deps: bump impl pin to close the CPU/CUDA kernels skew

[balbasty]

What

Moves the impl gitlink from a099a04 to 83fa026 (current fastfields-cuda-implmain). Gitlink-only — exactly one modified path.

The skew this closes

Traced by walking the actual gitlinks on main rather than trusting the submodule branch labels:

CPU path : lib -> cpu-lib dbd3274 -> cpu-impl 41d66b3 -> kernels be7be08
CUDA path: lib -> cuda-lib 6ae52bd -> cuda-impl a099a04 -> kernels b09b284

The two backends genuinely resolved to different kernels commits, with CUDA three behind (b09b284..be7be08):

kernels PR	commit	what CUDA was missing
#74	32951da	heap over-read from the past-the-end FaceIterator — a memory-safety fix
#76	3e38c85	make the host BVH/normal builders visible to nvcc's host pass — a CUDA-specific fix that the CUDA path itself lacked
#78	be7be08	drop the pointless virtual destructors from Node/BoundingSphere

A detail worth recording: cuda-impl's own main already pinned 3e38c85, so the stale link was this repo's impl pin, which was 11 commits behind cuda-implmain. That is why the CUDA path resolved to a third kernels commit that neither leaf repo pointed at directly.

What this pin picks up

Beyond closing the kernels skew (b09b284 → 1df9fd3), advancing impl by 11 commits also brings in the mesh sdt launcher/dispatch work and cuda-impl#48 (the eight missing #pragma once guards).

a099a04 is an ancestor of 83fa026, so this is a fast-forward of the pin.

Resulting state

cuda-lib (this PR) -> cuda-impl 83fa026 -> kernels 1df9fd3
cpu-lib 1fb2b37 -> cpu-impl 5e2c78e -> kernels 1df9fd3

Both backends on one kernels commit. The matching fastfields-lib bump follows, after which all six repos agree.

Verification

Per this project's standard — shared kernel math via the CPU suite, CUDA-only glue via nvcc compile+link, since there is no GPU in CI:

• Shared math: full CPU suite over kernels 1df9fd3 — 13/13 suites, 0 failures. The new boundary-symmetry regression tests (cpu-lib#89) fail with 328 + 1696 failures against the pre-fix kernels and pass at 1df9fd3.
• CUDA compile+link: fastfields-cuda-impl's build-via-cuda-lib job already passed against exactly this cuda-impl content — twice, on #48 (5c68cf9) and #49 (83fa026), each a full make -C . -j2 CXX=clang++ CXXFLAGS="-std=c++14 -O1" run. This PR's own build job re-confirms it from this side.
• -O1 and the split MODULES list (reg_field/reg_field_rls/reg_flow/reg_flow_rls separately, to cap ptxas memory) are untouched.

Pin hygiene

83fa026 is the squash-merge commit on cuda-implmain, not a pre-merge branch head — verified it resolves before opening.

🤖 Generated with Claude Code

https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z

---

Generated by Claude Code

[github-actions]

Submodule staleness (non-blocking -- fastfields-lib#15)

submodule	pinned commit	tracks	status	behind by
impl	83fa026	fastfields/fastfields-cuda-impl@main	up to date	0