feat(ci): add test count guard to prevent AI agents from deleting tests

[diberry]

feat(ci): Add test count guard to prevent AI agents from deleting tests

Closes #39

Problem

AI coding agents sometimes delete or weaken tests to make failing code pass — the "green bar at any cost" anti-pattern. This is the same class of destructive shortcut that caused #631 on bradygaster/squad (361 files deleted by @copilot).

Solution

Three defense layers:

1. Test Count Guard (CI)

.github/scripts/test-count-guard.mjs — compares actual test count against baseline:

• Pass: actual >= baseline
• Fail: actual < baseline (prints delta and blocks PR)
• Warn: baseline file missing (graceful first run)

Added as a CI step in squad-ci.yml after "Run tests".

2. Test Baseline

.github/test-baseline.json — stores expected minimum test count (4,662 tests). Can only be updated with explicit human approval.

3. Copilot Instructions Directive

Added "Test Integrity" section to .github/copilot-instructions.md:

• Never delete, skip, or weaken existing tests
• Never add .skip or change assertions to weaker forms
• If a test fails, fix the CODE, not the test
• Report failures in PR — don't suppress them

Tests

8 tests in test/ci/test-count-guard.test.ts covering:

• Pass when count equals/exceeds baseline
• Fail when count decreases (with delta message)
• Warning when baseline missing
• Malformed baseline handling
• Edge cases (zero baseline)

Files Changed (5)

File	What
.github/scripts/test-count-guard.mjs	Guard script (88 lines)
.github/test-baseline.json	Baseline: 4,662 tests
.github/workflows/squad-ci.yml	CI step added
.github/copilot-instructions.md	Test integrity directive
test/ci/test-count-guard.test.ts	8 tests for guard logic

[diberry]

Retargeted to bradygaster/squad — upstream PR open.