fix: 이식성 잔여 — python3, 개행 경계 순회, 캐시 resolver 정렬 2중 결함

.claude-plugin/marketplace.json

@@ -5,7 +5,7 @@
   },
   "metadata": {
     "description": "Personal Claude Code plugin collection with 24 specialized plugins. Codex 0.135 + Hermes Agent native — generated `.agents/plugins/marketplace.json` + per-plugin `.codex-plugin/plugin.json` (`scripts/sync-codex-manifests.mjs`) and Hermes `plugin.yaml` + `__init__.py` adapters (`scripts/sync-hermes-manifests.mjs`).",
-    "version": "2.7.2"
+    "version": "2.7.3"
   },
   "plugins": [
     {
@@ -61,7 +61,7 @@
       "name": "ml-toolkit",
       "source": "./plugins/ml-toolkit",
       "description": "ML/multimodal development principles, GPU parallel processing, Gradio CV apps, CV notebook generation, interactive CV data exploration",
-      "version": "1.4.3",
+      "version": "1.4.4",
       "category": "development"
     },
     {
@@ -124,7 +124,7 @@
       "name": "project-init",
       "source": "./plugins/project-init",
       "description": "Agent-harness project lifecycle. `new` — one-shot first-day bootstrap (interview, .claude/ scaffold, minimal CLAUDE.md, AGENTS.md with Codex GitHub cloud reviewer guidelines in general/ml/web variants, README/CHANGELOG seed, gh repo create + initial push) behind a hardened preflight guard that refuses a non-empty cwd. `wiring` — its inverse for existing repos: a read-only 14-axis setup diagnostic (FAIL / WARN / ASK / INFO / SKIP / OK) covering guidance files, .claude/rules paths: scoping defeated by @import, llm-wiki layout + staging backlog, Serena onboarding, memory posture, MCP duplicate registration, Codex CLI posture + AGENTS.md byte budget, spec homes, gws-sync, .tmp, core.hooksPath, .gitignore coverage. ASK answers persist to .claude/state/wiring.json. Shared read-only detector: scripts/project_state.sh.",
-      "version": "0.6.1",
+      "version": "0.6.2",
       "category": "planning"
     },
     {
@@ -152,7 +152,7 @@
       "name": "ppt-yeong-style",
       "source": "./plugins/ppt-yeong-style",
       "description": "yeong 스타일 강의·제안 덱 작성 규약 — ppt-master 빌드 엔진 위에 얹는 작성 레이어(엔진 자체가 아님). 스킬 3종: 메인 ppt-yeong-style(미감 시그니처 §0 'Editorial restraint, one committed accent'·md 소스 규약·작성 원칙 16종·밀도 리듬·역할 기반 색·codex-image vs SVG 경계·앱 UI 실물 강제·레버 조합 차별화·윤문·렌더 QA + references/ 6종 + 주입 페이로드) + lecture-deck(강의 덱 운영 — 실습 handouts 생성 규약·프롬프트 카드·placeholder→실캡처 스크린샷 슬롯·리넘버링 4중 동기화·전사 회고 루프·강사 노트 태그 + cc-common 47장 레퍼런스) + deck-review(관점별 리뷰 서브에이전트 4종 audience-fit·story-flow·fact-check·design-qa 병렬 오케스트레이션 + codex:rescue 교차 리뷰, 페르소나는 파라미터). 의존 스킬은 있으면 사용, 없으면 생략 + 설치 제안 문구(ppt-master만 prerequisite-stop). ppt-master로 그냥 'PPT 만들기'와 달리 yeong 규약이 필요할 때.",
-      "version": "0.9.3",
+      "version": "0.9.4",
       "category": "design"
     },
     {

plugins/ml-toolkit/.claude-plugin/plugin.json

@@ -1,5 +1,5 @@
 {
   "name": "ml-toolkit",
-  "version": "1.4.3",
+  "version": "1.4.4",
   "description": "ML/multimodal development principles, GPU parallel processing, Gradio CV apps, CV notebook generation, interactive CV data exploration"
 }

