🔧 fix: 検出器のファイル集合を repo-tracked に一致させる (#30)

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-kit",
-  "version": "0.4.5",
+  "version": "0.4.6",
   "description": "Shareable Claude Code agent kit: orchestration, batch-dispatch, review, ADR-authoring, and memory/skill-hygiene skills plus critic/implementer/code-reviewer agents. PR-workflow hooks ship separately as claude-kit-hooks.",
   "author": {
     "name": "Tomohito YABU",

.gitignore

@@ -4,6 +4,8 @@ memory/
 settings.json
 settings.local.json
 settings*.json
+# parked worktrees — otherwise `git ls-files --others` enumerates them as duplicate trees
+.claude/worktrees/
 # OS/editor noise
 .DS_Store
 __pycache__/

docs/claim-verification.md

@@ -76,6 +76,24 @@ lowercase filename back in, the grep fires and the next editor picks a dispositi
 mechanism working, not a regression. **Any doc quoting that example inherits the same
 constraint**, this one included.
 
+It still shipped with a blind spot, and the negative control is what missed it: the first form
+recursed with `rg`, which skips hidden directories, so `.claude/**` — rules, skills, agents — went
+unscanned, and the control sat at the repo root, inside the guard's existing reach, reddening
+without testing the question. **Scope a control to the claim's habitat, not just its pattern.**
+
+The deeper error was the *file set*: a recursive grep answers "which files lie here", the rule asks
+"which are repo-tracked", and the two diverge both ways — a tracked file under an ignored directory
+is missed, never-committed scratch is falsely flagged. `--hidden` closes that instance and leaves
+the class open; the fix was to enumerate with `git ls-files --cached --others --exclude-standard`,
+as `scripts/scrub-check.sh` already did — two mechanisms answering the same question differently is
+drift waiting to happen. **When a detector's file set is not the claim's file set, no flag fixes it.**
+
+It surfaced only when the detector ran against a repo it was not written in — the cheapest general
+form of this check, since the authoring repo is where a self-referential detector is least able to
+fail. The hit there was that repo's copy of the prose example, still spelling a concrete filename:
+the constraint above reaches consumer mirrors, which must import the placeholder rewrite along with
+the detector, or report their own rule file as a violation. Sequence and measurements: #30, #32.
+
 One shape has no in-session probe at all: a rules file created mid-session never injects in that
 session, so verify it from fresh subagent probes against a positive control. Mechanism, scope limits
 and re-runnable probes: `~/.claude/kit-docs/code-review-path-scoped-rules.md`.

rules/knowledge-layering.md

@@ -61,10 +61,12 @@ shared repo needs to carry itself.
 (`#N`, `ADR-NNN`). Memory refs are fine only in never-committed places (`~/.claude/CLAUDE.md`,
 this file, conversational scratch).
 
-**Detect** — new code must not *add* hits (the *reframe* disposition below, not a "must return 0"):
+**Detect** — new code must not *add* hits (the *reframe* disposition below, not a "must return 0").
+A recursive grep silently skips dot-dirs and tracked-but-ignored files — enumerate, don't recurse:
 
 ```sh
-rg -n 'memory `[a-z_]+\.md`' --glob '!**/memory/**' --glob '!.git/**'
+git ls-files -z --cached --others --exclude-standard \
+  | xargs -0 grep -nHE 'memory `[a-z_]+\.md`'
 ```
 
 ## Verify before you lock it