Carryover seed recovery — PR template · blog trigger · i18n residue gate (seeds #24/#23/#171/#26/#27/#30/#31)

Date
ImpactHigh

Goal

  • After recovering the security-sensitive single seed (#new7) in Sprint 174, recover the remaining inherited carryover seeds as independent per-seed PRs.
  • Three priorities: ① plan/PR template hardening (#24/#23/#171), ② ADR/blog CI hardening (#26/#27/#28), ③ i18n/lint (#30/#31).
  • Regression isolation: independent PR per seed + Critic (Codex) cross-review + Squash merge.

Decisions

D0. #28 is already implemented → excluded from work

The ROOT auto-detection in check-adr-links.mjs is already implemented via scripts/check-adr-links.mjs:33 (resolve(import.meta.dirname, '..')) + deriveOutRoot(), so no additional work is required. The ADR/blog hardening is reduced to the two items #26/#27.

D1. #26 — positive glob narrowing instead of a negation pattern

Verifying the dorny/paths-filter source (src/filter.ts) directly shows that each pattern is compiled into its own picomatch matcher and then combined with the default predicate-quantifier: some (OR). Therefore, adding a !docs/adr/README.md negation pattern to the blog filter would be neutralised because docs/adr/** matches first under OR (negation only works under the every quantifier, but every would break every filter in the single dorny step). Instead, narrow the filter to positive globs that match exactly what the blog SSG actually consumes.

  • Consumers: blog/src/lib/adr/loader.ts (DIR_KIND_MAP = [sprints, topics]) + blog/scripts/generate-search-index.mjs → read only root ADR-*.md + sprints/ + topics/, with README.md explicitly excluded (loader.ts:62, generate-search-index.mjs:86).
  • 'docs/adr/**' → 'docs/adr/ADR-*.md' + 'docs/adr/sprints/**' + 'docs/adr/topics/**'.
  • A picomatch simulation confirmed docs/adr/README.md skip / real ADR content trigger.

D2. #27 — out/ fail-fast guard

If next build (output:export) fails silently, an empty out/ is copied verbatim into Docker and an empty site could be deployed. Add a step that verifies out/ exists and is non-empty before the link check.

D3. #30/#31 — gate the translation source instead of the build artifact (shift-left)

A direct Hangul grep over the build artifact (out/en/**) would falsely flag legitimate KR-fallback banner pages (hasEnTranslation=false). Instead, gate the prose Hangul density of the translation source (docs/adr-en/**/*.md). Because the blog /en ADR body is rendered verbatim from docs/adr-en (loader.ts readLocalized), gating the source blocks build-artifact Korean residue shift-left while avoiding the banner false-positive.

  • i18n 3 layers: layer 1 KR SSOT exists → layer 2 EN file exists (check-adr-en-coverage --strict, existing) → layer 3 EN actually translated (check-i18n-residue --strict, new).
  • Code fences/inline code excluded (Korean log/commit examples are legitimate) + violation only when the ratio threshold (default 8%) AND the absolute floor (10 chars) are both met. The current corpus max prose Hangul is 2.19%, leaving 3.6x headroom.

Implementation (independent PR per seed)

PR #300 9b8f964 — PR template hardening (#24/#23/#171)

docs(ci): PR 템플릿에 문서 양면/소비처/누적카운트 체크리스트 추가

  • Added a「Docs/ADR change」section (KR+EN dual update #24, README cumulative count after rebase #23) and a「New artifact — adoption ≠ consumption」section (#171) to .github/pull_request_template.md, following the existing section style (> rationale blockquote).

PR #301 43fe210 — blog trigger + out/ verification (#26/#27)

ci(blog): blog 트리거 소비처 일치 + SSG out/ 산출물 검증

  • Narrowed the detect-changes blog filter in ci.yml to consumer-aligned positive globs (README excluded).
  • Added a Verify SSG output (out/) fail-fast step to the build-blog job.

PR #302 966fa56 — i18n residue gate (#30/#31)

feat(ci): EN ADR 한국어 잔재 게이트 (i18n 계층 3)

  • New scripts/check-i18n-residue.mjs — follows the en-coverage structure (ROOT auto-detect/--strict/exit 0·1·2/entry guard/exports), reuses collectAdrFiles (DRY). Modes: stats/--lint/--strict/--max-ratio=N.
  • Added a residue --strict step to the ci.yml quality-docs job + registered the script in the docs paths-filter.
  • Documented the i18n 3-layer model in docs/adr-en/README.md.

Critic cycle

  • PR1 (codex review --base main, session 019e442b-eaa7-7483-bbcc-9d61f8097889): 0 findings — "checklist guidance only, no runtime impact".
  • PR2 (session 019e4432-db17-7d62-a5a6-5bccdd71e8fa): 0 findings — "path-filter narrowing + SSG output check consistent, existing workflows unbroken".
  • PR3 (session 019e443b-ef00-7f81-b493-b6edd046ce0e): 0 findings — "new CI check/script works, integrates with the docs quality gate".

All three PRs passed in a single review round with zero P0–P3 findings.

Verification

Local

  • check-i18n-residue --strict: 122 EN files PASS (max prose Hangul 2.19% < 8%).
  • Exported-function unit checks 9/9 (English / stub 94.1% / code-fence excluded / inline excluded / absolute floor / arg parsing), error path exit 2, import side-effect-free.
  • check-doc-refs (311 files, 0 broken) / check-adr-en-coverage --strict (122/122) / check-regex-robustness clean.
  • Validated the ci.yml outer YAML + dorny inner filters YAML parse, plus a picomatch matching simulation passing all cases.

CI

  • PR #300/#301/#302 each merged via Squash after CI green (mergeStateStatus CLEAN).

New UAT (Sprint 175)

  • Real user direct: group analysis partial-recovery rendering on truncated responses (inherited from Sprint 174); visual check for Korean residue on the English blog /en ADR pages.

Result

  • Merge: origin/main 3f938fe → 966fa56 (PR #300 9b8f964 / #301 43fe210 / #302 966fa56, all squash merge)
  • Net change: +~250 (PR template 15 lines, ci.yml 24 lines, check-i18n-residue.mjs 197 lines, README 17 lines)

New patterns

  • Consumer-aligned trigger: a CI path filter is aligned with the exact set of paths the consuming code actually reads. In an OR-quantifier environment where negation patterns are neutralised, use positive glob narrowing to remove unnecessary triggers.
  • i18n 3-layer gate: separate existence (en-coverage) from quality (i18n-residue) to close the "file exists but is an untranslated stub" gap. The source gate blocks build-artifact residue shift-left.