plugins/ml-toolkit/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ml-toolkit",
-  "version": "1.4.3",
+  "version": "1.4.4",
   "description": "ML/multimodal development principles, GPU parallel processing, Gradio CV apps, CV notebook generation, interactive CV data exploration",
   "author": {
     "name": "YoungjaeDev"

plugins/ml-toolkit/plugin.yaml

@@ -1,5 +1,5 @@
 name: ml-toolkit
-version: "1.4.3"
+version: "1.4.4"
 description: "ML/multimodal development principles, GPU parallel processing, Gradio CV apps, CV notebook generation, interactive CV data exploration"
 author: "YoungjaeDev"
 kind: standalone

plugins/ml-toolkit/skills/cv-notebook/references/environment-setup.md

@@ -314,8 +314,11 @@ HF_HOME=/path/to/huggingface/cache
 
 #### Using venv
 ```bash
-# Create virtual environment
-python -m venv .venv
+# Create virtual environment (runs before activation, so it resolves against the
+# system PATH: macOS 12.3+ has no `python`, and a python.org Windows install has
+# no `python3` -- only `python` and the `py` launcher)
+python3 -m venv .venv        # Linux/Mac
+# py -3 -m venv .venv        # Windows
 
 # Activate
 source .venv/bin/activate  # Linux/Mac

plugins/ml-toolkit/skills/gpu-parallel-pipeline/SKILL.md

@@ -127,7 +127,7 @@ else
   SKILL_DIR="$HOME/.hermes/plugins/ml-toolkit/skills/gpu-parallel-pipeline"    # Hermes default install
 fi
 [ -d "$SKILL_DIR" ] || { echo "gpu-parallel-pipeline: skill dir not resolved" >&2; exit 1; }
-python "$SKILL_DIR/scripts/check_gpu_memory.py"
+python3 "$SKILL_DIR/scripts/check_gpu_memory.py"
 ```
 
 **Rule of thumb:**

plugins/ml-toolkit/skills/gpu-parallel-pipeline/scripts/check_gpu_memory.py

@@ -1,11 +1,11 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 """GPU memory check utility for parallel pipeline planning.
 
 Reports available GPU memory and recommends workers per GPU based on model size.
 
 Usage:
-    python check_gpu_memory.py
-    python check_gpu_memory.py --model-memory 5.0  # Specify model memory in GB
+    python3 check_gpu_memory.py
+    python3 check_gpu_memory.py --model-memory 5.0  # Specify model memory in GB
 """
 
 from __future__ import annotations

plugins/ppt-yeong-style/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ppt-yeong-style",
-  "version": "0.9.3",
+  "version": "0.9.4",
   "description": "yeong 스타일 강의·제안 덱 작성 규약 — ppt-master 빌드 엔진 위에 얹는 작성 레이어(엔진 자체가 아님). 스킬 3종: 메인 ppt-yeong-style(미감 시그니처 §0 'Editorial restraint, one committed accent'·md 소스 규약·작성 원칙 16종·밀도 리듬·역할 기반 색·codex-image vs SVG 경계·앱 UI 실물 강제·레버 조합 차별화·윤문·렌더 QA + references/ 6종 + 주입 페이로드) + lecture-deck(강의 덱 운영 — 실습 handouts 생성 규약·프롬프트 카드·placeholder→실캡처 스크린샷 슬롯·리넘버링 4중 동기화·전사 회고 루프·강사 노트 태그 + cc-common 47장 레퍼런스) + deck-review(관점별 리뷰 서브에이전트 4종 audience-fit·story-flow·fact-check·design-qa 병렬 오케스트레이션 + codex:rescue 교차 리뷰, 페르소나는 파라미터). 의존 스킬은 있으면 사용, 없으면 생략 + 설치 제안 문구(ppt-master만 prerequisite-stop). ppt-master로 그냥 'PPT 만들기'와 달리 yeong 규약이 필요할 때.",
   "skills": [
     "./skills/ppt-yeong-style",

