[mefrediani]

With fmoist_freq > 1, the first scheduled moisture update could miss the step-1 initialization, leaving class moisture and forcing history undefined. Initialize once from the first valid atmospheric state independently of cadence, apply zero elapsed moisture time during initialization, and restore that state on restart.

For example, with dt=4 and fmoist_freq=3, initialization is at 4 s, the update at 12 s uses 8 s elapsed time, and later updates use 12 s intervals. Time-based updates retain their existing phase. Invalid or unavailable current/previous forcing fails explicitly. The NUOPC rainfall accumulator is initialized before its first import.

Checkpoints now record initialization status/time and moisture activation/cadence. Loading restores the lifecycle and validates class state, forcing history, and matching settings. Prescribed legacy checkpoints remain readable. PR #39 time-based checkpoints are accepted when their metadata establishes compatible initialized state; ambiguous legacy step-based checkpoints fail with a specific diagnostic. Configuration documentation and twelve focused lifecycle scenarios cover the correction. No new namelist options, ordinary diagnostic fields, or humidity-formula changes are included.

Refs #42. Its remaining extinction, controls, and diagnostic work remains open.

Review base and dependencies

• This draft targets develop at ead119eca6dfb576d181a5560273d70035f780a3, which includes the squash merge of Add restart capability to the fire behavior model #39. The original local restart merge bc205bc and upstream ead119e have identical file trees but different ancestry.
• Head a87b23dae1c8b4a282ed503202cbb34579f0536f contains six commits and six files, without historical restart commits in its commit range. The final atmospheric defaults use the original patch bounds (ifps:ifpe, jfps:jfpe) in Set_vars_to_default; state/state_mod.F90 now matches develop exactly and drops out of the feature diff.
• The review follow-up sets the four atmospheric fill values in Set_vars_to_default, initializes the precipitation-rate accumulator in NUOPC Realize while preserving restored rainfall, and groups external/intrinsic imports before CFBM modules. The component member remains initialized; its NetCDF attribute remains fmc_initialized.
• P00 / Add focused restart regression coverage #61 at ed18f91 remains an explicit validation dependency absent from this feature's ancestry. The remaining foundation merge order is this PR, then Add focused restart regression coverage #61. Update and validate P00 separately before its merge.
• The earlier replayed implementation dec7eb2 was byte-for-byte identical to validated 3e472d8. Initialization-placement validation used separate local integration d144a898a436efb7ef411c781368e6853c0415b3, combining P00 and P02 at 439474f; all seven feature files matched that revision exactly. The subsequent patch-bound restoration changes only the four atmospheric default assignments. Original branches and validation evidence remain preserved.

Validation

Earlier validation at combined revision 554d150 used GNU Strict builds on Derecho with GNU 12.2.0, Cray MPICH 8.1.25, NetCDF 4.9.2, and ESMF 8.5.0. Serial, OpenMP, MPI, hybrid, and ordinary NUOPC/ESMX execution passed.

• Twelve focused initialization/forcing/timing scenarios per build; 45 Python tests; formatting and diff checks passed.
• All 15 restart cases passed, including the five delayed-moisture cases previously failing in P00. There were 65 continuous/segmented model runs and 30 loader probes.
• All 320 file comparisons and 12 cross-layout comparisons had zero maximum absolute error; 450 checkpoint checks passed. Loader probes comprised five accepted schema-compatible resumes and 25 expected rejections.
• Existing PR suite: 36 model runs, 40 execution comparisons, and 36 approved-reference comparisons passed. Largest cross-driver absolute difference was fgrnhfx = 0.125 W m-2, consistent with the existing baseline. No reference or tolerance changed.
• Original PR Add restart capability to the fire behavior model #39-era checkpoints generated with P00 were also checked: prescribed/time-based resumes agreed exactly with corresponding P02 checkpoints, and ambiguous step-based state was rejected.

PBS jobs: 7890398.desched1, 7890541.desched1, 7890596.desched1. Full retained evidence: /glade/derecho/scratch/frediani/tmp/cfbm-P02-20261021T034953Z/, including attempt-02/results/restart-z8yehdcl/summary.json, attempt-02/results/pr-hw75p1af/summary.json, and legacy-originals/summary.json.

P01/P02/P04 combined acceptance remains pending, including shared clock precision and fractional/long-duration timing. Direct WRF-host execution/restart, NUOPC/ESMX restart, and transfer of checkpoints between MPI layouts remain untested here. Ordinary NUOPC/ESMX execution and comparisons among independent MPI layouts do not establish those restart paths. GitHub CI must pass on the final PR revision before merge; its normal PR suite does not run P00's optional restart suite.

Initialization-placement follow-up validation

Feature 439474f matches tested P00/P02 integration d144a898a436efb7ef411c781368e6853c0415b3 exactly. Four fresh GNU Strict builds (serial, OpenMP, MPI/NUOPC/ESMX, hybrid/NUOPC/ESMX) passed with the environment above.

• All 15 restart cases pass: 65 continuous/segmented runs, 30 loading probes, 320 zero-error file comparisons (315 continuity and five accepted legacy resumes), 12 zero-error cross-layout comparisons, and 450 checkpoint checks. The restart suite's 67 CTest invocations include the twelve focused P02 lifecycle scenarios in each build.
• The existing 36-run PR matrix, 40 execution comparisons, and 36 approved-reference comparisons pass. Maximum cross-driver difference remains fgrnhfx = 0.125 W m-2 (maximum percentage error 0.0006368705741347137%), within unchanged tolerances.
• A connected, nonzero precipitation-rate ESMX_Data check passes with one MPI rank, four MPI ranks, and four ranks by four OpenMP threads. All 33 output-file comparisons against the previously validated executable are exact. Thirty accumulated-rain checks have maximum absolute error 1.862645149230957e-09 in model accumulation units.
• Formatting and regression both passed in GitHub workflow 39690399727 on 439474f. Local formatting and whitespace checks pass. No approved reference or regression tolerance changed.

PBS job 7900048.desched1 completed the builds and both suites successfully, then exited 1 because the scratch rate-check script looked for its completion marker in stdout instead of the ESMF log. After correcting only that checker, job 7900164.desched1 reran the six rate-check model executions and exited 0. Failed-attempt evidence and original scripts remain preserved.

Evidence: /glade/derecho/scratch/frediani/tmp/cfbm-P02-init-review-20261021T225746Z/, including attempt-01/results/restart-hl_i2yr_/summary.json, attempt-01/results/pr-jpmyi5r2/summary.json, and attempt-01/rate-probe-02/summary.json. The combined P01/P04, coupled-restart, and direct WRF-host validation gaps above remain.

Patch-bound restoration

The four atmospheric defaults retain their original ifps:ifpe, jfps:jfpe patch sections. The temporary explicit memory-bound spelling in f6a6206 is superseded by a87b23d. The final assignments match the original expressions, and state/state_mod.F90 matches develop exactly. Defaults remain in Set_vars_to_default; the NUOPC cold-start rainfall initialization in Realize is unchanged.

Whitespace checks pass. The detailed HPC results above are from 439474f; the optional restart suite has not been rerun after restoring patch bounds. GitHub CI on a87b23d is pending.