feat(i18n): deterministic integrity/typography checks and real per-language style guides

.github/workflows/i18n-lint.yml

@@ -48,3 +48,8 @@ jobs:
         run: |
           python3 -m pip install --quiet pyyaml
           python3 hack/i18n/test_i18n.py
+      - name: Typography lint on published translations
+        # Advisory: reports per-language typography drift (hand edits, pages
+        # translated before a rule existed) without failing the build. Flip to
+        # --strict per language once its backlog is clean.
+        run: python3 hack/i18n/lint_translations.py

hack/check-i18n.sh

@@ -4,9 +4,17 @@
 #
 # Subcommands:
 #   check           (default) Run every guard; exit non-zero on any gap.
-#   update-digests  Recompute and rewrite `source_digest` in every translated
-#                   page from its English source. Run this after editing an
-#                   English source or after adding/refreshing a translation.
+#                   Hand-localized pages (l10n: transcreate) get a ::warning::
+#                   instead of an error when they drift: the pipeline never
+#                   regenerates them, so drift is a report for a human, not a
+#                   build failure.
+#   update-digests [file...]
+#                   Recompute and rewrite `source_digest` in translated pages
+#                   from their English source. Run after editing an English
+#                   source or after adding/refreshing a translation. Without
+#                   arguments, hand-localized pages are SKIPPED (a wholesale
+#                   re-stamp would silence their drift report); pass the file
+#                   explicitly to re-stamp one you refreshed by hand.
 #
 # Guards run by `check`:
 #   1. i18n key parity  — every i18n/<lang>.toml must define exactly the same
@@ -67,6 +75,14 @@ translated_digest_files() {
     | sort
 }
 
+# A page a human localized by hand (l10n: transcreate). The pipeline never
+# regenerates these, so a stale digest here is a drift REPORT, not a build
+# failure — and re-stamping it wholesale would silence the report while the
+# drift persists.
+is_transcreate() {
+  grep -m1 -E '^l10n:' "$1" 2>/dev/null | grep -qE '^l10n:[[:space:]]*"?transcreate"?[[:space:]]*$'
+}
+
 # Map content/<lang>/<rel> -> content/en/<rel>
 en_source_for() {
   local f="$1"
@@ -129,28 +145,55 @@ check_digest_freshness() {
     actual="$(grep -m1 -E '^source_digest:' "$f" | sed -E 's/.*sha256:([0-9a-fA-F]+).*/\1/')"
     checked=$((checked + 1))
     if [ "$actual" != "$expected" ]; then
-      rc=1
-      echo "::error::stale translation: $f"
-      echo "    English source: $src"
-      echo "    recorded digest: sha256:${actual}"
-      echo "    current  digest: sha256:${expected}"
-      echo "    fix: refresh the translation, then run 'hack/check-i18n.sh update-digests'"
+      if is_transcreate "$f"; then
+        # Advisory by design: a drifted transcreation is refreshed when a
+        # human decides to, and must not block unrelated PRs meanwhile.
+        echo "::warning::hand-localized page drifted from its English source: $f"
+        echo "    refresh the transcreation by hand, then re-stamp it with:"
+        echo "    hack/check-i18n.sh update-digests $f"
+      else
+        rc=1
+        echo "::error::stale translation: $f"
+        echo "    English source: $src"
+        echo "    recorded digest: sha256:${actual}"
+        echo "    current  digest: sha256:${expected}"
+        echo "    fix: refresh the translation, then run 'hack/check-i18n.sh update-digests'"
+      fi
     fi
   done < <(translated_digest_files)
   [ "$rc" -eq 0 ] && echo "translation freshness: OK ($checked translated pages match their English source)"
   return "$rc"
 }
 