plugins/ppt-yeong-style/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ppt-yeong-style",
-  "version": "0.9.3",
+  "version": "0.9.4",
   "description": "yeong 스타일 강의·제안 덱 작성 규약 — ppt-master 빌드 엔진 위에 얹는 작성 레이어(엔진 자체가 아님). 스킬 3종: 메인 ppt-yeong-style(미감 시그니처 §0 'Editorial restraint, one committed accent'·md 소스 규약·작성 원칙 16종·밀도 리듬·역할 기반 색·codex-image vs SVG 경계·앱 UI 실물 강제·레버 조합 차별화·윤문·렌더 QA + references/ 6종 + 주입 페이로드) + lecture-deck(강의 덱 운영 — 실습 handouts 생성 규약·프롬프트 카드·placeholder→실캡처 스크린샷 슬롯·리넘버링 4중 동기화·전사 회고 루프·강사 노트 태그 + cc-common 47장 레퍼런스) + deck-review(관점별 리뷰 서브에이전트 4종 audience-fit·story-flow·fact-check·design-qa 병렬 오케스트레이션 + codex:rescue 교차 리뷰, 페르소나는 파라미터). 의존 스킬은 있으면 사용, 없으면 생략 + 설치 제안 문구(ppt-master만 prerequisite-stop). ppt-master로 그냥 'PPT 만들기'와 달리 yeong 규약이 필요할 때.",
   "author": {
     "name": "YoungjaeDev"

plugins/ppt-yeong-style/plugin.yaml

@@ -1,5 +1,5 @@
 name: ppt-yeong-style
-version: "0.9.3"
+version: "0.9.4"
 description: "yeong 스타일 강의·제안 덱 작성 규약 — ppt-master 빌드 엔진 위에 얹는 작성 레이어(엔진 자체가 아님). 스킬 3종: 메인 ppt-yeong-style(미감 시그니처 §0 'Editorial restraint, one committed accent'·md 소스 규약·작성 원칙 16종·밀도 리듬·역할 기반 색·codex-image vs SVG 경계·앱 UI 실물 강제·레버 조합 차별화·윤문·렌더 QA + references/ 6종 + 주입 페이로드) + lecture-deck(강의 덱 운영 — 실습 handouts 생성 규약·프롬프트 카드·placeholder→실캡처 스크린샷 슬롯·리넘버링 4중 동기화·전사 회고 루프·강사 노트 태그 + cc-common 47장 레퍼런스) + deck-review(관점별 리뷰 서브에이전트 4종 audience-fit·story-flow·fact-check·design-qa 병렬 오케스트레이션 + codex:rescue 교차 리뷰, 페르소나는 파라미터). 의존 스킬은 있으면 사용, 없으면 생략 + 설치 제안 문구(ppt-master만 prerequisite-stop). ppt-master로 그냥 'PPT 만들기'와 달리 yeong 규약이 필요할 때."
 author: "YoungjaeDev"
 kind: standalone

plugins/ppt-yeong-style/skills/lecture-deck/scripts/render-qa.sh

@@ -43,7 +43,14 @@ b_hits=$(grep -REl '>[^<]*\[[^]]*\][^<]*<' "$SVG_DIR" 2>/dev/null || true)
 if [ -n "$b_hits" ]; then
   fail=1
   echo "  FAIL — bracket placeholder(s) leaked into rendered text:"
-  for f in $b_hits; do
+  # Iterate on newline boundaries: `for f in $b_hits` word-splits on IFS and
+  # glob-expands, so an absolute DECK_ROOT containing a space arrives here as two
+  # fragments, the grep below fails on both, and its stderr is discarded -- the
+  # FAIL detail degrades to half-paths with (0) counts. The gate itself is
+  # unaffected (fail=1 is already set above); this is the report, which is what
+  # the user reads to find the leak.
+  printf '%s\n' "$b_hits" | while IFS= read -r f; do
+    [ -n "$f" ] || continue
     c=$(grep -Eo '>[^<]*\[[^]]*\][^<]*<' "$f" 2>/dev/null | wc -l | tr -d ' ')
     echo "    $f ($c)"
   done

