fix: force the RLS weight-channel stride to 0 for broadcast reads

[balbasty]

Summary

field_matvec_rls/field_diag_rls/field_relax_rls (the wc == 1
"single weight shared across all nc field channels" RLS path, for
absolute/membrane/bending) computed the weight tensor's
per-channel stride (wsc) from its real, contiguous last-dim stride
and then indexed wgt[wsc*c] inside a loop over the field'snc
channels. That's the right thing to do for the genuine per-channel
(JRLS, wc == nc) path, but wrong for RLS: the weight tensor
genuinely has only one element in that dimension, so for c > 0 the
read walks off the wc=1 buffer into the next spatial voxel's
weight (or out of bounds at the very last voxel, landing on
zero-initialized memory and reading -0).

Since these nine functions (matvec/diag/relax_ ×
absolute/membrane/bending) are only ever dispatched to when
wc == 1, the fix forces wsc = 0 unconditionally instead of
deriving it from stride_wgt — exactly the broadcast semantics the
RLS path is supposed to have. wgt[0*c] == wgt[0] for every c, so
this needs no change to the kernels-layer math.

Root cause, precisely

Confirmed with a minimal, direct field_matvec_rls harness at
C=2, wc=1, order=1 (absolute-only, no neighbour terms, so the bad
read is maximally localized) with a non-constant weight map (the
existing regression test uses an all-ones map, which masks the bug
almost everywhere by coincidence): channel 1 read the next spatial
voxel's weight instead of its own, at every interior voxel, and 0
(out of bounds) at the last voxel — exactly the wgt[wsc*c]
stride-indexing bug described above. This is a distinct bug from
kernels#40 (which only touched make_kernel_bending_rls, i.e.
order=3) — it explains the order=1/order=2 failures kernels#40
never could have.

flow_matvec_rls/diag_rls/relax_rls are not affected: flow's
wgt is architecturally always wc=1 (a CHECK_SAME(wgt.shape[...] == 1) at the cpu-lib API boundary — there is no flow JRLS variant),
and the flow kernels never index by a per-component weight stride in
the first place, so there's no equivalent bug to trigger.

A second, separate bug found while tracing this (not fixed here)

The _jrls (wc == nc, genuine per-channel-weight) counterparts
never thread a per-channel stride through to the kernels at all —
they read a single weight value once and apply it to every field
channel, silently ignoring channels 1..nc-1. That's real but
untested (the existing tests either use identical weights across
channels or only check symmetry/self-consistency, both blind to it),
needs kernel-level changes rather than a stride fix, and is out of
scope for this PR — will be filed as a follow-up issue.

Verified

make -C fastfields-cpu-lib test CXX=clang++

All 12 suites pass, 0 failures (was 29/8440 failing in
test_reg_field before this fix):

running ./build/test_reg_field
reg_field module CPU tests
checks: 8440, failures: 0
PASSED
running ./build/test_reg_flow
reg_flow module CPU tests
checks: 11415, failures: 0
PASSED

(plus distance/distance_mesh/distance_spline/posdef/pushpull/
pushpull_backward/reg_op/resize/restrict/splinc, all green.)

part of fastfields/fastfields-cpu-lib#62

Workstream: claude-jitfields-to-fastfields

🤖 Generated with Claude Code

---

Generated by Claude Code

[balbasty]

Review summary (triage, jitfields-to-fastfields): merged as squash.

Root cause confirmed precise and narrow: the 9 RLS driver functions (matvec/diag/relax_ × absolute/membrane/bending) computed the weight channel-stride as stride_wgt[nall] — correct for JRLS (wc == nc) but wrong for RLS, which is the only mode these functions are ever called in (wc == 1, one weight broadcast across all field channels). For c > 0 this walked off the 1-element weight buffer into the next voxel's weight (or OOB at the last voxel, landing on zeroed memory — explains the exact -0 the issue's order=1 failure showed). Fix forces wsc = 0, which is exactly correct broadcast semantics.

Verified independently, not just trusted: real kernel-level harness at C=2/wc=1/order=1 confirmed channel 1 was reading the neighboring voxel's weight before the fix; test-via-cpu-lib CI green.

Workstream: claude-jitfields-to-fastfields

---

Generated by Claude Code