Follow-up to D6: one-shot clang-format reformat of the six C++ repos

[balbasty]

Follow-up to #18 item D6. The .clang-format + lint gate PRs
(fastfields-lib#48 and its five siblings) deliberately did not reformat
any existing source, so their clang-format CI check is scoped to the lines a
PR touches. This issue tracks the one-shot whole-tree reformat that would let
the check be widened to clang-format --dry-run --Werror over everything.

Why it was split out

Applying the committed .clang-format to the current trees rewrites
essentially the whole codebase:

repo	changed lines	files differing
fastfields-kernels	24 727	47 / 54
fastfields-cpu-impl	7 528	12 / 12
fastfields-cuda-impl	9 948	12 / 12
fastfields-cpu-lib	11 517	32 / 32
fastfields-cuda-lib	6 852	18 / 18
fastfields-lib	3 181	20 / 20

(Measured with the committed config. Stock LLVM and Google styles are both
worse, ≈89 k changed lines each, which is why the config was derived from the
existing code instead — 4-space indent, 80 columns, T * p / T & r middle
alignment, Stroustrup braces.)

That is not residual sloppiness that a better config could absorb. Most of it
is deliberate hand column alignment that clang-format cannot reproduce by
design:

void sym_matvec(
 DLTensor & out ,
 const DLTensor & hessian ,
 const DLTensor & inp ,
 int stream )

and the backslash-aligned dispatch macro blocks in the *-lib layers. So the
reformat is a real judgement call, not a mechanical one.

What needs deciding

1. Is it wanted at all? The aligned parameter lists and macro tables are
   readable and consistent; clang-format's output is not obviously better. A
   permanent // clang-format off around the macro blocks is a middle option.
2. When. A whole-tree reformat invalidates every open PR's diff and makes
   git blame noisy across the layers where the math lives. Best done when few
   branches are in flight, and worth pairing with .git-blame-ignore-revs.
3. Scope. Possibly the *-lib dispatch layers only, or kernels only,
   rather than all six at once.

Not blocking anything

The gate that landed already prevents new drift, so this can sit until
someone wants it. Deliberately not doing it as part of the tooling PRs, per
CONTRIBUTING's "never bundle".

Workstream: claude-jitfields-to-fastfields

---

🤖 Generated with Claude Code

[balbasty]

Triage for the 0.0.dev0 release: confirmed non-blocking, deferred. This is a pure formatting/whitespace change with no semantic or correctness impact, and the issue's own text already says as much ("Not blocking anything... the gate that landed already prevents new drift"). The judgement calls it raises (whether to reformat at all, when, and whether to scope it to specific layers) are exactly the kind of thing worth deciding when few branches are in flight — i.e., after the initial release settles, not before. Not actioning it. Leaving open.

---

Generated by Claude Code

[balbasty]

Closing as won't-implement per maintainer decision: this is a pure cosmetic clang-format reformat across six repos with no functional/correctness benefit, and reformatting now would create unnecessary diff noise against in-flight work (including the parallel claude-fastfields-to-teeny refactor). Not planned.

---

Generated by Claude Code