Progress update — field_relax (the top "Remaining" item) is now implemented: C++ layers already had it; the Python-stack wiring landed as a 5-PR series (fastfields-dlpack#21 merged, fastfields-cupy#17 merged, fastfields-numpy#22/fastfields-torch#24/fastfields#26 open but code-complete — held only because their CI needs a released fastfields-dlpack to install against, which is whl#1's blocker, not a code problem).
Also: the RLS/JRLS "weighted symmetry oracle" this issue lists as a prerequisite for the last "(later)" item is now done at the C++ level — field_matvec_rls/diag_rls/relax_rls dispatch wired with a symmetry test oracle, plus two real bugs found and fixed since this issue was written (fastfields-kernels#40 membrane-scale bug, fastfields-cpu-lib#65 JRLS per-channel-weight bug). What's still open on that front is whether the RLS/JRLS variants are exposed through the Python wrappers the same way field_matvec/field_relax are — not yet confirmed, worth a follow-up check before closing this umbrella.
Remaining, unchanged: field_kernel accumulate variants (niche), and the Python-level RLS/JRLS exposure check above.
Workstream: claude-jitfields-to-fastfields
Progress update —
field_relax(the top "Remaining" item) is now implemented: C++ layers already had it; the Python-stack wiring landed as a 5-PR series (fastfields-dlpack#21merged,fastfields-cupy#17merged,fastfields-numpy#22/fastfields-torch#24/fastfields#26open but code-complete — held only because their CI needs a releasedfastfields-dlpackto install against, which iswhl#1's blocker, not a code problem).Also: the RLS/JRLS "weighted symmetry oracle" this issue lists as a prerequisite for the last "(later)" item is now done at the C++ level —
field_matvec_rls/diag_rls/relax_rlsdispatch wired with a symmetry test oracle, plus two real bugs found and fixed since this issue was written (fastfields-kernels#40membrane-scale bug,fastfields-cpu-lib#65JRLS per-channel-weight bug). What's still open on that front is whether the RLS/JRLS variants are exposed through the Python wrappers the same wayfield_matvec/field_relaxare — not yet confirmed, worth a follow-up check before closing this umbrella.Remaining, unchanged:
field_kernelaccumulate variants (niche), and the Python-level RLS/JRLS exposure check above.Workstream: claude-jitfields-to-fastfields