plugins/project-init/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "project-init",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "description": "Agent-harness project lifecycle. `new` — one-shot first-day bootstrap (interview, .claude/ scaffold, minimal CLAUDE.md, AGENTS.md with Codex GitHub cloud reviewer guidelines in general/ml/web variants, README/CHANGELOG seed, gh repo create + initial push) behind a hardened preflight guard that refuses a non-empty cwd. `wiring` — its inverse for existing repos: a read-only 14-axis setup diagnostic (FAIL / WARN / ASK / INFO / SKIP / OK) covering guidance files, .claude/rules paths: scoping defeated by @import, llm-wiki layout + staging backlog, Serena onboarding, memory posture, MCP duplicate registration, Codex CLI posture + AGENTS.md byte budget, spec homes, gws-sync, .tmp, core.hooksPath, .gitignore coverage. ASK answers persist to .claude/state/wiring.json. Shared read-only detector: scripts/project_state.sh.",
   "author": {
     "name": "YoungjaeDev"

plugins/project-init/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "project-init",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "description": "Agent-harness project lifecycle. `new` — one-shot first-day bootstrap (interview, .claude/ scaffold, minimal CLAUDE.md, AGENTS.md with Codex GitHub cloud reviewer guidelines in general/ml/web variants, README/CHANGELOG seed, gh repo create + initial push) behind a hardened preflight guard that refuses a non-empty cwd. `wiring` — its inverse for existing repos: a read-only 14-axis setup diagnostic (FAIL / WARN / ASK / INFO / SKIP / OK) covering guidance files, .claude/rules paths: scoping defeated by @import, llm-wiki layout + staging backlog, Serena onboarding, memory posture, MCP duplicate registration, Codex CLI posture + AGENTS.md byte budget, spec homes, gws-sync, .tmp, core.hooksPath, .gitignore coverage. ASK answers persist to .claude/state/wiring.json. Shared read-only detector: scripts/project_state.sh.",
   "author": {
     "name": "YoungjaeDev"

plugins/project-init/references/gh-repo-create-flow.md

@@ -66,8 +66,20 @@ The `--license <name>` flag of `gh repo create` can auto-generate a LICENSE file
 
 ```bash
 # Auto-seed the license
+# ${VISIBILITY,,} is a Bash 4 expansion; macOS /bin/bash is 3.2 and fails with
+# "bad substitution". Normalize with the POSIX tr helper the executable
+# procedure already uses (new-procedure.md Phase 4/6), and match on a token
+# rather than passing the raw answer straight into a CLI flag.
+to_lower() { printf '%s' "$1" | tr '[:upper:]' '[:lower:]'; }
+case "$(to_lower "$VISIBILITY")" in
+  *private*)  VIS_FLAG="private" ;;
+  *public*)   VIS_FLAG="public" ;;
+  *internal*) VIS_FLAG="internal" ;;
+  *) echo "unknown visibility: $VISIBILITY" >&2; exit 1 ;;
+esac
+
 gh repo create "${OWNER}/${PROJECT_NAME}" \
-  --${VISIBILITY,,} \
+  --${VIS_FLAG} \
   --description "${ONE_LINER}" \
   --license "${LICENSE}" \
   --source=. --remote=origin --push
@@ -88,8 +100,20 @@ git add .claude/ CLAUDE.md AGENTS.md README.md CHANGELOG.md
 git commit -m "chore: bootstrap project skeleton via project-init"
 
 # 3. gh repo create + auto push
+# ${VISIBILITY,,} is a Bash 4 expansion; macOS /bin/bash is 3.2 and fails with
+# "bad substitution". Normalize with the POSIX tr helper the executable
+# procedure already uses (new-procedure.md Phase 4/6), and match on a token
+# rather than passing the raw answer straight into a CLI flag.
+to_lower() { printf '%s' "$1" | tr '[:upper:]' '[:lower:]'; }
+case "$(to_lower "$VISIBILITY")" in
+  *private*)  VIS_FLAG="private" ;;
+  *public*)   VIS_FLAG="public" ;;
+  *internal*) VIS_FLAG="internal" ;;
+  *) echo "unknown visibility: $VISIBILITY" >&2; exit 1 ;;
+esac
+
 gh repo create "${OWNER}/${PROJECT_NAME}" \
-  --${VISIBILITY,,} \
+  --${VIS_FLAG} \
   --description "${ONE_LINER}" \
   --source=. --remote=origin --push
 ```

