[h0ffmann]

Before. The resident Kokkos preset of the fork (WW3@80644ba9) was byte-identical to the Fortran on ww3_shel's smoke and full, but not production-ready:

• ww3_multi gave wrong numbers. Its exchanges between grids wrote VA on the host between W3WAVE calls, while the device copy was marked current. So mww3_test_02 to _04 differed from the Fortran in 4 to 10 files.
• W3WAVE copied the whole spectrum back to the host at every return.
• Partitioning, boundary output and track output each copied the whole spectrum at every output time.
• WW_KOKKOS_ALL=1 and the W3SRCE gate fell back to the Fortran without a word, so a "pass" could compare the Fortran with itself.
• The preset lived in the lab script. ww_kokkos could not be installed without GoogleTest, NetCDF and Fortran.
• No CUDA build had been checked for this preset.

After. WW3/ is pinned at 50a10730 (fork branch ww3-gpu/integ-fixes), and the tree is rebased onto main c04035e, with #192, #193 and #194.

• No whole-spectrum copy in steady state. W3WAVE no longer reads VA at its return. ww3_multi's exchanges, the boundary output and the track output sync only the columns they read or write. W3CPRT (partitioning) and W3OUTG's partition fields run on the resident VA. On full, the copies are 1 host-to-device (the bind) and 0 device-to-host (v). That is now a hard rule in AGENTS.md, "Resident spectrum".
• ww3_multi is byte-identical on mww3_test_01 to _05 (SHRD), also with WW_KOKKOS_VA_POISON=1 (see the regtest table).
• The ESMF cap reads VA once before its exports (not compiled here ⚠).
• WW_KOKKOS_STRICT=1 turns every fallback into a stop with its reason. Without it, the run prints one line with the count of calls left on the Fortran. The regtest driver's Kokkos runs use strict mode.
• WW_KOKKOS_PRESET=resident is read through W3INIT, from one table (presets.txt) shared by the shim, the script and the driver. CPRT is in it.
• An install-only ww_kokkos (WW_KOKKOS_BUILD_TESTS=OFF) needs Kokkos only, and its package config pins the Kokkos version and backends. ADR-0018 (Proposed) asks where ww_kokkos lives for upstream. The licence statement is corrected: 249 of 347 files under kokkos/src are LGPL-3.0-or-later.
• The CUDA build of the preset compiles (--build-only, no device run). A checklist for your supervised GPU run is in docs/GPU_SUPERVISED_RUN_CHECKLIST_202610.md.
• Plans: 002 (this PR), 003 (the first supervised GPU run, and the five plans after it), 004 (the pin to production/GFS.v17, in its own PR).

Goal abbb0456-6043, Claude Code session e664b977-8f49-44ff-a914-0cf9abbb0456.

Evidence

Machine: i9-14900, 32 threads, OpenMP backend, gfortran and g++ 16.2.0, Kokkos 5.2.2. The machine was shared during these runs, so every wall time here is ⚠, and none is for a table.

Check, on daaef07	Result
just kokkos-test serial-debug, openmp-release	11/11, 11/11
just ww3-gpu-matrix-selftest, bash scripts/ww3-gpu.sh --self-test, port_score.py --check	ok
python3 -m unittest discover tests	88 OK in a clean worktree. In a checkout that holds agent worktrees under .claude/worktrees/, test_results fails because results.py check also scans those untracked copies
WW_KOKKOS_VA_POISON=1 WW_KOKKOS_VA_COUNT=1 bash scripts/ww3-gpu.sh --ww3-gpu smoke --cpu --kokkos resident	every model file byte-identical to the Fortran of the same executable; every file except the NetCDF byte-identical to --ww3, and nccmp-tol passes on the NetCDF
the same with full	the same verdicts; VA copies H2D 1, D2H 0
WW3GPU_KOKKOS=resident WW3GPU_LAB_SWITCH=ON just ww3-gpu-matrix 4 serial (the owner, worktree ~/code/ww3-gpu-pr191)	entries 1 and 2, ww3_tp2.6 ST4 and ST4_ASCII: FAIL, left as FAIL by the owner's decision, ran fortran-only, kokkos_spectra 0. The Kokkos run stops at start-up with W3KOKKOSMD : WW_KOKKOS_XYP3=1, BUT ITS KOKKOS TRANSLATION DOES NOT COVER THIS SWITCH FILE (logs/1.K.log:160): LIMON is an unstructured grid, which W3KOKKOS_PROP3_SETUP refuses, and the ww3_ounf IOSTAT=2 after it is a downstream effect. main records the same as refused by design (docs/log/2026-10-18-resident-regtest-lab-switch.md:51). Selection, verdict and parts unchanged. Entries 3 to 7 PASS: ww3_tp2.15 default, -g 5km, -i input_const, -i input_rho (test.comp 0/23, 0/23, 0/23, 0/25, nccmp-tol PASS on dir,dp,fp,hs), ww3_ts2 ST4_PR3_UQ (0/22); ctest 9/11 (v, the run's summary.tsv)
security-reviewer	nothing at confidence 8 or above
CI	green

Regtests on daaef07, WW3@50a10730 (unit 3; docs/log/2026-10-18-u3-regtests-v2.md)

All runs used WW3GPU_LAB_SWITCH=ON, WW3GPU_STRICT=ON, WW_KOKKOS_STRICT=1 in the Kokkos run, WW_KOKKOS_VA_POISON=1 and WW_KOKKOS_VA_COUNT=1. ww3_self was never needed, because no test.comp difference appeared. D2H counts whole-VA device-to-host copies.

