chore: delete the unused impl/kernels/vector/ headers

[balbasty]

"is vector/ used anywhere? We can get rid of it if not (it will be
superseded by teeny in the future)."

It is not. 13 files, −2,421 lines: the 11 headers, the scratch program that
was their only consumer, and the Makefile target that built it.

The proof, re-run on de288a9

Not inherited from an earlier survey — main has moved (#143 and #146 merged
since), so this was re-measured at the commit the deletion is made against.

The decisive check is the preprocessor's own dependency graph, not grep.
clang++ -MM over all 46 translation units the build compiles resolves
zero of the 11 headers, directly or transitively:

	TUs scanned	reach impl/kernels/vector/
host (src/lib, src/lib-cpu, tests/lib, tests/lib-cpu)	34	0
nvcc flags (src/lib-cuda, tests/impl-cuda)	12	0

Control, so the scan is not silently finding nothing:

$ clang++ -std=c++11 -I include -MM tests/kernels/vector/test.cpp | grep vector/
impl/kernels/vector/abstract_ptr.h
impl/kernels/vector/abstract_sized.h
impl/kernels/vector/abstract_vector.h
impl/kernels/vector/concrete_vector.h ...

The scratch program does show up. The method detects a real reference.

Why grepping for type names is not conclusive, and why the include closure
is the right question.vector/ declares type, value, pointer,
iterator, Vector, Impl, final_type, this_type — all of which occur
in unrelated code (reg_flow.h has its own Impl, posdef.h its own
guess_type, mesh_utils.h its own final_type/this_type). A symbol grep
returns ~1,966 hits, none of them real. In a header-only library the only way
to reach a declaration is to include the header carrying it, so the include
closure settles it and the symbol grep cannot.

Not part of the gate.Makefile called it "a hand-run scratch program …
compiled (not run)"; it appears nowhere in tools/test-baseline.expected and
in no CI job.

What is lost

Recoverable at de288a9, the last commit where the files exist:

git show de288a9:include/fastfields/impl/kernels/vector/abstract_vector.h

Four things worth knowing before anyone goes looking:

• It was never developed here. Imported wholesale in 7da9e7e ("Copy all
  lib files from jitfields", 2025-07-18), given a mechanical CUDA-annotation
  pass the same day in 865d1c5, and never functionally touched again. The
  only three commits since are tree-wide sweeps — the consolidation path
  rewrite, the FF_ macro prefixing, and refactor: spell the public interface <fastfields/...> #146's include spelling. Thirteen
  months, zero functional changes.
• The one idea with no equivalent in the surviving tree is
  SwitchStride<T, S, D> and its DynamicStride specialisation: stride as a
  compile-time-or-runtime axis — the same pattern bound::type::Dynamic and
  spline::type::Dynamic apply to boundary conditions and spline orders.
  distance/mesh_utils.h, which otherwise duplicates this hierarchy, has no
  such axis. If the teeny work wants a reference for a strided-view
  abstraction, that is the piece to read.
• Also going: an owning container (concrete_vector.h, new T[size])
  alongside the non-owning views — everything in mesh_utils.h is a view over
  caller memory — and stream.h's ostream printing for debugging.
• The two hierarchies are stylistic siblings, not a redesign. I expected
  to find that mesh_utils.h replaced virtuals with CRTP and that this was
  why it won; it did not. Both use FF_CUHOSTDEV virtual abstract bases (34
  virtuals here, 30 there). Whatever decided which survived, it was not that.

Deliberately not touched

• tools/consolidate.sh still names these paths in four places. It is the
  frozen record of the six-repo migration and describes a tree as it stood on
  2026-08-19; editing it would falsify history rather than tidy it.
• ci.yml's ^tests/kernels/ path filter stays.
  tests/kernels/atomic/test.cpp remains and is built by the tsan leg, so
  the filter is still load-bearing. Only the vector/ subdirectory goes.

Verification

• All 46 TU include graphs still resolve after the deletion.
• make test-atomics — the other program under tests/kernels/ — still
  builds and passes (4 checks).
• The gate must not move, and a non-gating target's removal leaving
  59,886 / 13 / 0 untouched is itself the confirmation. CI is the authority
  here; result recorded in a comment below.

Note for the teeny agent

teeny is 159 commits ahead and may still reference these headers. I have not
consulted or accommodated that branch — its agent owns its own merges — but
flagging the deletion here so it is not a surprise at merge time.

---

Generated by Claude Code

[balbasty]

can merge