feat: implement proper orchestrator workflow with reconciliation loop

Rewrite OrchestratorWorkflow as true reconciliation loop:
- PlanningActivity decides what tasks to dispatch
- Fan-out TaskUnit workflows for parallel execution
- Each TaskUnit runs Implementer → Test → Judge → Commit
- Judge reviews code quality, retries on failure with lessons
- Fan-in waits for all TaskUnits
- Board update and squash merge on success
- continue-as-new for long-running workflows
- Proper error handling and signal support

Key changes:
- statemachine/orchestrator.go: Reconciliation loop (Plan → Dispatch → Review → Repeat)
- statemachine/taskunit.go: Task execution with retry loop & judge review
- statemachine/types.go: Updated TaskUnitInput/Output for new workflow
- cmd/worker/main.go: Register RunIntegrationTestActivity
- action/integration.go: Renamed from integration_test.go (fix Go build issue)

Models:
- Planner: reasoning (OpenAI-compatible from local LLM API)
- Judge: reasoning (reviews diff + tests, gates success)
- Implementer: ornith:35b (executes tasks)

Verification: go build ./cmd/worker ./cmd/starter ✓

statemachine/types.go

@@ -69,20 +69,23 @@ type OrchestratorOutput struct {
 
 // TaskUnitInput is the input to the TaskUnit workflow.
 type TaskUnitInput struct {
-	TaskID          string
-	TargetRepoPath  string
-	JudgeSpec       PromptSpec
-	ImplementerSpec PromptSpec
-	BaseTimeout     time.Duration
-	MaxJudgeRetries int
+	TaskID         string
+	RemoteURL      string
+	TargetRepoPath string
+	Milestone      string
+	Config         OrchestratorConfig
+	DryRun         bool
 }
 
 // TaskUnitOutput is the output of the TaskUnit workflow.
 type TaskUnitOutput struct {
 	TaskID   string
-	Verdict  string // "pass" or "fail"
-	Critique string
+	Status   string // "success" or "failed"
+	Verdict  string // "pass" or "fail" from judge
+	Critique string // feedback from judge
 	Branch   string
+	Reason   string // error reason if failed
+	Changes  string // summary of changes
 }
 
 // SkillRef references a skill source.