[gregoryfoster]

CR round-1 findings from the #39/#41 review, all eight actioned.

The findings clustered on one theme: the session fixed a hook whose failure mode was silence, and three of the changes reintroduced silence somewhere new.

#	Finding	Action
1	socraticode-health.sh installed as a copy while both sibling hooks are symlinks — the #16 drift, invisible to doctor.sh (which only reports dangling links)	Re-linked into skills-vendor/; contradiction filed as skills#179
2	test_skills_hook.py pinned the symlink but not the registration — the half that failed in #39 was the half nothing guarded	Two tests added, both verified red against a stripped registration first
3	The skills#177 workaround lived only in gitignored settings.local.json	SOCRATICODE_DRIVER / SOCRATICODE_PROBE_FILE moved to committed .claude/settings.json
4	#177 note hand-added to docs/SOCRATICODE.md, six lines under its own "a re-run overwrites this wholesale" header	Moved to AGENTS.md → ## Code Exploration Notes (repo-specific)
5	Neither socraticode hook documented anywhere	docs/SKILLS.md now covers all three .claude/hooks/ entries
6	docs/SOCRATICODE.md hardcoded /home/exedev/replicator	Now . and "$SOCRATICODE_DRIVER"
7	Health hook measures . → a worktree session reports a healthy index as broken	Documented; filed as skills#180
8	Inventory table unchanged after a 153-commit bump	Noted that it lists what is linked, and that new skills are unevaluated

The two that were more than tidying

Finding 2 is the regression guard for #39 itself. The old test's docstring reasoned "Nothing else catches a re-copy, so it is pinned here" — that argument now applies more strongly to the registration than to the link, since the registration is the half that actually went missing for nine days. Proven red before green:

FAILED test_the_hook_is_registered_in_settings_json
FAILED test_the_installer_agrees_that_both_halves_are_present
  install-refresh.sh --check reported a half-installed hook:
    hook symlink:       .claude/hooks/skills-submodule-update.sh -> ...
    SessionStart entry: MISSING — the hook is on disk but Claude Code never runs it

Finding 3 meant the committed repo shipped the bug this one machine was patched against: on CI, a fresh clone, or any worktree, the health hook resolved the broken skills/ symlink, the driver exited 0 printing nothing, and the hook reported clean without measuring anything.

Budget

The write guard caught the first draft of the finding-4 relocation at 6,068 tokens, 68 over. Rather than ship over budget, the detail went to docs/SKILLS.md and only the operative rule stayed inline — AGENTS.md lands at 5,949.

Verification

• uv run pytest: 685 passed (2 new), 99.83% coverage
• uv run ruff check . and ruff format --check .: clean
• context surface: policy 5,949 / 6,000; every doc under 10,000; no dead links or anchors
• install-refresh.sh --check: exit 0, both artifacts

🤖 Generated with Claude Code