[h0ffmann]

Before. ADR-0010 removed every unit test of a ported routine and kept the tool tests. Since PR #191, every resident run depends on C++ that ports no WW3 routine:

• the resident spectrum's state machine, its column syncs and cache (resident_shim.cpp, 599 lines);
• the flags, presets, strict mode and fallback slots (part_flags.cpp, 135 lines).

That code had no test faster than a WW3 regtest. The review of #191 found its gaps by reading.

After. ADR-0019 (Proposed) would allow GoogleTest suites for that infrastructure only. Each suite goes through the public C ABI, on a host backend, and with a planted-defect check per test. Ported routines keep ADR-0010 unchanged.

Plan 006 covers:

• the four parts, ranked by what a silent defect costs: the resident shim, the flags, presets.txt against its three readers, and the package config;
• how every state is observed without a test hook;
• the test cases, and the order relative to Production single-node Kokkos runtime: no-copy resident VA, strict mode, preset, packaging, GPU-run plan #191's follow-up and Port W3XYPUG and W3XYPFSN2 (unstructured explicit propagation) to Kokkos, CPU and CUDA #196.

Nothing is built before the owner accepts ADR-0019.

Goal abbb0456-5370, Claude Code session e664b977-8f49-44ff-a914-0cf9abbb0456.

Check	Result
nix shell nixpkgs#vale -c scripts/vale.sh on the ADR, the ADR index and the plan	rc 0, 0 errors
python3 scripts/adrs.py	lists ADR-0019 Proposed
python3 -m unittest discover tests	88 OK

For the owner: accept, change or reject ADR-0019. Then, in plan 006's T2, decide whether WW_KOKKOS_STRICT=true or yes should turn strict mode on.

🤖 Generated with Claude Code

[h0ffmann]

Handover: this session is stopped, and another session takes over

The full record is in docs/log/2026-10-18-goal-abbb0456-handover.md on this branch (4eb7488).

What worked. #191 merged (5cd0140, pin 50a10730, and ww3-gpu/kokkos-pnt fast-forwarded to it):

• the ww3_multi defect is fixed;
• full makes no whole-spectrum copy in steady state;
• strict mode and the preset are in;
• partitioning and the output readers run on the resident spectrum;
• ww_kokkos installs on its own;
• the CUDA build compiles.

WW3's regtests on the rebased tree: resident serial 5 PASS (the two ww3_tp2.6 FAILs are by design), OMPG 10/10, mww3_test_01 to _05 and ww3_tnc1 0 files different.

Open work:

1. goal/abbb0456-8241-strict-outg-poison, 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. Pushed, no PR. Already shown in its log (docs/log/2026-10-18-strict-outg-poison.md):

   • smoke and full are byte-identical with POISON on;
   • a strict run stops on the OUTG fallback;
   • under all, ww3_tnc1 runs again.

   Left to do: the resident matrix result, port_score --write, the security review, and just pr. Note for whoever continues: smoke asks for WBT/SKW, so a strict resident smoke stops (37 OUTG fallbacks); ww3-gpu.sh sets no strict mode, so it still passes there.

2. This PR (docs: ADR-0019 (Proposed) and plan 006 #198): the owner accepts or changes ADR-0019, and answers plan 006's strict-parsing question.

3. goal/abbb0456-6043-fortran-ab (e9366f9): the Fortran A/B harness. No PR; the owner decides where it goes.

Waiting on the owner:

• the review lists of Production single-node Kokkos runtime: no-copy resident VA, strict mode, preset, packaging, GPU-run plan #191;
• the restart pair (~/code/ww3-gpu-3/test-191-restart.sh);
• the quiet-machine timing of main, then the NOAA record's facts;
• fork branch cleanup after the follow-up merges.

Runs: every run of this session is stopped. The one ww3_shel still running is in ~/code/ww3-gpu-2, another session's, and was left alone.