feat(provider): 端点机制去厂商化 —— 适配器声明 endpoint_stream

内核不再知道 Gemini 的流式动词。之前 provider.ChatURL 对
":generateContent" 做字符串替换（注释自认 "Gemini is the only such
adapter"），把厂商协议知识放进了调度器。

现在适配器自己声明：

    adapter.endpoint_stream = "/v1beta/models/{model}:streamGenerateContent"

内核只比较两个不透明字符串，四级优先级：

    source.endpoint_stream > source.endpoint > adapter.endpoint_stream
                                          > adapter.endpoint

- staticInfo.endpointStream + VM.StreamEndpoint() 编译期提取
- config.Source.StreamEndpoint（yaml: endpoint_stream）
- gemini.lua 声明 endpoint_stream
- 内核代码 generateContent 计数归零（仅注释/适配器持有）

首版实现踩了一个回归并当场修掉：源自定义 endpoint 时适配器的 stream
endpoint 会错误覆盖用户的显式路径。TestEndpointPrecedence 把四级优先级
钉死，防止再犯。

验证：go test -tags luajit ./... 全绿；golden 字节级未漂移。

internal/lua/adapters/gemini.lua

@@ -18,6 +18,10 @@ adapter.version = "2.0.0"
 -- can never work. A source that overrides `endpoint:` bypasses the template
 -- entirely and must then spell the whole path itself.
 adapter.endpoint = "/v1beta/models/{model}:generateContent"
+-- 流式走同一路径的不同动词。这个事实以前写在 Go 内核里
+-- （provider.ChatURL 对 ":generateContent" 做字符串替换），属于把厂商协议
+-- 知识放进调度器；现在由适配器自己声明，内核只比较两个不透明字符串。
+adapter.endpoint_stream = "/v1beta/models/{model}:streamGenerateContent"
 adapter.headers = {}
 
 -- Gemini API: POST /v1/models/{model}:generateContent

internal/lua/vm.go

@@ -67,7 +67,18 @@ type staticInfo struct {
 	name     string
 	version  string
 	endpoint string
-	headers  map[string]string
+	// endpointStream is the streaming counterpart of endpoint. It exists so the
+	// KERNEL never has to know a vendor's streaming path: Gemini streams over the
+	// same path with a different verb (:generateContent -> :streamGenerateContent),
+	// and the verb is part of the path, so the two cannot both be static.
+	//
+	// Before this field the kernel carried `strings.Replace(ep,
+	// ":generateContent", ":streamGenerateContent", 1)` — vendor protocol
+	// knowledge inside the dispatcher. It now lives here, where it belongs:
+	// the adapter declares both endpoints and the kernel just picks one.
+	// An adapter that omits endpoint_stream uses endpoint for both (most do).
+	endpointStream string
+	headers        map[string]string
 }
 
 // adapterPool owns the worker states for one adapter. Idle workers are held in
@@ -923,6 +934,22 @@ func (v *VM) Endpoint(name string) string {
 	return p.static.endpoint
 }
 
+// StreamEndpoint returns the adapter's streaming endpoint, or "" when it
+// declares none (in which case the caller keeps using Endpoint).
+//
+// This is the whole mechanism that keeps vendor streaming verbs out of the
+// kernel: an adapter whose stream path differs declares endpoint_stream, and
+// the dispatcher only ever compares two opaque strings.
+func (v *VM) StreamEndpoint(name string) string {
+	p := v.pool(name)
+	if p == nil {
+		return ""
+	}
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.static.endpointStream
+}
+
 // ---- static inspection (compile-once at load time) ----
 
 func inspectScript(name, code string) (*staticInfo, error) {
@@ -943,6 +970,7 @@ func inspectScript(name, code string) (*staticInfo, error) {
 	}
 	info.version = readString(L, -1, "version")
 	info.endpoint = readString(L, -1, "endpoint")
+	info.endpointStream = readString(L, -1, "endpoint_stream")
 	if getFieldBatch(L, "headers") {
 		luaTableToMap(L, -1, info.headers)
 		L.Pop(1)