 update_digests() {
+  # With explicit files, re-stamp exactly those — the escape hatch for a
+  # transcreation a human just refreshed. With no arguments, re-stamp every
+  # machine-translated page but SKIP hand-localized ones: their stale digest
+  # is the only signal that the transcreation drifted, and a wholesale
+  # re-stamp would silence it while the drift persists.
   local f
-  while IFS= read -r f; do
+  {
+    # `|| true`: with zero stamped translations the underlying grep exits 1,
+    # which under `set -euo pipefail` would silently kill the whole script.
+    if [ "$#" -gt 0 ]; then printf '%s\n' "$@"; else translated_digest_files || true; fi
+  } | while IFS= read -r f; do
     [ -z "$f" ] && continue
+    if [ "$#" -eq 0 ] && is_transcreate "$f"; then
+      echo "skip $f — hand-localized (l10n: transcreate); refresh it by hand, then: hack/check-i18n.sh update-digests $f"
+      continue
+    fi
     local src expected tmp
     src="$(en_source_for "$f")"
     if [ ! -f "$src" ]; then
       echo "::warning::skip $f — English source missing: $src"
       continue
     fi
+    # The awk below only REWRITES an existing line; without this check a page
+    # missing the field would pass through untouched yet be reported "updated".
+    if ! grep -qE '^source_digest:' "$f"; then
+      echo "::warning::skip $f — no source_digest line in the front matter; add one, then re-run"
+      continue
+    fi
     expected="$(sha256 "$src")"
     tmp="$(mktemp)"
     awk -v d="sha256:${expected}" '
@@ -159,7 +202,7 @@ update_digests() {
     ' "$f" > "$tmp"
     mv "$tmp" "$f"
     echo "updated $f -> sha256:${expected}"
-  done < <(translated_digest_files)
+  done
 }
 
 cmd="${1:-check}"
@@ -171,10 +214,11 @@ case "$cmd" in
     exit "$rc"
     ;;
   update-digests)
-    update_digests
+    shift
+    update_digests "$@"
     ;;
   *)
-    echo "usage: $0 [check|update-digests]" >&2
+    echo "usage: $0 [check|update-digests [file...]]" >&2
     exit 2
     ;;
 esac

hack/i18n/README.md

@@ -20,6 +20,10 @@ content/en/**  ──▶ worklist.py        (missing | stale, via source_digest)
                 back-translate ─▶ compare vs original   (meaning-drift check)
                        │
                        ▼
+     deterministic checks (version/flag/do-not-translate integrity,
+                per-language typography — lib.py, no model calls)
+                       │
+                       ▼
         review ×2 (native speakers of the target language)
           • technical editor    → fluency / register / terminology
           • Cozystack maintainer → technical correctness vs source
@@ -163,7 +167,8 @@ succeeds. A page that reproducibly fails is a signal to translate it by hand.
 | `prompts/review-maintainer.md` | native Cozystack-maintainer reviewer (technical correctness) |
 | `prompts/revise.md` | revise a translation against reviewer findings |
 | `style-guides/<lang>.md` | per-language tone/register conventions |
-| `lib.py` | shared helpers (config, discovery, digest, front matter, protect/restore) |
+| `lib.py` | shared helpers (config, discovery, digest, front matter, protect/restore, integrity + typography checks) |
+| `lint_translations.py` | typography lint for already-published translations (advisory; `--strict` to gate) |
 | `worklist.py` | diff detector |
 | `translate.py` | translate + review-gate driver; also removes orphaned translations |
 | `run-daily.sh` | daily runner (subscription, until limit, commit + PR) |
@@ -183,9 +188,15 @@ succeeds. A page that reproducibly fails is a signal to translate it by hand.
 | Field | Meaning | Read by |
 |-------|---------|---------|
 | `source_digest` | sha256 of the English source the translation was made from — the freshness signal | `hack/check-i18n.sh`, `worklist.py` |
-| `l10n` | *how* the page was localized: `mt` (machine) or `transcreate` | humans triaging review |
+| `l10n` | *how* the page was localized: `mt` (machine) or `transcreate` | `worklist.py`, `translate.py`, `hack/check-i18n.sh`, plus humans triaging review |
 | `translation_review` | *ratification state*: `auto-reviewed`, `auto-reviewed-with-findings`, or `ratified` | `translation-banner.html` |
 
+`l10n: transcreate` is load-bearing for the pipeline, not just a review hint:
+mark a page `transcreate` → `worklist.py` skips it and `translate.py` refuses to
+overwrite it → `check-i18n.sh` downgrades its source drift to a warning instead
+of failing CI → you refresh the wording by hand → run `update-digests <file>` to
+re-stamp `source_digest` so the warning clears.
+
 `translation_status: current` appears on pages from the i18n PoC. It is not
 written or read by anything — `source_digest` computes freshness properly — so
 this pipeline does not write it.