refactor(pushpull): delete 5,554 lines of unreachable legacy pushpull kernels

[balbasty]

What

The teeny-based single-voxel path (pushpull/teeny.h, vox::pull/push/count/grad) is now the only pushpull implementation the impl layers use. Everything else under pushpull/ was left behind by that migration and is unreachable from every live dispatch path. This PR removes it.

file	lines removed	what
pushpull/nd.h	-489	whole file
pushpull/3d.h	-1814	whole file
pushpull/2d.h	-1229	whole file
pushpull/1d.h	-1180	the Spline<Z>/<L>/<Q>/<C> (Nearest/Linear/Quadratic/Cubic) partial specialisations only
pushpull/utils.h	-842	PushPull1D/2D/3D, PushPullSelect/PushPull, the InFOV<> block + FF_EXTRAPOLATE_TINY, PushPullAnyUtils, PushPullUtils<Z|L|Q|C>
pushpull.h	-2	umbrella now includes only pushpull/1d.h
CLAUDE.md	+2/-1	layout note updated
total	-5,554 (kernel code)

Kept (still reachable): Config, the primary Kernels declaration, the generic Kernels<Config<one, Spline<I>, Bound<B>, ABS>> in 1d.h, and PushPullMaybe / SplineBufSize / the generic PushPullUtils<S,B,ABS> in utils.h.

Verification methodology

Re-derived from scratch, symbol by symbol, over fastfields-kernels, -cpu-impl, -cpu-lib, -cuda-impl, -cuda-lib, -lib and -bind-py (sources and tests):

1. Include graph.pushpull/{1d,2d,3d,nd}.h and pushpull/utils.h are reachable only through the pushpull.h umbrella, and the only #include "../pushpull.h" in the entire tree is distance/spline.h. nd.h is not included even by the umbrella — and it specialises a type PushPullConfig that does not exist anywhere in the tree, so it could not compile if it were.

2. Instantiation sites. The only instantiation of pushpull::Kernels anywhere is distance/spline.h:38:

   usingPP = pushpull::Kernels<pushpull::Config<1, Spline<S>, Bound<B>>>;

   D == 1 is hard-coded → Kernels<Config<two,…>> / <three,…> / the nd form can never be instantiated.

3. Which S/B? Every caller of that — cpu-impl/distance_spline.h and cuda-impl/distance_spline.h, driven by cpu-lib/distance.cpp and cuda-lib/distance.cpp_dt_spline_{table,brent,gaussnewton} — instantiates with spline_t::Dynamic, bound_t::Dynamic. Dynamic matches neither Z/L/Q/C nor a static bound, so overload resolution always picks the genericSpline<I>/Bound<B> specialisation. The static-order specialisations in 1d.h (and correspondingly PushPullUtils<Z|L|Q|C>) were unreachable. The generic one already handles a static S/B correctly (I == Dynamic ? runtime : I), so no behaviour is lost — only an unused compile-time fast path.

4. The live pushpull ops don't touch this code at all.pushpull::pull/push/count/grad (cpu-impl/pushpull.h, cuda-impl/pushpull.h), dispatched by cpu-lib/pushpull.cpp, cuda-lib/pushpull.cpp and fastfields-lib/pushpull.cpp, route exclusively to kernels/pushpull/teeny.h; they never name Kernels, Config or PushPull*.

5. Residual-reference sweep.grep for PushPull1D, PushPull2D, PushPull3D, PushPullSelect, PushPullAnyUtils, PushPullUtils<Z|L|Q|C>, InFOV, PushPullConfig and the pushpull/{2d,3d,nd}.h include paths across all seven repos: zero remaining references after the deletion (the only surviving InFOV / FF_EXTRAPOLATE_TINY hits are teeny.h's own macro and one comment).

6. No test depended on deleted code.tests/test_pushpull.cpp goes through the lib API → teeny.h. tests/test_distance_spline.cpp goes through dt_spline_* → the generic 1d Kernels, which is kept and still covered (704 checks).

Test results

make -C fastfields-cpu-lib clean test on the same base commit, before and after:

test	before	after
test_distance	2352 / 0	2352 / 0
test_distance_mesh	4622 / 0	4622 / 0
test_distance_spline	704 / 0	704 / 0
test_posdef	5092 / 0	5092 / 0
test_pushpull	326 / 0	326 / 0
test_reg_field	465 / 4 *	465 / 4 *
test_reg_flow	854 / 16 *	854 / 16 *
test_reg_op	186 / 0	186 / 0
test_resize	4130 / 0	4130 / 0
test_restrict	191 / 0	191 / 0
test_splinc	4577 / 0	4577 / 0

Byte-identical for bothCXX=clang++ and CXX=g++.

* the diag_bending failures are pre-existing on this base and unrelated to pushpull — they are fixed on the separate claude/regularisers-diag-bugfix branch. They fail identically before and after this change.

CUDA: nvcc -std=c++17 compile of fastfields-cuda-lib/distance.cpp (the CUDA TU that transitively includes pushpull.h) still succeeds (warnings unchanged). No GPU available, so CUDA is compile-verified only, as usual.

Note vs. the earlier estimate

The prior review estimated ~4,910 dead lines. The actual verified figure is 5,554 — the difference is utils.h, where the dead region turned out to be 842 lines (the four PushPullUtils<Z|L|Q|C> specialisations alone are ~550) rather than the ~200 implied. Nothing that was flagged as dead turned out to be live.

Consumers

No changes were needed in fastfields-cpu-impl, -cpu-lib, -cuda-impl, -cuda-lib or -lib — none of them referenced any deleted symbol.

---

Generated by Claude Code

[balbasty]

Agent:claude-fastfields-to-teeny — opened from the fastfields teeny-refactor session (claude/fastfields-teeny-refactor-js42id). Flagging for disambiguation from the concurrent claude-jitfields-to-fastfields agent (claude/jitfields-fastfields-migration-v5r416), which is porting jitfields to fastfields directly, without teeny.

---

Generated by Claude Code