Next MYNN tendency and predictor oracle fixtures.
Source: NOAA-EMC/HRRR tag v4.1.21, sorc/hrrr_wrfarw.fd/WRFV3.9/phys/module_bl_mynn.F.
Source SHA256: 42178a22693ad14b1a3adcf3e44be3047f051abf80f3a64c0ce9ed9ac2ec6d8a.
Built on the node CPU with GNU Fortran at -O0, from the unmodified module.

run_transport_cloud_gsd41.F90 calls mynn_tendencies for 12 columns and 12 layers.
The first eight reproduce every committed clear-transport output as FP32 bits.
Four additional columns carry liquid and ice gradients, moist air, a negative
cloud forcing and enhanced diffusion. Outputs include all three water tendencies
and theta. The original thl input is recorded as thl0. The legacy thl column keeps
the earlier clear fixture's convention and is not the cloud-case input.

run_predict_gsd41.F90 calls mym_predict at source levflag 2 for eight columns and
12 layers. Cases bind the 1e-4 floor, absence of the 150 cap, top Neumann row,
negative initial TKE, negative production, stronger diffusion and longer step.
The as-written top solve differs from its adjacent level by one FP32 ULP.

The predictor validates the P20 port. The corrected cloud fixture validates
P13 and P22 under the explicit source-defect form and the P21 conversion.
All six tendencies match the source as float bits over 144 levels on an RTX 5090.

Validation: 144 cloud transport rows and 96 predictor rows are finite. All
8 prior clear columns are unchanged. Uncapped case maximum is 190.9942322;
floor case minimum is the FP32 representation of 1e-4. Exact hashes and
validation are retained in the lane's handover receipts.
SHA256:
dac3c6a484550b3b29ebdcc0ffa19de5f405f980ea28b2282568802e7d0b9a99  predict-gsd41.csv
d9d957c008faf81c81919244c20997b47d05563007235289b1b4b6760991606b  transport-cloud-gsd41.csv
52d960947db3c797d49ce3095183a169cfaec35c6f5e0a0b83ec6c614ec2d71c  run_predict_gsd41.F90
c62717594327b7bd803fc9f8020b5968dc1b89773f949579ad1fd79fb588a032  run_transport_cloud_gsd41.F90

Adversarial correction: the first harness bound qcd to qsq and carry_ice
to FLAG_QNC. The corrected call matches the source signature at 2742-2772,
passes the three variance arrays before tcd/qcd, and puts carry_ice in
FLAG_QI. All 96 prior clear rows retain every float bit. The corrected
case 11 consumes the negative cloud forcing and binds condensate clipping.
