fix(trace): an action names the generator that produced it

The setup exclusion landed for the model arm only, because only a model
pick stamped a source. A seeded run returned setup's action through the
same entry with no marker, so its denominator still counted the login
while the model arm's did not, and the two are compared.

serializeAction names setup and seeded on the wire, so both arms are
counted by one rule. An already-recorded trace names nothing and keeps
exactly the count it was reported with; unattributed_actions counts those
steps so the old denominator cannot pass as the new one. TraceVersion is
deliberately unbumped: oracle-reduction refuses a differing version, and
a bump would make all 169 recorded runs unreplayable.

internal/runner/llm_source.go

@@ -65,14 +65,11 @@ type llmSource struct {
 	// can stamp the trace. lastSource is "llm" only when the LLM (not setup)
 	// chose the action; lastReasoning is the model's rationale. lastChoice is the
 	// 1-based number it picked and lastChosenAction the description it echoed, so
-	// the trace shows what the model believed it was doing. lastFromSetup says
-	// the spec's setup produced the action, which is the app being put in
-	// position rather than the generator exploring it.
+	// the trace shows what the model believed it was doing.
 	lastSource       string
 	lastReasoning    string
 	lastChoice       int
 	lastChosenAction string
-	lastFromSetup    bool
 }
 
 // llmSelection is the outcome of one LLM selection call.
@@ -99,14 +96,12 @@ func (s *llmSource) NextAction(ctx context.Context, stepIndex int) (verifier.Act
 	s.lastReasoning = ""
 	s.lastChoice = 0
 	s.lastChosenAction = ""
-	s.lastFromSetup = false
 	s.history.completeLast(s.verifier.CurrentScreen())
 
 	// Setup precedence only: the LLM replaces the seeded action root, so we run
 	// setup (e.g. login) first but never the weighted picker.
 	action, err := s.verifier.SetupAction()
 	if err == nil {
-		s.lastFromSetup = true
 		s.record(stepIndex, trace.LLMCall{Outcome: trace.LLMOutcomeSetupAction})
 		s.history.add(describeAction(action, s.verifier.Tree()))
 		return action, nil
@@ -501,9 +496,10 @@ func (h *actionHistory) recent() []historyEntry {
 	return h.entries
 }
 
-// stampActionSource records the backend that chose an action on the trace.
-// Only an LLM-selected action (not a setup action the JS path produced) carries
-// source="llm" and the model's reasoning.
+// stampActionSource names the model as the producer of an action it chose. The
+// spec's two generators name themselves on the wire (runtime-entry.ts) and
+// traceActionFor has already carried that over; a model pick is built here from
+// the candidate list, so nothing on the wire could have named it.
 func stampActionSource(traceAction *trace.Action, source ActionSource) {
 	if traceAction == nil {
 		return
@@ -518,14 +514,13 @@ func stampActionSource(traceAction *trace.Action, source ActionSource) {
 	traceAction.LLMChosenAction = llm.lastChosenAction
 }
 
-// generatorChoseAction reports whether the action the source just returned came
-// from the generator rather than from the spec's setup driving the app into
-// position. Only the model source can tell them apart: the seeded picker
-// resolves setup precedence inside the one JS call it makes, so everything it
-// returns counts as the generator's.
-func generatorChoseAction(source ActionSource) bool {
-	llm, ok := source.(*llmSource)
-	return !ok || !llm.lastFromSetup
+// generatorChoseAction reports whether this step drove the app on the action
+// generator's behalf rather than on setup's, which is the exposure a per-action
+// rate divides by. Both policies are read the same way, off the stamped trace
+// action: an unstamped one is a source the runner cannot name, and counting it
+// as setup would report a run that explored as having explored nothing.
+func generatorChoseAction(traceAction *trace.Action) bool {
+	return traceAction != nil && traceAction.Source != trace.ActionSourceSetup
 }
 
 // screenshotDataURL downscales the PNG and encodes it as a data URL for the

internal/runner/llm_source_test.go

@@ -968,13 +968,13 @@ func TestRunner_SetupAndGeneratorBothDrivingIsAHealthyRun(t *testing.T) {
 		}
 	}
 	for _, line := range lines[:2] {
-		if line.NextAction.Source != "" {
-			t.Errorf("step %d action source = %q, want none: setup chose it",
+		if line.NextAction.Source != trace.ActionSourceSetup {
+			t.Errorf("step %d action source = %q, want setup: setup chose it",
 				line.Step, line.NextAction.Source)
 		}
 	}
 	for _, line := range lines[2:] {
-		if line.NextAction.Source != "llm" {
+		if line.NextAction.Source != trace.ActionSourceModel {
 			t.Errorf("step %d action source = %q, want llm", line.Step, line.NextAction.Source)
 		}
 	}

internal/runner/runner.go

@@ -82,9 +82,8 @@ type Summary struct {
 	// GeneratorActions counts the dispatched actions the generator chose. The
 	// spec's setup drives the app into its starting position before the
 	// generator is consulted, so a run at zero here explored nothing however
-	// many actions its login fired. Only the model generator separates the two:
-	// the seeded picker resolves setup precedence inside the one JS call it
-	// makes, so everything it returns counts as the generator's.
+	// many actions its login fired. Both generators separate the two the same
+	// way, by the producer each action names on the trace.
 	GeneratorActions int
 }
 
