Audit agent guidance against the Opus 5 prompting guide

[dekobon]

Audits the repository's agent-facing configuration against the
Opus 5 prompting guide,
and fixes two tooling bugs the audit uncovered along the way.

Agent guidance

Three things the guide calls for were absent entirely, and are now in
CLAUDE.md: narration cadence during a task, a length rule for written
deliverables, and a delegation policy.

The delegation policy caps concurrent isolation: "worktree" agents at
six. improve-crate and cleanup-crate said "launch all area
agents in parallel"; batch-fix had no cap at all, and its waves were
bounded only by distinct crate count — this workspace has twelve crate
manifests, so a wave could start a dozen workspace clones. All three now
encode the same six.

The guide warns that explicit re-verification compounds with the model's
own checking rather than adding to it. Three such loops are removed (two
in batch-fix, one in doc-author). What stays is everything that
looks at the work from outside the context that produced it: the
domain checklist, batch-fix's Phase 3 review, and doc-author's
fresh-context reviewer. simplify-rust now scales its fan-out to the
diff instead of always spawning three agents for a two-file change.

Complexity feedback is now explicitly scoped to the edit that triggered
it — the guide's task-scope point applied to this repo's own hook — in
all three copies of the shipped guidance block plus the condensed
opencode strings.

Left alone deliberately: the review and audit-* skills already
do find-then-filter without severity gating, which is exactly what the
guide recommends. The guidance block's "do not …" framing also stays: it
names specific gaming moves, and the guide's preference for positive
examples is about narration style, not prohibitions.

Bugs found

make book-pot has been broken under mdBook 0.5. It passed a
relative -d po, which 0.4 resolved against the book root and 0.5
resolves against the cwd, so the pot landed in ./po/ at the repository
root and book-po-update failed on a missing file.

The book's sample Claude Code hook would break for anyone who pasted
it — unguarded ${CLAUDE_PROJECT_DIR} under set -euo pipefail, where
the repo's own shipped hook guards it. The sample had drifted.

Japanese

ja.po read as "3348 translated, 0 fuzzy" only because the broken
book-pot meant msgmerge had not run since those English edits
landed. Against a correct pot the real drift was 42 entries, now all
translated: 3382 translated, 0 fuzzy, 0 untranslated, 0 obsolete.

A version guard was added because getting this wrong is quiet.
mdbook-i18n-helpers 0.3.x under mdBook 0.5.x aborts loudly, but helpers
0.4.x extracts fenced code blocks one entry per line instead of one per
block — extraction succeeds, and the next msgmerge shreds ja.po
against msgids the pinned toolchain never produces. That happened during
this work and was caught by installing the pinned pair; book-pot now
refuses to run outside mdBook 0.4.x, and translations.md pins both
halves and describes both failure modes.

Validation

markdown-lint, makefile-check, shellcheck, toml-lint,
actionlint, snapshot-anchors, check-versions, check-manpage-assets,
grammar-marker-sync, self-scan, and msgfmt --check all pass.
shellcheck and bash -n pass on the sample hook extracted from its
fence. The Japanese book builds and renders the new content, and
make book-po-update is idempotent.

The cargo stages of make pre-commit could not be run: cargo fmt --all walks into the vendored grammar crates, which are not workspace
members, so manifest discovery escapes any worktree under
.claude/worktrees/ and resolves to the main checkout. This predates the
branch — the diff touches no .rs or Cargo.toml file — but it means
make pre-commit cannot complete in the worktree mode that batch-fix,
improve-crate, cleanup-crate, and audit-crate all document. Worth
its own issue; CI will cover the cargo gates for this PR.