Align spans with the OTel GenAI convention (N3 WP4)

Why: acceptance clause 3 requires span names and attributes to follow
the convention; today names were bare words with two ad-hoc attribute
keys, and every span got a fresh random trace with no parentage.
Evidence: GOWORK=off go vet + go test -count=1 ./... green;
TestSpanNamesFollowGenAIConvention (name suffixes, gen_ai.* key enum,
intValue usage, service.name resource, prompt/reserved-key exclusion),
TestLoopSpansFollowGenAIConvention (invoke_agent root with shared
trace, chat/execute_tool children with parent links, usage attrs on
chat spans); the pre-existing prompt-body denylist test passes
unchanged.
Impact: span names become invoke_agent <agent> / chat <model> /
execute_tool <tool>; attributes move to gen_ai.* with usage as
intValue; one Run shares one trace with invoke_agent as root. New
Config.AgentName/ProviderName (documented defaults keep old hosts
meaningful); telemetry StartFunc gains WithServiceName/WithEndpoint
options (defaults preserve current behavior).
Refs: Agent-System-Documents#43 (N3 plan WP4)

runtime/loop.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"strconv"
 	"strings"
 	"sync"
 	"time"
@@ -12,6 +13,7 @@ import (
 	"agent-native/runtime/llm"
 	"agent-native/runtime/reminders"
 	"agent-native/runtime/session"
+	"agent-native/runtime/telemetry"
 	"agent-native/runtime/tools"
 
 	"logging-system/logsys"
@@ -116,6 +118,15 @@ type Config struct {
 	// 本次恢复写入的配对摘要以系统消息注入（不持久化，已在 journal 里）。
 	// 为假（默认）保持既有语义：Session 为权威上下文。
 	JournalOnly bool
+	// AgentName 是 invoke_agent span 的名字后缀（OTel GenAI 约定
+	// "invoke_agent <agent>"）。空 = "agent-native"。
+	AgentName string
+	// ProviderName 是 gen_ai.system 属性值（如 openai / anthropic）。
+	// 空 = "openai"。
+	ProviderName string
+	// traceCtx 是本次 Run 的 span 关联（Run 启动时建立；指针随 Config
+	// 副本传播到各轮与各工具）。
+	traceCtx *traceIDs
 	// Hooks 是进程内钩子。空总线不改变停止、提醒和拒绝次数。
 	Hooks *hooks.Bus
 	// PermissionAsk 在钩子对一次工具调用表态 ask 时向宿主请求裁决。
@@ -176,7 +187,12 @@ func Run(ctx context.Context, cfg Config, user llm.Message) (Result, error) {
 		cfg.ToolExecution = ExecSequential
 	}
 	serializeHostCallbacks(&cfg)
-	endRun := startSpan(cfg, "invoke_agent", map[string]string{"model": cfg.Model})
+	cfg.traceCtx = newTraceContext()
+	endRun := startSpan(cfg, "invoke_agent "+agentName(cfg), map[string]string{
+		"gen_ai.system":          providerName(cfg),
+		"gen_ai.agent.name":      agentName(cfg),
+		"gen_ai.request.model":   cfg.Model,
+	}, true)
 	defer endRun()
 	var messages []llm.Message
 	if cfg.Session != nil && !cfg.JournalOnly {
@@ -438,7 +454,11 @@ func systemText(cfg Config) string {
 }
 