plugins/project-init/references/new-procedure.md

@@ -30,14 +30,19 @@ if [ -z "$PLUGIN_ROOT" ]; then
   # Best-effort Codex cache lookup. Order: latest version directory under the
   # configured marketplace name; user override via $CODEX_PLUGIN_CACHE.
   cache_root="${CODEX_PLUGIN_CACHE:-$HOME/.codex/plugins/cache}"
-  # Prefer `sort -V` (GNU) for proper semver ordering; macOS / BSD sort lacks
-  # -V and would fail silently, so fall back to plain lexicographic sort.
-  # Picks fine for typical X.Y.Z under 10 — covers the realistic version range
-  # the plugin cache will ever hold for a single plugin.
+  # Sort on the version basename, not the full path: with two marketplace dirs a
+  # name like zeta/project-init/0.4.0 would otherwise outrank alpha/project-init/0.6.0.
+  # `sort -V` orders X.Y.Z properly and is present on GNU and on Apple's FreeBSD
+  # sort (10.13+), but the probe stays for userlands that predate it; the fallback
+  # is a numeric dotted-field sort, because plain lexicographic ranks 0.6.0 above
+  # 0.10.0 and would resolve to an older cached version. Same form as
+  # plugins/github-dev/skills/cr-fix/SKILL.md, whose regression test guards it.
   if sort -V </dev/null >/dev/null 2>&1; then
-    candidate=$(ls -1d "$cache_root"/*/project-init/* 2>/dev/null | sort -V | tail -1)
+    candidate=$(ls -1d "$cache_root"/*/project-init/* 2>/dev/null \
+      | awk -F/ '{print $NF "\t" $0}' | sort -V | tail -1 | cut -f2-)
   else
-    candidate=$(ls -1d "$cache_root"/*/project-init/* 2>/dev/null | sort | tail -1)
+    candidate=$(ls -1d "$cache_root"/*/project-init/* 2>/dev/null \
+      | awk -F/ '{print $NF "\t" $0}' | sort -t. -k1,1n -k2,2n -k3,3n | tail -1 | cut -f2-)
   fi
   [ -n "$candidate" ] && [ -d "$candidate" ] && PLUGIN_ROOT="$candidate"
 fi
@@ -61,7 +66,11 @@ CWD=$(pwd)
 DIR_NAME=$(basename "$CWD")
 HAS_GIT=$([ -d .git ] && echo "yes" || echo "no")
 HAS_CLAUDE=$([ -d .claude ] && echo "yes" || echo "no")
-HAS_CODE=$(find . -maxdepth 2 -type f \( -name "*.py" -o -name "*.ts" -o -name "*.js" -o -name "*.go" -o -name "*.rs" -o -name "*.java" \) 2>/dev/null | head -1 | wc -l)
+# yes/no like its neighbours, not a count: BSD `wc -l` right-pads its output, so
+# the old `| head -1 | wc -l` form yielded "       1" here while HAS_GIT and
+# HAS_CLAUDE were clean tokens. `-print -quit` also stops at the first match
+# instead of walking the whole tree.
+HAS_CODE=$([ -n "$(find . -maxdepth 2 -type f \( -name "*.py" -o -name "*.ts" -o -name "*.js" -o -name "*.go" -o -name "*.rs" -o -name "*.java" \) -print -quit 2>/dev/null)" ] && echo "yes" || echo "no")
 
 # Collect GitHub owner candidates
 bash "$PLUGIN_ROOT/scripts/infer-github-context.sh"