feat: tool logging (Option B), workflows track QA, and new advanced track

[3t-dev]

What changed

Four commits delivering the tool-validation architecture (Option B from docs/discussion/architecture-alternatives.md), a QA pass of the workflows track, and a brand-new advanced track. Every exercise validated end-to-end.

1. Option B — hook-based tool logging (3e0c6c4)

• hooks/hooks.json: plugin-bundled PostToolUse hook (matcher-scoped to Edit|Write|Read|Grep|Glob|Bash|Skill) — activates automatically when the plugin is enabled, no user settings edits
• tools/log-tool.sh: POSIX-only stdin-JSON logger; records tool usage only for calls touching ~/.cclab/workspace/, appending to the exercise's .tool_log.jsonl
• /cclab:reset now clears the tool log so stale tool usage can't pass process validation
• cc-009 "Master the Edit Tool": first exercise that validates the process (Edit used, not Write) in addition to the artifact, with graceful guidance when logging is inactive

2. Workflows track QA hardening (3b9a012)

A reviewer-agent pass of wf-001..wf-008 against June 2026 Claude Code docs found and fixed:

• Reset bug in 5 exercises: setup.sh didn't remove learner-created files, so reset silently left exercises solved
• Cheat paths: wf-002 passed with .env rules in allow (the opposite of the lesson); wf-004 frontmatter checks satisfiable from body text; wf-007 accepted single-underscore MCP tool names
• Content accuracy: Bash(npm *) word-boundary semantics, description as the auto-invocation trigger, Python 3 prereq for wf-007; verified argument-hint: and 0-indexed $0/$1 are correct per current docs (left unchanged)
• Execute bits set on all setup/validate scripts

3. New advanced track — adv-001..adv-006 (076b708)

Six exercises (+ docs/curriculum/CURRICULUM-mvp3.md) that go one level deeper than workflows without duplicating it:

ID	Title	Deepens
adv-001	Specialist Squad	subagent specialization: tools/disallowedTools/model/permissionMode
adv-002	Iron Gate	blocking PreToolUse hooks — validator executes the learner's hook
adv-003	Server Smith	authoring a stdio MCP server — validator speaks JSON-RPC to it
adv-004	Parallel Worlds	two parallel worktrees, merged back to main
adv-005	Forked Context	skills with named arguments + context: fork
adv-006	Plugin Architect	capstone: package skill + bundled hook into a plugin

/cclab:start banner/completion and CLAUDE.md updated for 3 tracks / 23 exercises.

4. Docs (1488d0f)

Architecture discussion documents moved into docs/discussion/.

Why

Artifact-only validation couldn't verify how a learner worked ("use the Edit tool", "use Grep"). Plugin-bundled hooks solve this with zero install friction — and unblock the whole class of process-validated exercises the new tracks rely on. The workflows QA and advanced track complete the curriculum through MVP3.

How to test

# hook logger unit behavior (gates on workspace path, ignores everything else)
bash tools/log-tool.sh < <payload.json with cwd/file_path inside ~/.cclab/workspace/cc-009>

# any exercise end-to-end (same pattern for all 23)
bash exercises/fundamentals/cc-009/setup.sh
bash exercises/fundamentals/cc-009/validate.sh   # exit 1 (fresh)
# ...apply the reference solution from hints.md level 3...
bash exercises/fundamentals/cc-009/validate.sh   # exit 0

# full matrix already run: fresh-fail → solution-pass → cheat-fail → post-reset-fail
# for all 9 fundamentals + 8 workflows + 6 advanced exercises

🤖 Generated with Claude Code