 func consume(ctx context.Context, cfg Config, messages []llm.Message) (llm.Event, error) {
-	endSpan := startSpan(cfg, "chat", map[string]string{"model": cfg.Model})
+	spanAttrs := map[string]string{
+		"gen_ai.system":        providerName(cfg),
+		"gen_ai.request.model": cfg.Model,
+	}
+	endSpan := startSpan(cfg, "chat "+cfg.Model, spanAttrs, false)
 	defer endSpan()
 	var decls []llm.ToolDecl
 	if cfg.Tools != nil {
@@ -475,6 +495,11 @@ func consume(ctx context.Context, cfg Config, messages []llm.Message) (llm.Event
 			sawEnd = true
 		}
 	}
+	// usage 在 span 结束前补进属性（结束时过滤导出，见 telemetry 工厂）。
+	if sawEnd {
+		spanAttrs["gen_ai.usage.input_tokens"] = itoa(end.Usage.PromptTokens)
+		spanAttrs["gen_ai.usage.output_tokens"] = itoa(end.Usage.CompletionTokens)
+	}
 	if !sawEnd {
 		return llm.Event{Message: llm.Message{StopReason: llm.StopError}}, errString("stream ended without a message")
 	}
@@ -686,7 +711,10 @@ func runOne(ctx context.Context, cfg Config, call llm.ToolCall) llm.Message {
 		emit(cfg, Event{Type: "tool_execution_end", Tool: call.Name, Text: msg.Content})
 		return msg
 	}
-	endSpan := startSpan(cfg, "execute_tool", map[string]string{"tool": call.Name})
+	endSpan := startSpan(cfg, "execute_tool "+call.Name, map[string]string{
+		"gen_ai.system":    providerName(cfg),
+		"gen_ai.tool.name": call.Name,
+	}, false)
 	defer endSpan()
 	res := tools.Result{Content: "no tools configured", IsError: true}
 	known := cfg.Tools != nil && cfg.Tools.Has(call.Name)
@@ -864,10 +892,38 @@ func commitToolIntent(cfg Config, call llm.ToolCall) error {
 	})
 }
 
-func startSpan(cfg Config, name string, attrs map[string]string) func() {
+func itoa(n int) string { return strconv.Itoa(n) }
+
+// traceIDs 是一次 Run 的 span 关联：根 trace + invoke_agent 根 span。
+type traceIDs struct {
+	traceID string
+	rootID  string
+}
+
+// newTraceContext 建立一次 Run 的关联（Trace 为 nil 时也给空壳，调用方免判空）。
+func newTraceContext() *traceIDs {
+	return &traceIDs{traceID: telemetry.NewTraceID(), rootID: telemetry.NewSpanID()}
+}
+
+// startSpan 开一个 OTel GenAI span：root=true 时作 invoke_agent 根（无 parent），
+// 否则作子 span（parent = 根）。attrs 由调用方持有引用——存续期间可补
+// （usage 在结束时随过滤导出）。
+func startSpan(cfg Config, name string, attrs map[string]string, root bool) func() {
 	if cfg.Trace == nil {
 		return func() {}
 	}
+	if attrs == nil {
+		attrs = map[string]string{}
+	}
+	if cfg.traceCtx != nil {
+		attrs[telemetry.AttrTraceID] = cfg.traceCtx.traceID
+		if root {
+			attrs[telemetry.AttrSpanID] = cfg.traceCtx.rootID
+		} else {
+			attrs[telemetry.AttrSpanID] = telemetry.NewSpanID()
+			attrs[telemetry.AttrParentSpanID] = cfg.traceCtx.rootID
+		}
+	}
 	end := cfg.Trace(name, attrs)
 	if end == nil {
 		return func() {}
@@ -875,6 +931,21 @@ func startSpan(cfg Config, name string, attrs map[string]string) func() {
 	return end
 }
 
+// agentName / providerName 是 OTel 属性的默认值解析。
+func agentName(cfg Config) string {
+	if strings.TrimSpace(cfg.AgentName) != "" {
+		return cfg.AgentName
+	}
+	return "agent-native"
+}
+
+func providerName(cfg Config) string {
+	if strings.TrimSpace(cfg.ProviderName) != "" {
+		return cfg.ProviderName
+	}
+	return "openai"
+}
+
 func clearOldResults(cfg Config, messages []llm.Message) []llm.Message {
 	window := cfg.ContextWindow
 	if window <= 0 {