RTX 4060 (sm_89) after the profile rename, 2026-10-16

Verdict: every step of final_f2_4060.sh exits 0 (run.txt, 23:11 to 23:43 PDT): the preflight (18 of 18), T3 (every per-file run, t3/summary.txt), the four test files of the hopper preset's kernels, the full accumulators of every GPU kernel, T3p (6 passed), the three split-K envelope sweeps, the GPU tier of T4, the bench, and the replay of the H100 PCIe probe through h4 and h1. Every result equals that of the same script on fa0fd90 (../rtx-4060-sm89/README.md) once the old profile names are read as the new ones (below); only timings, paths and the commit differ.

• Hardware and software: the RTX 4060 host of ../../2026-10-14/rtx4060-host/README.md. Preflight 18 of 18 PASS (preflight.out, preflight/preflight.json): driver 570.211.01, Python 3.12.2, torch 2.7.1+cu128, cuBLASLt 12.8.3.14, the 32 packages of the reference lock, fatbins current and equal to the manifest's sha256, profile ada matches the device. The expected commit of preflight_4060.py was set to 2648bba.
• Code under test: 2648bba, a git checkout; mmaemu/, tools/ and tests/ unmodified before and after the run (run.txt). Its code is that of 63b7395, which renamed the profiles (hopper-h100pcie-torch -> hopper, hopper-h100pcie-te -> hopper-te, hopper-h100pcie-ws0 -> hopper-ws0, ada-rtx4060-torch -> ada, ada-rtx4060-te -> ada-te, ada-rtx4060-ws0 -> ada-ws0) and changed their descriptions; 2648bba adds logs only. Its fatbins are those of fa0fd90 (../rc-fatbin/README.md).
• Script: final_f2_4060.sh, the final_4060.sh of fa0fd90 writing to logs-f2/ and running the sweep scripts with the new profile names (sweeps/name_mapping.diff). One step after another, every GPU step under the GPU lock, with the ground truth and the audit data present. Paths: ~/ is the host's home directory.

Equal to the run on fa0fd90

Compared after replacing the old profile names by the new ones and dropping timings, paths and commits:

Output	Result
t3/summary.txt	the same passed, skipped and deselected counts in every file (test_backend_selection exits 5 in both: a CPU-only file, nothing collected under -m gpu)
t3/di_full_gpu_kernels.log	84 passed: the 21 Evo 2 calls of v1, v4, v9 and tc3 bit for bit equal to the RTX 4060's recorded accumulators
hopper-kernels/summary.txt	test_kernels_ieee 167, test_cuda_runtime 19, test_h100_mini 435, test_backend_selection 26 passed
replay-h4/mmaemu_replay.json (--backend auto)	every record equal: all 1,423 live calls equal the H100 (1,049 split, 374 unsplit; 1,043 on h4, 380 mixed-format on the reference backend), the 27 rejected calls raise ReferenceRejectsError, 168 of 168 Evo 2 outputs on h4 equal; "PASS": true
replay-h1/mmaemu_replay.json (--backend cuda:h1)	every record equal: the 1,043 E4M3 x E4M3 live calls and the 168 Evo 2 outputs on h1 equal; the 380 mixed-format calls raise BackendUnavailableError, so "PASS": false by construction, as on fa0fd90
t3p_profile_device.log	6 passed, the same six tests
sweeps/	the same query, split and max-f counts (306,859, 663,936 and 666,624 queries); the per-query row files (not kept, 127 MB) are byte-identical to those of fa0fd90 once its profile names are mapped
t4/mutants-gpu-rtx4060.json	PASS: the 7 kernel mutants killed, the identity control survives, each with its earlier outcome
mmaemu_info.txt	the same lines apart from the profile names; h1 and h4 "validated here"

Bench (bench-*-gauss40.json, .log)

tools/bench_kernels.py --kernels v1,v4,v9,tc3,h1,h4 --shape 1b|7b|40b --M 2048 --data gauss40, every kernel from the release fatbins, 0 mismatches in every exactness check. TMAC/s, the median of 5 windows, within 0.01 of the run on fa0fd90:

Kernel	1B (K 1920, N 5760)	7B (K 4096, N 12288)	40B (K 8192, N 24576)
tc3	2.94	2.95	2.94
v9	1.54	1.55	1.55
h4	1.34	1.40	1.41
v4	1.34	1.36	1.37
h1	0.69	0.69	0.69
v1	0.63	0.62	0.63