test(hooks): add CLAUDE.md filename-reference regression coverage

Issue #608 reports the commit-msg validator falsely rejected legitimate
file references like "update CLAUDE.md routing index". The active
three-pattern attribution regex (introduced for #480) actually already
allows these — the original substring "CLAUDE" check has been gone for
a while — but the test suite never asserted the filename-allow surface
explicitly, so the behavior was undocumented and at risk of regression.

This commit closes that gap by adding two test groups to both layered
test files (PreToolUse commit-message-guard and the git commit-msg hook):

  1. Filename references allowed (issue #608)
     - "docs: update CLAUDE.md routing index"
     - "fix: correct CLAUDE.local.md gitignore path"
     - "refactor: move CLAUDE config loader" (narrative inline)
     - "chore: bump CLAUDE.md version field" (mid-subject)
     - "docs(claude): clarify routing in CLAUDE.md" (scope + filename)

  2. Known reject forms still blocked (issue #608)
     - "feat: generated with Claude" (prose pattern)
     - "fix: created by Anthropic team" (prose pattern)
     - "docs: written using Claude assistance" (prose pattern)

The validator library hooks/lib/validate-commit-message.sh is unchanged;
no behavior shift, only freshly-asserted regression coverage.

global/commit-settings.md gains a one-line clarification stating that
filename and narrative references to the project root config are
allowed, with a pointer to the three-pattern design comment in the
validator library for the precise rules.

Closes #608

Author: kcenon

global/commit-settings.md

@@ -3,4 +3,6 @@
 No AI/Claude attribution in commits, issues, or PRs.
 Enforced by `settings.json` (`attribution: ""`), the `commit-message-guard` PreToolUse hook (Claude-side feedback loop), and the `commit-msg` git hook installed by `hooks/install-hooks.sh` (terminal-side gate).
 
+Filename references to project root config files (`CLAUDE.md`, `CLAUDE.local.md`) and narrative mentions of the config (e.g. "the CLAUDE config loader") are allowed in commit subjects — only attribution patterns (`Co-Authored-By:` trailers, bot emoji adjacent to Claude/Anthropic, "generated/created/authored {with|by|using} {Claude|Anthropic}" prose) are rejected. See the three-pattern design comment in `hooks/lib/validate-commit-message.sh` for the exact rules.
+
 All GitHub Issues and Pull Requests must be written in English.

tests/hooks/test-commit-message-guard.sh

@@ -88,6 +88,23 @@ assert_allow '{"tool_input":{"command":"git commit -m \"feat: add claude integra
 assert_allow '{"tool_input":{"command":"git commit -m \"fix: anthropic API fallback\""}}' "casual 'anthropic API' → allow"
 assert_allow '{"tool_input":{"command":"git commit -m \"docs: clarify Claude API behavior\""}}' "casual 'Claude API behavior' → allow"
 
+echo ""
+echo "[AI attribution — config-filename references allowed (issue #608)]"
+# Both enforcement layers (PreToolUse guard and git commit-msg hook) source
+# the same library, so the filename-allow surface must be covered here too.
+# These cases exercise the -m extraction path of the PreToolUse guard.
+assert_allow '{"tool_input":{"command":"git commit -m \"docs: update CLAUDE.md routing index\""}}' "filename ref: CLAUDE.md → allow"
+assert_allow '{"tool_input":{"command":"git commit -m \"fix: correct CLAUDE.local.md gitignore path\""}}' "filename ref: CLAUDE.local.md → allow"
+assert_allow '{"tool_input":{"command":"git commit -m \"refactor: move CLAUDE config loader\""}}' "narrative ref: 'CLAUDE config' → allow"
+
+echo ""
+echo "[AI attribution — known reject forms still blocked (issue #608)]"
+# Regression guard: the false-positive surface above must not loosen the
+# reject surface. One representative per attribution prose pattern.
+assert_deny '{"tool_input":{"command":"git commit -m \"feat: generated with Claude\""}}' "prose: 'generated with Claude' → deny"
+assert_deny '{"tool_input":{"command":"git commit -m \"fix: created by Anthropic team\""}}' "prose: 'created by Anthropic' → deny"
+assert_deny '{"tool_input":{"command":"git commit -m \"docs: written using Claude assistance\""}}' "prose: 'written using Claude' → deny"
+
 echo ""
 echo "[emoji detection]"
 EMOJI_PARTY=$(printf '\xf0\x9f\x8e\x89')

tests/hooks/test-commit-msg.sh

@@ -103,6 +103,26 @@ assert_accept "feat: add claude integration" "casual 'claude integration'"
 assert_accept "fix: anthropic API fallback" "casual 'anthropic API'"
 assert_accept "docs: clarify Claude API behavior" "casual 'Claude API behavior'"
 
+echo ""
+echo "[AI attribution — config-filename references allowed (issue #608)]"
+# Project root config filenames are legitimate file references, not author
+# attribution. The validator must allow them in subjects so commit messages
+# that genuinely modify these files can name them precisely.
+assert_accept "docs: update CLAUDE.md routing index" "filename ref: CLAUDE.md in subject"
+assert_accept "fix: correct CLAUDE.local.md gitignore path" "filename ref: CLAUDE.local.md in subject"
+assert_accept "refactor: move CLAUDE config loader" "narrative ref: 'CLAUDE config' inline"
+assert_accept "chore: bump CLAUDE.md version field" "filename ref: CLAUDE.md mid-subject"
+assert_accept "docs(claude): clarify routing in CLAUDE.md" "scope + filename in subject"
+
+echo ""
+echo "[AI attribution — known reject forms still blocked (issue #608)]"
+# Regression guard: tightening the false-positive surface above must not
+# loosen the deliberate reject surface. Each of these is a representative
+# from the three attribution patterns (trailer / emoji / prose).
+assert_reject "feat: generated with Claude" "prose: 'generated with Claude'"
+assert_reject "fix: created by Anthropic team" "prose: 'created by Anthropic'"
+assert_reject "docs: written using Claude assistance" "prose: 'written using Claude'"
+
 echo ""
 echo "[emoji detection]"
 EMOJI_PARTY=$(printf '\xf0\x9f\x8e\x89')