Exercise material notice outcomes in the browser fixture

Why: The task-owned mechanism must reproduce stage progress, partial and empty failures, and source-bound choices without a paid provider call.

Evidence: Backend vet/full Go tests and live-tag terminal tests passed; six task-owned HTTP/Playwright cases passed with this fixture.

Impact: Evaluation fixture only; model comparison still disables notice turns unless the explicit mechanism browser flag is set.

Refs: U3 r0.4 Task 6

internal/server/work_checkpoint_eval_test.go

@@ -36,10 +36,13 @@ import (
 	core "agent-system-core"
 )
 
-// evalForeground excludes the existing automatic notice model turn in BOTH
-// baseline and candidate. It is not evidence for notice quality; normal notice
-// behavior is covered separately by the production regression suite.
-type evalForeground struct{ delegate runtime.Runner }
+// Paid baseline/candidate comparisons exclude the automatic notice model turn
+// equally. Only the task-owned mechanism browser fixture enables it to check
+// the user-visible choice and failure paths without provider requests.
+type evalForeground struct {
+	delegate    runtime.Runner
+	allowNotice bool // only the task-owned, zero-paid browser mechanism enables this
+}
 
 func (e evalForeground) Name() string { return "evaluation-foreground" }
 
@@ -56,7 +59,7 @@ func evalA01Budget(label string) (int, bool) {
 
 func (e evalForeground) Execute(rc runtime.RunCtx) error {
 	history := rc.History()
-	if len(history) > 0 && history[len(history)-1].Role == core.RoleSystem {
+	if !e.allowNotice && len(history) > 0 && history[len(history)-1].Role == core.RoleSystem {
 		return rc.SetAssistantParts([]core.Part{{Type: core.PartText, Text: "后台结果已回投任务卡（评测不再调用通知模型）"}})
 	}
 	return e.delegate.Execute(rc)
@@ -65,26 +68,53 @@ func evalMechanismStream(ctx context.Context, model string, msgs []agentllm.Mess
 	out := make(chan agentllm.Event, 2)
 	defer close(out)
 	checkpoint := false
+	notice := false
+	choiceReady := false
+	sourceText := ""
+	toolReplies := 0
 	for _, d := range decls {
 		if d.Name == "publish_work_checkpoint" {
 			checkpoint = true
 		}
 	}
+	for _, m := range msgs {
+		if m.Role == agentllm.RoleUser {
+			sourceText += m.Content
+		}
+		if m.Role == agentllm.RoleTool {
+			toolReplies++
+		}
+		if strings.HasPrefix(strings.TrimSpace(m.Content), "<system-reminder>") && strings.Contains(m.Content, "后台子代理「") {
+			notice = true
+			choiceReady = strings.Contains(m.Content, runtime.MaterialChoiceReadyMarker)
+		}
+	}
+	if notice {
+		content := "后台整理状态已回投工作卡；请按卡片中的结果继续。"
+		var calls []agentllm.ToolCall
+		if choiceReady {
+			// Deliberately supply unrelated options: the runtime must replace them
+			// with the host-extracted, source-bound options from the notice.
+			calls = []agentllm.ToolCall{{ID: "notice-choice", Name: "confirm_choice", Arguments: `{"question":"模型提出的交付物？","options":[{"id":"fake-a","label":"虚构报告","description":"无来源","impact":"虚构"},{"id":"fake-b","label":"虚构方案","description":"无来源","impact":"虚构"}]}`}}
+		}
+		reason := agentllm.StopEnd
+		if len(calls) > 0 {
+			reason = agentllm.StopToolCalls
+		}
+		out <- agentllm.Event{Kind: agentllm.KindMessageEnd, Message: agentllm.Message{Role: agentllm.RoleAssistant, Content: content, ToolCalls: calls, StopReason: reason}, Usage: agentllm.Usage{PromptTokens: 10, CompletionTokens: 10, TotalTokens: 20}}
+		return out
+	}
 	content := "已收到材料，正在后台整理；请结合阶段发现继续工作。"
 	var calls []agentllm.ToolCall
 	if checkpoint {
-		sourceText := ""
-		count := 0
-		for _, m := range msgs {
-			if m.Role == agentllm.RoleUser {
-				sourceText += m.Content
-			}
-			if m.Role == agentllm.RoleTool {
-				count++
-			}
-		}
+		count := toolReplies
 		matches := mechanismCheckpointSources(sourceText)
 		matches = filterCheckpointSources(matches)
+		if (count == 0 && strings.Contains(sourceText, "[fixture:zero-failure]")) ||
+			(count == 1 && strings.Contains(sourceText, "[fixture:partial-failure]")) {
+			out <- agentllm.Event{Kind: agentllm.KindMessageEnd, Message: agentllm.Message{Role: agentllm.RoleAssistant, StopReason: agentllm.StopEnd}, Usage: agentllm.Usage{PromptTokens: 10, CompletionTokens: 10, TotalTokens: 20}, Err: fmt.Errorf("task-owned mechanism failure after %d checkpoint(s)", count)}
+			return out
+		}
 		if count < 2 && len(matches) > 0 {
 			var block int
 			fmt.Sscan(matches[0][2], &block)
@@ -108,7 +138,14 @@ func evalMechanismStream(ctx context.Context, model string, msgs []agentllm.Mess
 		} else {
 			content = "整理结束，日期与验收条件需要核对；完整范围以原件为准。"
 		}
-		time.Sleep(200 * time.Millisecond)
+		if strings.Contains(sourceText, "[fixture:slow-stages]") {
+			time.Sleep(1500 * time.Millisecond)
+		} else {
+			time.Sleep(200 * time.Millisecond)
+		}
+	}
+	if toolReplies >= 2 && strings.Contains(sourceText, "[fixture:choice]") {
+		content = "整理结束，日期与验收条件需要核对；完整范围以原件为准。\n\n## 建议后续处理\n1. **会议摘要**\n   - 依据：材料记载上线六月。\n2. **验收清单**\n   - 依据：验收条件需要补齐。\n3. **联调计划**\n   - 依据：三项需求依赖外部接口。"
 	}
 	if len(calls) == 0 {
 		out <- agentllm.Event{Kind: agentllm.KindTextDelta, Text: content}
@@ -208,7 +245,7 @@ func TestServeWorkCheckpointEval(t *testing.T) {
 		front.Stream = evalMechanismStream
 		sub.Stream = evalMechanismStream
 	}
-	mgr := runmgr.New(st, b, registry, prof, settingSvc, evalForeground{front}, sub, &runtime.MockRunner{}, &runtime.MockSubagent{})
+	mgr := runmgr.New(st, b, registry, prof, settingSvc, evalForeground{delegate: front, allowNotice: mode == "mechanism" && os.Getenv("WORK_CHECKPOINT_MECHANISM_NOTICES") == "1"}, sub, &runtime.MockRunner{}, &runtime.MockSubagent{})
 	mgr.SetWorkspaceRoot(filepath.Join(dir, "workspaces"))
 	mgr.SetInstanceID("checkpoint-eval")
 	st.SetInstanceID("checkpoint-eval")