From the epic #278 feature-parity audit (Fortran-coverage pass). Two related reporting gaps vs amica15.f90, unrecorded:
write_nd (amica15.f90:658, 1760-1761, output at 2398-2429): the per-component, per-iteration squared-gradient trajectory. pamica exposes only the scalar RMS aggregate (ndtmpsum feeding the min_nd stop); the legacy NumPy backend keeps a public per-iteration nd list that torch/MLX lack. Diagnosing WHICH sources converge slowly is not possible on the primary backends.- Per-iteration run log (Fortran's
outstep-gated console + unit-20 log line: iter/lrate/LL/nd/Dsum/timing-ETA, amica15.f90:1005-1012): pamica's verbose tqdm shows LL/lrate only, and no on-disk log exists on any array backend.
Also from the audit, for the record when this is picked up: .context/feature_parity.md's "History tracking ... grad norms" row implies a tracked gradient-norm trajectory that does not publicly exist on torch/MLX. Scope decision (port, partial, or record-as-deliberate) left open.
From the epic #278 feature-parity audit (Fortran-coverage pass). Two related reporting gaps vs amica15.f90, unrecorded:
write_nd(amica15.f90:658, 1760-1761, output at 2398-2429): the per-component, per-iteration squared-gradient trajectory. pamica exposes only the scalar RMS aggregate (ndtmpsumfeeding themin_ndstop); the legacy NumPy backend keeps a public per-iterationndlist that torch/MLX lack. Diagnosing WHICH sources converge slowly is not possible on the primary backends.outstep-gated console + unit-20 log line: iter/lrate/LL/nd/Dsum/timing-ETA, amica15.f90:1005-1012): pamica'sverbosetqdm shows LL/lrate only, and no on-disk log exists on any array backend.Also from the audit, for the record when this is picked up:
.context/feature_parity.md's "History tracking ... grad norms" row implies a tracked gradient-norm trajectory that does not publicly exist on torch/MLX. Scope decision (port, partial, or record-as-deliberate) left open.