@@ -506,7 +505,7 @@ func Run(ctx context.Context, options Options) (Summary, error) {
 		}
 		if nextErr == nil && !applySkipped {
 			summary.DispatchedActions++
-			if generatorChoseAction(actionSource) {
+			if generatorChoseAction(traceAction) {
 				summary.GeneratorActions++
 			}
 		}
@@ -1580,7 +1579,12 @@ func driverIsAndroid(ctx context.Context, options Options, logger *slog.Logger)
 }
 
 func traceActionFor(action verifier.Action, tree *hierarchy.Tree) *trace.Action {
-	traceAction := &trace.Action{Kind: string(action.Kind), X: action.X, Y: action.Y}
+	traceAction := &trace.Action{
+		Kind:   string(action.Kind),
+		X:      action.X,
+		Y:      action.Y,
+		Source: action.Source,
+	}
 	switch action.Kind {
 	case verifier.ActionKindTap, verifier.ActionKindDoubleTap, verifier.ActionKindLongPress:
 		traceAction.Selector = action.On

internal/runner/runner_test.go

@@ -223,6 +223,62 @@ func TestRunner_SeededRunRecordsNoModelCalls(t *testing.T) {
 	}
 }
 
+// seededLoginSetupSpec drives the first two steps from setup, the way a
+// login-fronted spec does, and leaves the rest to the seeded action root.
+const seededLoginSetupSpec = `
+import { always, actions, taps, typing, weighted, Tap } from "@sanderling/spec";
+globalThis.properties = { ok: always(() => true) };
+let setupTapsLeft = 2;
+globalThis.setup = actions(() => (setupTapsLeft-- > 0 ? [Tap({ on: "id:Submit" })] : []));
+globalThis.actions = weighted([1, taps], [1, typing]);
+`
+
+// TestRunner_SeededSetupActionsAreNotTheGeneratorDrivingTheApp: a seeded run's
+// login steps used to be indistinguishable from its exploration, so the arm
+// divided its defect rate by every action it dispatched while the model arm
+// divided by the ones its policy chose. The two rates were then compared.
+func TestRunner_SeededSetupActionsAreNotTheGeneratorDrivingTheApp(t *testing.T) {
+	state := newHarnessWithSpec(t, seededLoginSetupSpec)
+	state.mock.HierarchyJSON = llmTreeJSON
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	summary, err := Run(ctx, Options{
+		Duration:    30 * time.Second,
+		IdleTimeout: 20 * time.Millisecond,
+		MaxSteps:    4,
+		Driver:      state.mock,
+		Verifier:    state.verifier,
+		TraceWriter: state.writer,
+		Logger:      slog.New(slog.NewTextHandler(io.Discard, nil)),
+	})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if summary.DispatchedActions != 4 {
+		t.Errorf("DispatchedActions = %d, want 4: every step drove the app",
+			summary.DispatchedActions)
+	}
+	if summary.GeneratorActions != 2 {
+		t.Errorf("GeneratorActions = %d, want 2: the picker drove the two steps setup left it",
+			summary.GeneratorActions)
+	}
+	lines := readTraceLines(t, state.writer.Directory())
+	if len(lines) != 4 {
+		t.Fatalf("wrote %d trace lines, want 4", len(lines))
+	}
+	for _, line := range lines[:2] {
+		if line.NextAction == nil || line.NextAction.Source != trace.ActionSourceSetup {
+			t.Errorf("step %d action = %+v, want one named setup", line.Step, line.NextAction)
+		}
+	}
+	for _, line := range lines[2:] {
+		if line.NextAction == nil || line.NextAction.Source != trace.ActionSourceSeeded {
+			t.Errorf("step %d action = %+v, want one named seeded", line.Step, line.NextAction)
+		}
+	}
+}
+
 func TestRunner_MaxStepsStopsAfterExactlyNSteps(t *testing.T) {
 	state := newHarness(t)
 

internal/runner/testdata/trace.jsonl

@@ -1,4 +1,4 @@
-{"extractor_changes":{"extractor_0":{"prev":null,"curr":0}},"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120}},"residuals":{"balanceNonNegative":{"op":"true"}},"step":1,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
-{"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120}},"residuals":{"balanceNonNegative":{"op":"true"}},"step":2,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
-{"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120}},"residuals":{"balanceNonNegative":{"op":"true"}},"step":3,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
-{"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120}},"residuals":{"balanceNonNegative":{"op":"true"}},"step":4,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
+{"extractor_changes":{"extractor_0":{"prev":null,"curr":0}},"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120},"source":"seeded"},"residuals":{"balanceNonNegative":{"op":"true"}},"step":1,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
+{"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120},"source":"seeded"},"residuals":{"balanceNonNegative":{"op":"true"}},"step":2,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
+{"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120},"source":"seeded"},"residuals":{"balanceNonNegative":{"op":"true"}},"step":3,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}
+{"hierarchy":{"elements":[{"resourceId":"HomeScreen","bounds":{"left":0,"top":0,"right":0,"bottom":0},"attrs":{"editable":"false","resource-id":"HomeScreen"}},{"resourceId":"next","clickable":true,"enabled":true,"bounds":{"left":40,"top":80,"right":240,"bottom":160},"attrs":{"bounds":"[40,80,240,160]","clickable":"true","editable":"false","enabled":"true","resource-id":"next"}}],"depths":[0,1]},"next_action":{"kind":"Tap","selector":"id:next","resolved_bounds":{"x":40,"y":80,"width":200,"height":80},"tap_point":{"x":140,"y":120},"source":"seeded"},"residuals":{"balanceNonNegative":{"op":"true"}},"step":4,"timestamp":"0001-01-01T00:00:00Z","trace_version":1}