[dekobon]

Closes #1724.

What

mdBook 0.4.40 + mdbook-i18n-helpers 0.3.6 → 0.5.4 + 0.4.0, in pages.yml, the Makefile book targets, the Dockerfile, the deploy script and docs/development/translations.md together. The -std=gnu17 CFLAGS override from 36a9dec is gone.

Two commits: the toolchain bump, and a separate make book-po-update refresh of ja.po (990 fuzzy / 309 new / 203 obsolete — ten commits of English doc drift since 2026-10-21, not toolchain churn; msgmerge against the old toolchain's pot yields a byte-identical file).

The part the issue did not anticipate

Helpers 0.4.0's lockfile has no onig_sys at all: upstream switched syntect to the pure-Rust fancy-regex (google/mdbook-i18n-helpers#267). That removes the GCC 15 problem, but under fancy-regex the shell/YAML/TOML grammars drop lines from consecutive # comment runs (google/mdbook-i18n-helpers#321, open):

```bash / # a1 / # a2 / # a3	extracted
0.4.40 + 0.3.6 (previous CI)	# a1\n# a2\n# a3\n
0.5.4 + 0.4.0 --locked (issue's plan)	# a1\n, # a3\n — # a2 dropped
0.5.4 + 0.4.0 --locked --features syntect/regex-onig	# a1\n# a2\n# a3\n

On this book the --locked-only pot loses 90 translated entries and eleven /ja/ pages fall back to English. With Oniguruma the pot is byte-identical (modulo POT-Creation-Date) to the previous toolchain's. So the helper is installed with --features syntect/regex-onig; the two onig crates resolve fresh under --locked (69.9.3, which carries oniguruma's GCC 15 fix) while everything else stays pinned.

Both builds report the same --version, so new utils/check-xgettext-backend.sh tells them apart behaviourally on a four-line book; make book-pot and the build-book job both run it before extracting anything. Verified to reject the fancy build, a stub emitting two split entries, and a missing helper; to accept both Oniguruma builds.

Other things 0.5 exposed

• -d is now cwd-relative. The workflow's -d book/ja would have written to ./book/ja and shipped a Pages artifact with no /ja/ tree, silently. Fixed in the workflow and make book-ja; the workflow now also asserts the /ja/ tree exists and carries a known translation across its pages (a missing/mis-built gettext preprocessor is optional and yields English under lang="ja" with no error).
• book.toml unchanged. 0 heading-id changes in either language; fragment-link breakage unchanged (3 en / 8 ja, pre-existing); 0 non-punctuation text differences in English. Visible deltas: 0.5's smart-punctuation default (kept), and \| inside table code spans now renders correctly on /ja/.

Not verified locally

GCC 15 / C23 compilation of onig_sys 69.9.3 — host is aarch64/GCC 11, x86 emulation of rustup crashed. Inferred from the oniguruma HISTORY entry in the vendored tree. The build mdBook job on this PR is the first real test. If it fails, the fallback is a scoped CFLAGS override again; a tooling lockfile is not available to cargo install.

Review

Three astra-review rounds plus one independent review (8 findings total, all fixed; round three clean). make pre-commit green. Evidence (pots, fixture, both toolchains' en/ja builds, probe scripts): ~/bca-evidence/1724/.

[dekobon]

The one thing I could not verify locally is now verified in CI. The build mdBook job on this PR (run 41142624729, ubuntu-26.04) was a cold install — no rust-cache hit carried a prebuilt helper — and its log shows:

Downloaded onig_sys v69.9.3
 Compiling onig_sys v69.9.3
  Finished `release` profile [optimized] target(s) in 40.40s

with no CFLAGS override, so oniguruma 6.9.10's C23 fix does compile under GCC 15. Every new step passed on the first run: the backend probe (check-xgettext-backend: ok — consecutive comment lines extract as one block), both book builds, and the /ja/ translation sentinel. The whole Pages workflow (including bca self-scan) is green.