fix(gemini): endpoint 自相矛盾导致预置模板必失败

gemini.lua 里 adapter.endpoint = "/v1/models"，而它自己的注释写的是
  POST /v1/models/{model}:generateContent
两者矛盾，而 Go 侧是静态拼接（provider.URL = base_url + endpoint），拼不出
模型名。预置模板 "Google Gemini"（base_url=.../v1beta）于是会 POST 到
  https://generativelanguage.googleapis.com/v1beta/v1/models
既多一段 /v1，又缺 :generateContent——那是 Gemini 的模型**列表**端点，
对 POST 返 405。所以任何用户从模板建这个源，拿到的都是必定失败的源。

实测确认影响范围：线上 21 个源里没有 gemini（openai×15 / trae / sensenova /
opencodezen / deepseek / anthropic / agentrouter），所以是潜伏缺陷。

修法：endpoint 改成模板 `/v1beta/models/{model}:generateContent`，新增
provider.ChatURL(model, stream)：
  - 用 **PathEscape** 替换 {model}——模型 id 进的是 URL 路径，不转义的话
    一个 "/" 就会静默指向另一个资源（判据里用 RequestURI 而非 URL.Path
    断言，因为后者是解码后的，看不出 %2F）；
  - 流式把 ":generateContent" 换成 ":streamGenerateContent"（同一个路径、
    不同动词，也在路径里）。替换刻意只认这个精确后缀，免得别的适配器
    仅仅提到这个词就被改写；
  - source 自己设的 endpoint: 仍然优先，模板被整体跳过。
Chat / ChatStream / probeChat 三处调用点改为传本次请求真实的 model——AUTO
按槽位把 req.Model 钉死，所以 URL 必须跟随**请求**的模型，用源默认模型会让
多模型源每次都打同一个（还记到别的模型的账上）。

判定静态 endpoint 的其他 10 个适配器零影响（TestNonGeminiEndpointsAreUntouched）。

顺带：Stats 的 mutex 不是可重入的，导出方法自己加锁、*Locked 后缀要求调用
方持锁。持锁调导出方法会死锁——我的探针真卡死过一次（直到 10 分钟超时）。
补上 LOCKING 注释，并加判据把这条规则钉住（含一个 20 秒上限的行为判据，
让未来的重构撞死锁时快速失败而不是拖满整个套件）。

internal/provider/provider.go

@@ -531,7 +531,9 @@ func isAutoID(s string) bool {
 	return s == "" || strings.EqualFold(s, "AUTO")
 }
 
-// Endpoint resolves the upstream chat path.
+// Endpoint resolves the upstream chat path. It may contain the placeholder
+// "{model}"; use URL (or ChatURL) rather than calling this directly when the
+// path has to be usable.
 func (p *Provider) Endpoint() string {
 	if p.cfg.Endpoint != "" {
 		return p.cfg.Endpoint
@@ -553,8 +555,40 @@ func (p *Provider) ImageEndpoint() string {
 	return "/v1/images/generations"
 }
 
+// modelPlaceholder is the marker an adapter puts in its endpoint template when
+// the upstream carries the model id in the PATH rather than in the body.
+// Gemini is the only such adapter: POST /v1beta/models/{model}:generateContent.
+// Every other adapter's endpoint is a fixed path, so substituting is a no-op
+// for them.
+const modelPlaceholder = "{model}"
+
+// ChatURL resolves the full chat URL for a specific model.
+//
+// Two substitutions happen here, both driven by the adapter's endpoint template:
+//
+//   - "{model}" is replaced by the model id actually being sent. Without this
+//     the gateway would POST to Gemini's model-LIST endpoint, which answers 405.
+//   - for a streaming request, the ":generateContent" verb becomes
+//     ":streamGenerateContent". Gemini streams over the same path with a
+//     different verb, and the verb is part of the path, so the two cannot both
+//     be static. The rewrite is deliberately narrow: it only fires on the exact
+//     ":generateContent" suffix, so an adapter whose endpoint merely mentions
+//     the word keeps its path untouched.
+func (p *Provider) ChatURL(model string, stream bool) string {
+	ep := p.Endpoint()
+	if strings.Contains(ep, modelPlaceholder) {
+		// A model id is put in a URL path, so it must be escaped: an id with a
+		// slash would otherwise silently address a different resource.
+		ep = strings.ReplaceAll(ep, modelPlaceholder, url.PathEscape(model))
+	}
+	if stream {
+		ep = strings.Replace(ep, ":generateContent", ":streamGenerateContent", 1)
+	}
+	return strings.TrimRight(p.cfg.BaseURL, "/") + ep
+}
+
 func (p *Provider) URL() string {
-	return strings.TrimRight(p.cfg.BaseURL, "/") + p.Endpoint()
+	return p.ChatURL("", false)
 }
 
 func (p *Provider) ImageURL() string {
@@ -750,10 +784,11 @@ func (p *Provider) probeChat(ctx context.Context) (bool, string) {
 		body, err := json.Marshal(probe)
 		if err == nil {
 			var hdr http.Header
-			if hdrs, herr := p.buildHeaders(string(body), p.URL(), ""); herr == nil {
+			probeURL := p.ChatURL(model, false)
+			if hdrs, herr := p.buildHeaders(string(body), probeURL, ""); herr == nil {
 				hdr = hdrs
 			}
-			raw, status, derr := p.do(ctx, p.URL(), string(body), hdr)
+			raw, status, derr := p.do(ctx, probeURL, string(body), hdr)
 			if derr != nil {
 				msg = derr.Error()
 			} else if status == 200 {
@@ -1084,11 +1119,12 @@ func (p *Provider) Chat(ctx context.Context, req *types.ChatRequest) (*types.Uni
 	if err != nil {
 		return nil, err
 	}
-	hdrs, err := p.buildHeaders(body, p.URL(), req.ClientSession)
+	chatURL := p.ChatURL(model, false)
+	hdrs, err := p.buildHeaders(body, chatURL, req.ClientSession)
 	if err != nil {
 		return nil, err
 	}
-	raw, status, err := p.do(ctx, p.URL(), body, hdrs)
+	raw, status, err := p.do(ctx, chatURL, body, hdrs)
 	if err != nil {
 		// a client disconnect or cancelled context is neither a success nor
 		// a failure for scheduling purposes — only upstream errors count
@@ -1144,7 +1180,10 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 		p.Release()
 		return nil, err
 	}
-	hdrs, err := p.buildHeaders(body, p.URL(), req.ClientSession)
+	// stream=true so a path-carried model plus the streaming verb is resolved
+	// for THIS request's model, not the source default.
+	streamURL := p.ChatURL(model, true)
+	hdrs, err := p.buildHeaders(body, streamURL, req.ClientSession)
 	if err != nil {
 		p.Release()
 		return nil, err
@@ -1156,7 +1195,7 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 	}
 	rc := make(chan respOrErr, 1)
 	go func() {
-		resp, err := p.doRawStream(ctx, p.URL(), body, hdrs)
+		resp, err := p.doRawStream(ctx, streamURL, body, hdrs)
 		rc <- respOrErr{resp, err}
 	}()