhead by using occupied-only response-cache inputs, full remaining TD response memory, and cheaper restricted/Casida matvec array handling [Correctness passed, but the 0.334% primary-metric improvement is below the 2% acceptance threshold.]
15	c046c9ec1362	rejected	weighted_median_td_kernel_wall_seconds	29.967951294	Vectorize symmetric `_lr_eig.eigh` correction preconditioning for unconverged target residuals in one batched preconditioner call while preserving solver tolerances, linear-dependency filtering, and response batch selection. [Correctness passed, but the 0.167% primary-metric improvement is below the 2% acceptance threshold.]
16	2a188b646cff	rejected	weighted_median_td_kernel_wall_seconds	30.5035316269	Taper the accepted real_eig correction preconditioner shift downward for late-stage residuals below 100x tolerance while preserving solver tolerances and root filtering [Correctness passed, but the primary metric regressed 1.62% vs incumbent and total matvec applications increased to 178.]