Recast WOOF 1.0.2

Recast WOOF 1.0.2

WRF-fidelity fixes correct moisture pressure terms, edge vertical velocity,
diffusion and map factors, and the 2 m humidity cap. Forecast answers change.

Requires Python 3.11 or newer. Native tools ship for Linux x86_64
(manylinux_2_28). GPU forecasts need an NVIDIA card:
recast-woof[gpu] (CUDA 12) for regional and global forecasts,
recast-woof[gpu-cu13] (CUDA 13) for hex forecasts.

pip install --upgrade "recast-woof[gpu]"
woof fetch-tables
woof doctor

New

• [devices] in a single-domain or tree config, or woof go --devices N,
  splits a forecast across cards, with per-card memory checks and an
  execution receipt. domains selects split grids, defaulting to all;
  split nests take forcing each parent step and return two-way feedback
  to split parents. woof sim --devices N or --devices-table sets a
  prepared forecast's split without changing its configuration.
  woof check CONFIG --devices N prices every card, grid and pinned host
  store, exiting 2 for an unknown VRAM budget. Frames reach the host while
  the model steps on. Split parents over resident nests, moving nests and
  late-starting split grids are refused. Split grids refuse Noah mosaic
  and slope radiation (slope_rad = 1) before allocation, naming the
  missing state; resident grids retain both. Splitting is off by default.
• woof run-plan starts GFS and native HRRR single domains and trees from
  the first posted hours, as woof go does. Both start single hrrr-prs,
  rap, rrfs, icon-eu and gem-gdps domains the same way. A [tiles]
  root, including woof cyclone-setup trees, waits at each boundary seam,
  instead of for the whole cycle. Output is unchanged. An earlier forecast
  failure keeps its own error beside any later input timeout.
• WOOF_FETCH_POLICY=aws pins sources with an AWS host to that host without
  NOMADS fall-through; sources with no AWS host retain their usual host.
  latest selects the newest cycle whose first hours are on AWS;
  --whole-cycle retains the whole-cycle rule. Forecasts still stream from
  their first posted hours. Unset, fetching is unchanged. woof fetch
  probes hosts refusing HEAD with a one-byte GET.
• woof import-namelist accepts domain_wizard, mod_levs, plotfmt and
  bare 30s geography, and fills 17 of 33 formerly required keys from WRF
  and WPS defaults. Unknown sections are refused by name.
• diff_opt = 1 runs WRF's coordinate-surface horizontal diffusion with
  km_opt 2 or 4 and either mix_full_fields setting, including restarts
  and streamed forecasts. It is off by default (diff_opt = 2).
• topo_wind runs under YSU, excluding BEP and BEP+BEM;
  gwd_opt = 1 or 3 runs KIM or GSL gravity-wave drag under every PBL
  scheme. GPU column checks match WRF 4.7.1 bit for bit. Static preparation
  adds orographic statistics from WPS geography, and namelist imports keep
  both keys. Orographic statistics are byte-identical on Linux and Windows.
  A streamed [tiles] domain stops at its first drag call.
  Both options are off by default.
• woof ml-export converts regional histories, converted hex frames,
  global tapes or their ZIPs into one Zarr store per domain. It offers
  WeatherBench 2's 13 pressure levels, ERA5's 37 or model levels, ERA5
  and WeatherBench 2 names and units, earth-relative winds, a
  below_ground mask over ERA5's below-ground fill, an optional regular
  latitude-longitude grid and a ZIP that opens in place. Levels above the
  model top are omitted and listed. xarray.open_zarr works with
  zarr-python 2.18 and 3. On Windows, ZIP history files with WRF's colon
  filenames unpack with underscores; see docs/ml-export.md.
• GPU preparation rotates winds, interpolates vertically and builds the
  Thompson cold start with the CPU route's arithmetic, matching its
  prepared arrays. Forecasts prepared on the card change. A full HRRR
  forcing preparation took 154 s instead of 640 s.
• Acoustic, advection, horizontal diffusion and per-step glue kernels move
  less memory, with byte-identical output. The horizontal diffusion change
  alone reduced an hour of an 800 x 600 x 50 forecast from 95.7 to 88.4 s
  on an RTX 5090.

