[h0ffmann]

Before. After #191, a passing Kokkos run did not prove that W3OUTG ran on Kokkos. For the fields the translation leaves out (WBT, SKW), W3KOKKOS_OUTG ran the Fortran W3OUTG without telling strict mode. No script set the missed-sync detector WW_KOKKOS_VA_POISON. Under WW3GPU_KOKKOS=all, the matrix set strict mode, so every entry whose switch file a part does not cover stopped in W3INIT. The resident restart had never been regtested, and scripts/restart_pair.sh had never run: it read the Kokkos-parts line from the wrong file and exited 1 without a message. W3SDS4, 24.8 % of the one-thread batched W3SRCE in the S0 profile, ran one sea point at a time.

After. This PR combines three pieces of work from goal abbb0456 and goal 5afd9379. Each piece was checked at the old pin 50a10730, and the combined branch was checked again at the new pin fac74069.

1. Strict W3OUTG, POISON in every run, strict mode only when a run names its parts (the follow-up of Production single-node Kokkos runtime: no-copy resident VA, strict mode, preset, packaging, GPU-run plan #191's review findings 1 to 3). The fallback of W3OUTG to the Fortran goes through KOKKOS_FALLBACK: a strict run stops there, and a run without strict mode counts the fallback. This is the fork commit fac74069 on ww3-gpu/strict-outg, PATCH.md section 31, and the WW3/ pin moves to it. scripts/ww3-gpu.sh and the regtest matrix set WW_KOKKOS_VA_POISON=1 in every Kokkos run and unset it in the Fortran runs. Strict mode applies only to a Kokkos run that names its parts, so all is a coverage run again. .gitmodules names ww3-gpu/kokkos-pnt as the fork branch that holds the pin. Record: docs/log/2026-10-18-strict-outg-poison.md.
2. The resident restart pair, first run (readiness gap 10). scripts/restart_pair.sh now reads log.ww3. This fixes only the plumbing; what the script compares is unchanged. Case full under SHRD, OpenMP with 4 threads, strict mode and POISON on. A run from the 06:00 restart gives the same 12:00 output and restart as the continuous run, for the Fortran and for the resident Kokkos, and every Kokkos file equals the Fortran's (cmp). This holds at 50a10730 and again at fac74069. Not covered ⚠: MPI, CUDA, ww3_multi, a moving grid. Record: docs/log/2026-10-18-resident-restart-pair.md.
3. W3SDS4 with one sea point per vector lane (plan 005 unit L3, after kokkos: W3SIN4's stress scan with one sea point per vector lane… #199's L2). w3sds4_point is split into parts that hold its lines unchanged. w3sds4_lanes<8> runs the mean-square-slope sums, the saturation integral and the turbulence and cumulative loops on 8-wide GCC vectors. The libmvec sites stay one grouped call per point. The V8/M8 types move to st4_vec8.hpp, shared with W3SIN4. Record: docs/log/2026-10-18-srce-lanes-sds4.md, plan 005.

Evidence

All runs used the OpenMP backend on the CPU. Nothing ran on a device; --gpu appears only as --build-only.

Check	At fac74069 (combined HEAD)
WW_KOKKOS_SRCE_LANES={1,8} WW_KOKKOS_VA_POISON=1 bash scripts/ww3-gpu.sh --ww3-gpu {smoke,full} --cpu --kokkos resident	(v) every model file byte-identical to the Fortran of the same executable; every file but the NetCDF byte-identical to --ww3's; nccmp-tol PASS
The same with WW_KOKKOS_STRICT=1	full: (v) the same verdict. smoke: exit 1, stopped by design in W3OUTG because smoke asks for a skewness field the translation leaves to the Fortran
WW_KOKKOS_VA_POISON=1 OMP_NUM_THREADS=4 bash scripts/restart_pair.sh	(v) exit 0, every comparison identical, no call left on the Fortran for SRCE, CPRT, OUTG
just regtest-selected	(v) PASS (ww3_tp2.15 x4 and ww3_ts2 ST4_PR3_UQ; the L3 log has the per-entry test.comp counts, 0 files different)
WW3GPU_KOKKOS=resident WW3GPU_LAB_SWITCH=ON WW3GPU_OMP_THREADS=2 just ww3-gpu-matrix 4 serial (at fac74069, before L3)	(v) 5 PASS; the two ww3_tp2.6 entries FAIL by design (unstructured grid, kept FAIL by the owner)
just kokkos-test serial-debug, openmp-release; CUDA --build-only of the preset	(v) pass; build and link clean
gates	(v) bash -n, shellcheck, both --self-tests, ww3-gpu-matrix-selftest, port_score.py --check, unittest, wfip check, figures.py check, vale 0 errors; security-reviewer: no findings

Time of the one-thread srce.batch.kernel, full, taskset -c 2, Intel Core i9-14900, at 50a10730, median of three alternated runs: 104.99 s with 1 point per team, 80.55 s with 8 (L2 and L3 together, 23.3 %). Regtest pending ⚠: only the selected entries, not the whole matrix. The first run with 1 point per team started at a load of 7.85 ⚠. Command and loads: the L3 log.

Needs a person's review

AGENTS.md: a change to a comparison script, a gate or how they are invoked is reviewed by a person.

• POISON in every Kokkos run (scripts/ww3-gpu.sh, regtests/cmake/ww3gpu_common.cmake). A correct run writes the same bits with or without it, and a caller's =0 turns it off.
• Strict mode only when a run names its parts (ww3gpu_common.cmake, run_entry.cmake, selftest.cmake). This relaxes when strict mode applies: under all, a fallback to the Fortran is counted in the exit line rather than stopping the run. Confirm that this hides nothing you want stopped.
• The new stop: strict mode now stops a run whose output asks for WBT or SKW, as smoke does. A strict run of smoke with the resident preset now exits 1 until those fields are translated.
• scripts/restart_pair.sh: the plumbing fix described above. Is one SHRD case enough to call gap 10 covered on the CPU?
• W3SDS4 lanes: the loop nesting (IK and ITH outside, the lanes inside, one accumulator per direction and lane); GCC vector types rather than std::simd. COS/SIN(USDIR) is computed once per lane, not once per ITH as the Fortran text has it. It is the same pure call on the same argument, so the bits are the same. Say whether "translate, don't improve" wants the per-ITH call kept.
• Not covered by full ⚠: SSDSC(3) > 0, SDSDTH >= 180, cases 2 and 3 of W3SDS4, and the fallback when the transposed tables are missing. The first three were forced by a temporary override and gave the same bits with 1 and 8 points per team. That compares the Kokkos code with itself, not with the Fortran.

Decisions left to the owner

• The lane loops on the host call the same glibc libm and libmvec per lane, not the transcribed device libm (docs/GPU_LIBM_BITS_202610.md, decision 2).
• The zero weights of the saturation integral are multiplied, not skipped: skipping is not bit-neutral (0 × Inf = NaN).
• Done 2026-10-18: the fork branch ww3-gpu/kokkos-pnt, which the new .gitmodules names, was fast-forwarded 50a10730..fac74069, so it holds the pin.

Links: the abbb0456 handover, the c3e7eecf handover, #191, #199, gap 10 of docs/PRODUCTION_READINESS_202610.md. Goals abbb0456-8241 and 5afd9379 (session ed126573-8f64-46a0-9c43-4a8d5afd9379).

🤖 Generated with Claude Code