Selection, parts	Entries	test.comp	D2H
serial, resident (the owner's run and unit 3's agree)	ww3_tp2.15 x4, ww3_ts2 ST4_PR3_UQ	0/23, 0/23, 0/23, 0/25, 0/22	0
serial, resident	ww3_tp2.6 ST4, ST4_ASCII	FAIL, left as FAIL by the owner: refused at start-up on an unstructured grid	0
omp, SRCE,OUTG,IOPE,UPDV,CPRT (XYP3/KTP3 refuse moving grids)	ww3_ts3 x5, mww3_test_05 x5 (every propagation scheme)	0/20 each, 0/27 each	24 and 153 each, from the Fortran propagation's host write
serial, resident, ww3_multi	mww3_test_01 to _04	0/22, 0/23, 0/34, 0/25	12, 0, 3, 0 (before the fix: 24, 384, 216, 864)
omp, SRCE,OUTG,IOPE,UPDV	mww3_test_05 ST1, ST4_PR3_UQ_OMP	0/27 each	156 each
serial, mww3_test_07		FAIL: WW3's own ww3_grid fails in the Fortran run too (SDS4 namelist under the lab switch)	—
serial, SRCE,OUTG,IOPE,UPDV,CPRT, words PR0;SHRD	ww3_tnc1 (partition fields)	0/30, W3CPRT on Kokkos	0
the same with resident	ww3_tnc1	FAIL by design: stops in W3INIT naming XYP3 (PR0)	—

Not run: the restart-reproduction pair (scripts/restart_pair.sh, new; the owner's test-191-restart.sh), and mww3_test_08 (MPI, ADR-0017). mww3_test_01 to _04 have the source terms off in their grid inputs, so they test the exchanges, the propagation and the outputs on the resident VA, not W3SRCE.

Needs a person's review before merge (AGENTS.md, "Never cheat a regtest tool")

1. A gate that now prints instead of failing. scripts/ww3-gpu.sh failed a run whose VA was "still newer on the device at finalize". With the exit read gone, that state is normal. The check for a missed sync is now WW_KOKKOS_VA_POISON=1, which fills a stale host copy with NaN.
2. Strict mode in the regtest driver's Kokkos runs (regtests/cmake/ww3gpu_common.cmake, ww3gpu_kokkos_env). An entry that used to fall back silently now fails.
3. The selection fix (WW3GPU_NO_OPTS): an entry with no options keeps its place in the list. Every existing selection count is unchanged.
4. WW3GPU_KOKKOS_PREFIX, which links WW3 against an installed ww_kokkos.
5. The ESMF cap's read (wmesmfmd.F90, PATCH.md section 29). Not compiled here: there is no ESMF library.
6. Two FAIL rows that stay FAIL (the owner, 2026-10-18): ww3_tp2.6 ST4 and ST4_ASCII under the resident preset, refused by design at start-up on an unstructured grid (evidence table). They are not dropped from the selection or marked expected-to-fail.
7. scripts/restart_pair.sh (new, from unit 3): the restart-reproduction pair on a SHRD case. It passes bash -n, shellcheck and its --self-test, and has never run.
8. selection:count fails on a custom WW3GPU_SWITCH_WORDS list, because regtests/CMakeLists.txt has no expected count for it. The entries themselves ran and passed. Not changed; the owner decides.

Owner steps (not for agents)

• Run ~/code/ww3-gpu-3/test-191-restart.sh (the restart pair, never run), and test-fortran-ab.sh --controls --patch <the W3SDS4 diff> if the NOAA record needs the Fortran A/B.
• Fast-forward the fork's ww3-gpu/kokkos-pnt to 50a10730: git -C WW3 push origin 50a10730:refs/heads/ww3-gpu/kokkos-pnt.
• Timing: a median of three on a quiet machine of main before kokkos: three bit-neutral resident speed-ups, the CPU study PDF, a selected-regtests runner #193, main after it, and this PR. One loaded run on this tree had Kokkos slower than the Fortran (46.4 s against 31.8 s), against your earlier 15.5 s against 23.4 s ⚠. Then the bench/results/ record, and the speed sentence of docs/NOAA_PROGRESS_RECORD_202610.md.
• Whether the lab-switch regtests let the regtested cells of kokkos/ports.tsv move to "yes".
• Any GPU run (plan 003, the checklist).
• Whether to act on the profile's speed proposals. They change loop nesting or skip zero weights, so each one is its own PR with L2 evidence.
• Where the A/B harness for a Fortran change proposed to NOAA goes (branch goal/abbb0456-6043-fortran-ab, e9366f9).