Fixed

• Legacy RRTMG longwave radiation keeps the correct spacing above the
  model top. A pressure guard had compressed valid layers, changing their
  temperatures and heating. Forecast answers change.
• woof go --cycle accepts an exact match to a prepared bundle,
  checkpoint or declared forcing's actual initial time, recording a no-op.
  Conflicting or unreadable input times remain refused. Forcing fetched by
  the run is checked after acquisition and before preparation.
• Every shipped profile and configuration leaving moist_cq unset applies
  WRF's moisture correction to pressure terms, including passive vapour
  with microphysics off. Dry states bypass it; an explicit setting wins.
  Unmatched configurations had omitted it, and WSM6 and Kessler profiles
  had disabled it. Forecast answers change.
• The dynamical core corrects vertical-velocity damping and float32
  rounding, geopotential tendencies at open and specified outer rows,
  vertical advection at an open model top, mapped open-boundary map
  factors, and diffusion's sixth-order filtering, boundary deformation,
  outer-row TKE and prescribed surface heat flux. Routine output words
  are checked against compiled WRF 4.7.1. Forecast answers change.
• Surface vertical velocity at open and specified edges uses WRF's
  clamped terrain slope. Copying the interior slope outward had doubled
  its terrain-following term there. Periodic domains and interior cells
  are unchanged; edge answers change.
• Land points cap 2 m humidity at 1.05 times the lowest model level's
  vapour after the surface schemes, as WRF's surface driver does.
  Answers change for 2 m humidity.
• WRF wrfinput starts initialize unfrozen Noah soil liquid water from
  soil moisture and terrain-following vertical wind from surface flow,
  instead of retaining zeros. First-hour latent heat flux RMSE against WRF
  fell from 239 to 0.4 W/m2 on a 3 km comparison. With usemonalb off,
  Noah keeps WRF's seasonal background albedo instead of the file's monthly
  albedo. Answers change for these forecasts.
• woof downscale children of saved runs, including --point, build
  terrain, land use and soil at their own spacing, using Copernicus GLO-30
  terrain at 1 km or finer. Parent states and every boundary frame are
  blended and rebalanced as WRF's ndown does, within a few float32
  epsilons of compiled WRF 4.7.1 on column checks. A WPS_GEOG tree is
  required (woof fetch-geog or --geog-root); a missing tree is refused.
  --parent-terrain retains the old route, and [static] records the
  choice. Answers change for these children.
• Downscaled children and DA nests take steep-terrain acoustic substeps
  and the epssm floor from their own terrain, keeping automatic parent
  choices. Lower explicit epssm values are refused. Answers change when
  the child's terrain needs the rule and the parent's terrain did not.
• Adaptive steps land on history, radiation, cumulus, surface and
  boundary-layer times; history intervals need only be whole seconds.
  Prepared forecasts honor adaptive history alarms and delayed domains
  start on time. Answers change where those times fell between steps.
• Disk estimates count each domain's physics fields, history_vars and
  output window. Trimmed runs that fit start; insufficient space is
  refused before starting.
• BEP+BEM workspace estimates count urban columns, including forecasts
  starting while their source posts, keeping a forecast that fits on the
  card. Configuration checks explain their upper bound.
• A resident nest under a streamed or split parent restores from the
  shared scratch arena without a KeyError.
• Streamed final-state digests use the resident forecast's model time.
  Fractional adaptive steps no longer leave identical forecasts with
  different digests through a 0.27 ms time difference after 3 h.
• Resident, streamed and out-of-core checkpoint writers record the run
  configuration alike. Checkpoints of the same state no longer differ in
  their header.
• CPU-only validation and SASE resource accounting work without CuPy.
  Windows numerical receipts resolve their own paths.

Available since 1.0.1

• sf_urban_physics = 1 runs single-layer urban canopy under Noah or
  Noah-MP; 2 runs BEP with YSU or MYJ; 3 adds BEP's building energy
  model. Local Climate Zones and NLCD urban classes are retained. A
  single-layer canopy above the first model level is refused. Urban
  physics is off by default; BEP under YSU applies nonurban drag once.
• sf_surface_mosaic = 1 runs Noah land-use tiles per cell. With the
  single-layer canopy, mosaic_urban_canopy = "every_tile" also treats
  town tiles in mostly rural cells. Mosaic is off by default.
• The UW moist-turbulence PBL (bl_pbl_physics = 9) runs on the GPU and
  imports from WRF namelists. [shared] zadvect_implicit = 1 runs
  implicit-explicit vertical advection with w_crit_cfl, including the
  two corrected boundary terms. Implicit advection is off by default;
  answers change.
• Terrain accepts WPS GEOGRID.TBL smoothers or no smoothing through
  woof domain --terrain-smoothing, domain static settings or GEOGRID.TBL
  import. smooth_precision = "wps-float32" or
  --terrain-smoothing-precision matches default geogrid.exe smoothing;
  float64 remains the default.
• WRF namelists retain history_begin, history_end, smooth_cg_topo,
  topo_shading and slope_rad, and imports name all unsupported keys
  together. Restarts can change history windows where cadence can change.
  WUDAPT Local Climate Zones import with their urban land cover.
• Byte-identical physics speedups remain included: per call, legacy
  RRTMG 7x, RUC 7 to 22x, Noah-MP 5.8x, MYNN 4.5x, NSSL 1.9 to 3.1x,
  P3 1.5 to 2.2x, RTE-RRTMGP 1.6x, Thompson 1.3 to 1.6x, Morrison
  1.2 to 1.5x, and Milbrandt-Yau, Kain-Fritsch, New Tiedtke, Shin-Hong
  and Grell-Freitas 1.1 to 1.5x. Dycore bookkeeping and nest forcing
  use fewer launches.
• [shared] adaptive_nest_lattice = true selects the root step with the
  fewest nest cell-steps; the three-domain 250 m comparison was 6% faster.
  It is opt-in and changes answers.
• woof fetch takes posted leads in order and records their first
  availability or an already-up marker; late leads exit 75, AIGFS waits
  for its GDAS donor, and events carry every source's leads.
  GFS, native HRRR and mapped-source trees prepare their start and nests
  first, then stream later boundaries; storm-following trees start at
  that head too.
  woof sources shows posting schedules and lateness budgets; fetch
  latest, --readiness and --whole-cycle remain available, and
  woof go --cycle selects the cycle. HRRR DA uses expected posting
  times, and latest probes AIFS f360 from +5 h 26 min. Config preparation
  can pin its CPU or GPU backend with woof run --preprocess-backend or
  [case_data] preprocess_backend, and --preprocess-workers sets decode
  concurrency. A 48 h HRRR decode on 24 cores took 140 s instead of 480 s.