fix(model): DeepSeek 推理参数语义反转 + thinking 开关真发（按官方文档核对结果）

依据 DevDocs/design/04 §3.2「官方核对后的修正」9 条 + 本地 DevDocs/reference/deepseek/facts-for-cyrene.md（官方原句/URL）。
- 🔴 reasoning_effort 白名单改为官方集合 [none, low, high, max] + 别名映射(minimal→low, medium→high, xhigh→high, ultra→max)；**删除就近降级**，非法值一律丢弃+WARN。此前 none(关思考) 被降级成 low(开思考) 属语义反转
- 🔴 DeepSeek thinking 开关真发顶层 {"thinking":{"type":"enabled|disabled"}}（官方默认 enabled，不传≠关）；此前只打 WARN 导致 mode:off 静默无效
- DeepSeek 的 /anthropic 端点忽略 budget_tokens → 不再对其抬 max_tokens（会误导预算/日志），改为按官方用 output_config.effort
- 注释纠正：penalty 属接口级 deprecated（非仅思考模式）；top_p 仅思考模式生效且下限 0.95（不加该字段）
- tool_choice 雷登记：思考模式下 required/指定 tool 会 400（思考默认开）
- 旧模型名清理（deepseek-chat/reasoner 已于 2026-07-24 停用）
- 请求体级测试 + 变异验证（把 none 改回 low / 把 thinking 改回只 WARN → 对应断言 FAIL）

internal/runtime/model/model.go

@@ -158,6 +158,15 @@ type Request struct {
 	Model    string
 	Messages []Message
 	// Temperature is the sampling temperature. nil means the provider default.
+	//
+	// DeepSeek notes (design/04 §3.2 facts 3 and 4): temperature takes no effect
+	// while thinking is on. frequency_penalty and presence_penalty are deprecated
+	// at the API level — "传入该参数将不会产生任何效果" — in every mode, which is
+	// why Request has no field for them. top_p is deliberately not exposed
+	// either: it only applies while thinking is on and is clamped to >= 0.95
+	// there, is always 1.0 otherwise, so sending a configured value would be
+	// silently ineffective on one side of the switch and a hard floor on the
+	// other.
 	Temperature *float64
 	// MaxTokens caps the generated tokens. Zero means the provider default.
 	MaxTokens int
@@ -169,16 +178,28 @@ type Request struct {
 	// opaque provider-agnostic directive (e.g. {"type":"auto"} or {"type":"tool",
 	// "name":"x"}); empty means the provider default. Passed through verbatim to
 	// protocols that support it (Anthropic's tool_choice).
+	//
+	// DeepSeek red line (design/04 §3.2 fact 5): with thinking on — its default —
+	// a forcing directive ("required" / "any" / a named tool) is rejected with a
+	// 400, so thinking must be switched off explicitly first. The adapters only
+	// WARN about that combination; they never switch thinking off on their own.
 	ToolChoice any
 	// Thinking controls the provider's reasoning budget/switch (design/04 §3.2).
 	// nil means "do not intervene" and uses the provider default. A provider
 	// whose protocol has no mapping for it logs a WARN rather than silently
 	// ignoring it.
+	//
+	// "Do not intervene" is not the same as "off": DeepSeek's thinking defaults to
+	// enabled, so Enable=false is what closes it there (sent as
+	// {"thinking":{"type":"disabled"}}).
 	Thinking *Thinking
 	// ReasoningEffort is the OpenAI-compatible reasoning strength knob
 	// (reasoning_effort). Empty means "do not intervene". The value is validated
-	// against the detected provider's whitelist; an unsupported value is
-	// downgraded to the nearest allowed tier (or dropped) with a WARN.
+	// against the detected provider's whitelist; an unrecognized value is dropped
+	// with a WARN and never replaced by a different tier (design/04 §3.2 fact 1).
+	// DeepSeek additionally documents compatibility aliases (minimal→low,
+	// medium/xhigh→high, ultra→max), which are normalized to the tier the service
+	// applies.
 	ReasoningEffort string
 }