m2ui structural tests

Bash-based smoke tests that validate the skill's structural integrity. Run on every push (manually or via CI) to catch drift between releases — broken cross-references, missing required sections, atlas entries that lost their Quick check, etc.

These tests do NOT validate agent behavior (that's manual via fresh CC session). They validate the skill's text shape.

Running

From the plugin root:

bash tests/run-all.sh

Or run an individual test:

bash tests/test-mode-dispatch.sh
bash tests/test-required-refs.sh
bash tests/test-pre-emit-invariants.sh

Each test prints PASS or FAIL: <reason> per assertion and exits non-zero on any failure.

What each test covers

• test-mode-dispatch.sh — SKILL.md mode-detection numbered items, dispatch keywords, conditional-load matrix scope, EXTREMELY-IMPORTANT block presence, SUBAGENT-STOP guard.
• test-required-refs.sh — every skills/m2ui/reference/*.md cross-reference in SKILL.md, failure-atlas.md, screenshot.md, and m2ui-pre-emit-reviewer.md points to a file that exists on disk.
• test-pre-emit-invariants.sh — SKILL.md Pre-Emit Self-Review checklist has the expected number of items; every failure-atlas entry has a Quick check + a See-also cross-reference; mandatory anchors all present in reference/anchors/; mandatory ref files all present.

Adding a new test

Follow the same pattern as the three existing scripts:

1. Print test name + a separator at start.
2. For each assertion, print PASS or FAIL: <human description> on its own line.
3. Maintain a FAILURES counter; exit $FAILURES at the end.
4. Add a line invoking the new script to run-all.sh.
5. Document what the test covers in this README.

Keep assertions independent — one failed assertion should not prevent the next from running. The goal is a complete map of what's broken, not a stop-on-first-failure debugger.

Limitations

• These tests catch STRUCTURAL drift (missing sections, broken cross-refs, count mismatches). They do NOT catch SEMANTIC regressions (e.g., a rule that's worded incorrectly but still present). Behavioral validation requires running the skill against real prompts in a fresh Claude Code session.
• Personal-data scrub is NOT a test here — it's enforced at PR time via the git ls-files | xargs grep -l ... pattern in commit/PR plans. A test would risk false positives on the regex itself.