pushpull: route bound/spline dispatch through static/Dynamic selector macros + add SPLINEFLAGS

[balbasty]

Workstream: claude-jitfields-to-fastfields

Part of fastfields/fastfields-cuda-lib#30 (wire pushpull into cuda-lib's CI). This is the cpu-lib layer piece. Default policy is unchanged (everything static) — plumbing to let a different build policy (which cuda-lib now uses by default) actually take effect.

• Dispatch macros (PP_BOUND/PP_ORDER, both the full and FF_TEST_SPARSE variants) now instantiate FF_BOUND_<NAME>/FF_SPLINE_<NAME> instead of bound::type::<NAME>/spline::type::<NAME> directly, so a Dynamic build policy (BOUNDFLAGS/SPLINEFLAGS) routes the flagged conditions/orders through the shared bound::type::Dynamic/spline::type::Dynamic instantiation instead of a dedicated one — mirrors how reg_field.cpp/reg_flow.cpp already use FF_BOUND_<NAME> (reg_field/reg_flow: select static-or-Dynamic bounds per build (BOUNDFLAGS) #46).
• pull/push/count/grad build a bound::BoundVec + spline::SplineVec from the runtime bound/spline arguments and forward them through the leaf wrappers into the impl launchers, so whichever axis ends up Dynamic can recover its actual condition/order at run time.
• Added SPLINEFLAGS (default empty = all 8 orders static, unchanged behaviour), the spline-order analogue of BOUNDFLAGS, wired into both the library and test compile rules.

Depends on fastfields/fastfields-cpu-impl#38 (bumped the impl pin to it), which in turn depends on fastfields/fastfields-kernels#44.

Test plan

• test_pushpull: 308/308 under the default (unchanged) all-static policy.
• test_pushpull: 308/308 under -DFF_STATIC_SPLINES=0 -DFF_STATIC_SPLINE_{NEAREST,LINEAR,QUADRATIC,CUBIC}=1 -DFF_STATIC_BOUNDS=0 -DFF_STATIC_BOUND_{DCT2,DST2}=1 — the exact policy fastfields-cuda-lib now ships by default. This run originally caught two pre-existing kernels-layer bugs (fixed in pushpull: implement spline::type::Dynamic + fix two Dynamic-order/-bound bugs fastfields-kernels#44); 308/308 after the fix.
• test_pushpull: 308/308 under a fully-Dynamic policy (-DFF_STATIC_SPLINES=0 -DFF_STATIC_BOUNDS=0).

🤖 Generated with Claude Code

---

Generated by Claude Code

[balbasty]

Looks good, with one question:

Can we get rid of the bound/order sparse macros and fold everything using the per-bound/order defines? I.e in the make file, have a sparse (or test) target that uses a different set of bound/order defines than the release tag?

[balbasty]

Addressed the feedback on folding FF_TEST_SPARSE into the per-bound/order defines (pushed to claude/cuda-pushpull-dynamic-spline):

• Removed the #ifdef FF_TEST_SPARSE ... #else ... #endif duplicate PP_ORDER/PP_BOUND dispatch-macro block in pushpull.cpp entirely — there is now exactly one unconditional set of macros, always built from FF_SPLINE_<NAME>/FF_BOUND_<NAME> (the same per-bound/order define mechanism BOUNDFLAGS/SPLINEFLAGS already drive).
• Instead of the sparse macro, the test Makefile target now sets its own defaults for BOUNDFLAGS/SPLINEFLAGS (target-specific variables, e.g. test: BOUNDFLAGS = -DFF_STATIC_BOUNDS=0 -DFF_STATIC_BOUND_DCT2=1 -DFF_STATIC_BOUND_DST2=1) distinct from the top-level release default, while still yielding to a command-line override (make test BOUNDFLAGS=...). CI's 3 matrix legs (static/dynamic/cuda-default) now pass explicit BOUNDFLAGS/SPLINEFLAGS for both axes.
• -DFF_TEST_SPARSE itself is left alone in TESTCPPFLAGS where resize.cpp/restrict.cpp still use it independently — out of scope for this fold, confirmed via grep that pushpull no longer references it at all.
• Verified: full test suite (11 binaries) passes under all 3 policies (308/308).

Checked fastfields-cuda-lib for an analogous sparse-macro/test-target to fold — it has no FF_TEST_SPARSE and no test target at all, so there's nothing to consolidate on that side.

(GitHub's GraphQL review-thread API is rate-limited for this account right now, so I'm leaving this as a regular comment rather than an inline reply.)

---

Generated by Claude Code