After merge

docs/NOAA_PROGRESS_RECORD_202610.md then has stale facts:

• the resident preset has seven parts (CPRT joined);
• the fork pin is 50a10730;
• ww3_multi and the restart are now shown, if the table above says so.

🤖 Generated with Claude Code

[h0ffmann]

Open concerns for review

These are recorded while the PR is in draft. Each one says what was checked and what was not.

1. The CUDA build compiles its host C++ with another g++

• In the CPU build, both gfortran and g++ are 16.2.0.
• In the CUDA build (the #cuda shell), gfortran is still 16.2.0, so WW3's Fortran is compiled the same way. But nvcc_wrapper hands the host half of every .cpp to g++ 14.4.0. That covers the shims, the table set-up and every host branch of a kernel (v, docs/log/2026-10-18-cuda-build-only.md:25-26, :74-77).
• Every static parity argument in kokkos/ports.tsv read g++ 16.2.0's output. No host path compiled by g++ 14.4.0 has been compared with the Fortran ⚠.
• Why the shell pins g++ 14.4 is not recorded. Presumably it is the newest host GCC that nvcc 12.9 accepts ⚠ (not read in NVIDIA's support matrix).
• Device code is compiled by nvcc's device compiler with --fmad=false, -prec_div=1, -prec_sqrt=1 and -ftz=0, and no fast-math flag (v, same log).
• Next step: plan 003's phase A4 runs the CPU reference on the frozen commit. It does not cover this. Phase A3 should add the -O3 GIMPLE of one host-side shim path under both compilers, or a CPU-backend build of the library with g++ 14.4 run pair to pair, so the GPU run is not judged on an unknown host compile.

2. The total number of whole-spectrum copies has not dropped yet (U8)

• Boundary output and track output now read only their columns: 70 kB in all on ww3_tr1, and 4.4 kB of track columns on ww3_tp2.2 (v, docs/log/2026-10-18-resident-output-readers.md).
• The per-run count of full copies is unchanged, because another sync in the same step still copies VA:
  • W3WAVE's exit read on ww3_tp2.2. U1, still running, removes it.
  • The host write after the Fortran PR1 propagation on ww3_tr1, every step. This PR does not cover it.
• The PR is not done on the "Resident spectrum" rule until U1's WW_KOKKOS_VA_COUNT=1 numbers show the drop.

3. A fork defect found and fixed

No KOKKOS switch without PR3 compiled: W3KOKKOS_MAP3 and W3KOKKOS_MAPT used W3ADATMD names that exist only under PR3. Fixed on the fork in 31a09522 (branch ww3-gpu/u8-out-readers). It goes into the one pin bump at the end of this PR.

4. A regtest-driver bug, not fixed (needs a person's review)

In regtests/cmake/ww3gpu_common.cmake:337-343, ww3gpu_select_entries drops the empty options of the first selected entry when that entry has none. E_OPTS then has one item fewer than E_TESTS, and configure stops with "list index out of range" (seen on the SHRD;RTD selection). U8 ran ww3_tr1 through run_entry.cmake directly for that reason ⚠. The fix changes how entries are selected, so it is a change to a comparison's invocation: it waits for review (AGENTS.md, "Never cheat a regtest tool").

5. Licence statement corrected (U6)

Gap 12 said the C++ is MIT. The SPDX lines say otherwise:

• 249 of the 347 files under kokkos/src are LGPL-3.0-or-later;
• 22 are MIT;
• 76 have no SPDX line.

ADR-0018 (Proposed) builds on these counts. The licence of the 76 files without an SPDX line, and the legal reading of each option, are open ⚠.

6. Changes to comparisons or their invocation in this PR

A person reviews these before merge:

• the U2 strict mode in regtests/cmake/run_entry.cmake (coming);
• WW3GPU_KOKKOS_PREFIX in scripts/ww3-gpu.sh (U6), which changes which ww_kokkos WW3 links;
• any fix for item 4.

[h0ffmann]

Two FAIL rows that stay FAIL (the owner, 2026-10-18)

The owner's serial run on daaef07, WW3GPU_KOKKOS=resident WW3GPU_LAB_SWITCH=ON just ww3-gpu-matrix 4 serial (worktree ~/code/ww3-gpu-pr191):

• Entries 1 and 2, ww3_tp2.6 ST4 and ST4_ASCII: FAIL. ran is fortran-only and kokkos_spectra is 0.
• Cause: the Kokkos run stops at start-up with W3KOKKOSMD : WW_KOKKOS_XYP3=1, BUT ITS KOKKOS TRANSLATION DOES NOT COVER THIS SWITCH FILE (build/ww3-gpu-matrix-serial-kokkos-resident-lab/logs/1.K.log:160, v). LIMON is an unstructured grid, which W3KOKKOS_PROP3_SETUP refuses. The ww3_ounf IOSTAT=2 that follows is a downstream effect.
• Same as main: main already records this as refused by design (docs/log/2026-10-18-resident-regtest-lab-switch.md:51).
• Decision: both entries stay FAIL. The selection, the verdict and the parts are unchanged, and neither entry is dropped or marked expected-to-fail. The PR body's Tested section and its review list say the same; unit 3's records will too.

Entries 3 to 7 of that run follow when it ends.