feat(types): 统一格式落地为内核实际使用的类型 —— 私有字段泄漏归零

Phase 2 收尾（入向）。ChatRequest 从"客户端发什么就是什么"改为
UnifiedRequest（最大集合）的别名，内核真正用上了那个超集类型。

改掉的具体缺陷：

1. 私有字段泄漏归零（v1 的核心问题）
   v1 的 disable_thinking 是内核私有开关，却被序列化进交给适配器的 body，
   迫使 10 个适配器各自写 req.disable_thinking = nil 来清理。
   现在它是中立的 thinking 字段（{enabled, budget_tokens}），客户端与
   适配器都认这个词：
   - gateway/chat.go：thinkingFromClient() 把 disable_thinking 翻译成
     thinking.enabled=false；冲突时中立字段优先（它能带预算，布尔不能）
   - deepseek.lua：thinking.enabled=false → extra_body.thinking.type=disabled
   - anthropic.lua：thinking.enabled=true + budget → 原生 {type,budget_tokens}
     （enabled=false 时不传该字段——Anthropic 没有 type=disabled 这个写法）
   - 其余 7 个适配器：把"清理内核私有字段"改成"清理中立字段"

2. Tools/ToolChoice 改 json.RawMessage
   原为 []interface{}（OpenAI 原始形状 = 内核在假设某个协议）。
   内核从不解释工具 schema——它原样交给适配器，所以原始形状不损失任何
   内核侧统一性。ToolDef 结构化类型保留在包内供新适配器选用。

3. 删除 MarshalJSON 后门
   v1 手工 splice ExtraBody 进 OpenAI 形状的 body；现在 ExtraBody 是
   UnifiedRequest 上一个正正经经声明的字段。

渐进迁移：ChatRequest 保留为别名（不是新类型），49 处引用零改动编译，
语义变更不被机械改名淹没。

测试：新增 thinking_test.go（翻译链 4 例 + 泄漏回归 + tools 透传）；
vm_test.go 两处迁移到中立语义。golden 字节级未漂移。

internal/lua/adapters/anthropic.lua

@@ -254,12 +254,24 @@ function adapter.transform_request(raw_body)
         end
     end
 
-    -- ── thinking (opt-in via extra_body) ──────────────────────
-    -- Default: OFF. Client sends extra_body.thinking to enable.
-    -- Example: {"extra_body": {"thinking": {"type": "enabled", "budget_tokens": 4096}}}
-    if type(req.extra_body) == "table" and type(req.extra_body.thinking) == "table" then
+    -- ── thinking ─────────────────────────────────────────────
+    -- Anthropic 的原生字段就是 {type, budget_tokens}，所以中立语义几乎直译。
+    --
+    -- 中立格式（v2）：{ enabled = true/false, budget_tokens = N }
+    --   enabled=false → 不传该字段（Anthropic 默认关闭思考，传 type=disabled
+    --                  是非法的，这也是过去只能「不传」的原因）
+    --   enabled=true  → { type="enabled", budget_tokens=N 或默认 }
+    --
+    -- 仍接受 extra_body.thinking 的原生写法：已有部署的配置不因此失效。
+    local th = req.thinking
+    if type(th) == "table" and th.enabled ~= false then
+        local budget = th.budget_tokens
+        if budget == nil or budget <= 0 then budget = 4096 end  -- Anthropic 必填
+        anthropic_req.thinking = { type = "enabled", budget_tokens = budget }
+    elseif type(req.extra_body) == "table" and type(req.extra_body.thinking) == "table" then
         anthropic_req.thinking = req.extra_body.thinking
     end
+    req.thinking = nil
 
     -- ── system ───────────────────────────────────────────────
     if system ~= "" then

internal/lua/adapters/deepseek.lua

@@ -15,11 +15,18 @@ function adapter.transform_request(raw_body)
         req.model = "deepseek-v4-flash"
     end
     req.stream = req.stream or false
-    if req.disable_thinking then
+    -- 中立的 thinking 语义 → DeepSeek 自己的协议写法。
+    -- 客户端说「关掉思考」时 thinking.enabled == false；上游的表达方式是
+    -- extra_body.thinking.type = "disabled"。v2 之前这里读的是内核私有字段
+    -- req.disable_thinking，而那个字段曾被内核泄漏进本函数收到的 body。
+    if type(req.thinking) == "table" and req.thinking.enabled == false then
         req.extra_body = req.extra_body or {}
         req.extra_body.thinking = { type = "disabled" }
+    elseif type(req.thinking) == "table" and req.thinking.budget_tokens ~= nil then
+        req.extra_body = req.extra_body or {}
+        req.extra_body.thinking = { type = "enabled", budget_tokens = req.thinking.budget_tokens }
     end
-    req.disable_thinking = nil
+    req.thinking = nil
 
     -- V4 thinking 模式要求：带 tool_calls 的 assistant 消息必须回传 reasoning_content。
     -- OpenAI 兼容客户端不会发该字段，补空串即可通过校验。

internal/lua/adapters/github.lua

@@ -14,7 +14,13 @@ function adapter.transform_request(raw_body)
     req.temperature = req.temperature or 0.7
     req.max_tokens = req.max_tokens or 4096
     req.stream = req.stream or false
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/groq.lua

@@ -13,7 +13,13 @@ function adapter.transform_request(raw_body)
     req.temperature = req.temperature or 0.7
     req.max_tokens = req.max_tokens or 4096
     req.stream = req.stream or false
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/kimicode.lua

@@ -19,7 +19,13 @@ function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
     req.model = req.model or "kimi-k2"
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/mistral.lua

@@ -13,7 +13,13 @@ function adapter.transform_request(raw_body)
     req.temperature = req.temperature or 0.7
     req.max_tokens = req.max_tokens or 4096
     req.stream = req.stream or false
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     return json.encode(req)
 end

internal/lua/adapters/openai.lua

@@ -39,7 +39,13 @@ end
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     if req.messages then
         for _, msg in ipairs(req.messages) do

internal/lua/adapters/sensenova.lua

@@ -12,7 +12,13 @@ adapter.headers = {}
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     if req.messages then
         for _, msg in ipairs(req.messages) do

internal/lua/adapters/trae.lua

@@ -15,7 +15,13 @@ adapter.headers = {}
 function adapter.transform_request(raw_body)
     local ok, req = pcall(json.decode, raw_body)
     if not ok then return raw_body end
-    req.disable_thinking = nil
+    -- The gateway's private control fields are GONE from the body (v2): the
+    -- neutral `thinking` field carries that intent and each adapter maps it in
+    -- its own protocol terms. These upstreams have no thinking switch at all,
+    -- so the field is simply dropped here — which is what the nil-ing below
+    -- used to do, except the field being dropped is no longer the kernel's
+    -- private one.
+    req.thinking = nil
     req.extra_body = nil
     if req.messages then
         for _, msg in ipairs(req.messages) do