refactor(repo): move helper scripts into utils/

[dekobon]

What

Moves the twelve helper scripts that sat loose in the repository
root into utils/, where check-tools.sh and
deploy-book-to-gh-pages.sh already lived. The root now lists
project structure rather than maintenance tooling.

Moved to utils/	Purpose
check-versions.py + -test.py	Lockstep version invariant across owned crates and doc pins
check-snapshot-anchors.py	Blocks new bare insta::assert_json_snapshot! in src/metrics/
check-manpage-assets.py	Every man/bca*.1 present in the deb+rpm asset lists
check-grammar-marker-sync.py + -test.py	mozjs/mozcpp marker vs bundled parser.c (#400)
check-grammar-crate.py + -test.py + check-grammars-crates.sh	Grammar-bump metric diff; the test guards EXTENSIONS against src/langs.rs
check-enums-codegen-drift.sh + -test.py	enums codegen drift (#405)
verify-name-only-churn.py	Proves a bulk snapshot regen is path-only churn

recreate-grammars.sh stays at the root — nothing moved is coupled
to it. The two shell scripts move rather than splitting each from
the test that exercises it (check-enums-codegen-drift.sh) and the
script it calls (check-grammars-crates.sh).

The parts that are not a rename

Three things a plain git mv plus a path rewrite would have broken
silently:

• Repo-root anchoring. Every script derived its paths from
  Path(__file__).resolve().parent, which was the repository root
  only by accident of where the file sat. That becomes parents[1],
  so each still resolves the root from its own location and runs
  correctly from any cwd.
• The two self-tests that stage a copy of their gate into a
  tempdir (check-grammar-marker-sync-test.py,
  check-enums-codegen-drift-test.py) relied on the same
  coincidence: the copy sat at the fixture root. Under parents[1]
  that resolves to the tempdir's parent, so the copy now stages under
  <tmpdir>/utils/. Both suites failed loudly at the intermediate
  step, which is how this surfaced.
• .pre-commit-config.yaml files: regexes, not just the
  entry: commands. Several are ^-anchored, so a hook such as
  grammar-marker-sync-test would have silently stopped firing on
  edits to its own script.

Callers updated

Makefile, .pre-commit-config.yaml, .github/workflows/ci.yml,
.github/workflows/release.yml, .taskcluster.yml, and
check-grammars-crates.sh's own call sites. Prose in AGENTS.md,
CONTRIBUTING.md, RELEASING.md, tests/README.md,
docs/development/output_name_normalization_design.md, the book's
recipes/exporting-data.md, and .claude/skills/add-lang/SKILL.md.
The --update hints the gates print to users carry the new prefix
too, as does the header written into
.grammar-marker-baseline.toml. CHANGELOG.md history is left
alone.

Reviewer note on .snapshot-anchor-baseline.txt

Regenerated because its header embeds the script path. The regen
also corrects two rows that were stale before this change, which is
the one non-rename content change in the diff:

• a phantom src/metrics/exit.rs entry left behind by the rename to
  nexits.rs (count 0 either way);
• src/metrics/npa.rs ratcheting 13 → 12, after an anchor landed
  without a baseline refresh.

Both are corrections toward the live measurement. Say the word if
you would rather the move commit be a pure path change and these
land separately.

Validation

make pre-commit green end-to-end, plus a targeted lint re-run for
the few edits that landed after that run started. Each gate and each
self-test was also run individually from the new location.

ja.po is untouched per the translation policy — the changed
exporting-data.md paragraph falls back to English on /ja/ until
the next make book-po-update.