docs: retire the criterion this version disproved, everywhere it survived

The reasoning_tokens docstring was still teaching downstream to treat
None or 0 as no reasoning. The changelog and the schema page had both
been corrected; the docstring had not, and it is the copy that ships in
the wheel and shows up on hover. Someone writing a report from it would
have counted every real MiniMax reasoning call as not reasoning, which
is issue #16 all over again with the tests green.

The original wording stays, since reading pre-1.3.1 rows still needs
it. What follows it now says when it expired and what to read instead.

Two more places had drifted the same way: the changelog and the
architecture doc described the throttle and the cache fallback as they
were before this review, which is to say as the opposite of what the
code now does.

The claim that the two throttle sets would suppress each other does not
survive checking, as the mutation testing showed: their key spaces do
not overlap. Keeping them apart is still right, but for the honest
reason, which is that the two warnings have unrelated lifetimes.

CHANGELOG.md

@@ -68,7 +68,7 @@ issue 判定「M3 开启推理静默失效,模型不推理」。实测推翻了
 | 开启 | `absent` | 已注入开启参数,上游却明确上报未推理 |
 | 开启 | `unknown` | 已注入开启参数,但本路径观测不到;若为非流式,推理内容可能已计费却不回传 |
 
-`关闭 × unknown` 与「调用方没提要求」两类**有意不表态**: 前者没有证伪力,拿它报警等于每次关闭调用都喊一遍(M3 关闭档恒落此档),噪声即等于没有告警。同一 `(模型, 方向)` 只喊一次。
+`关闭 × unknown` 与「调用方没提要求」两类**有意不表态**: 前者没有证伪力,拿它报警等于每次关闭调用都喊一遍(M3 关闭档恒落此档),噪声即等于没有告警。同一 `(源, 模型, 方向)` 只喊一次,文案点名出问题的源——多源多账号下同一模型跨 N 个源是常态,键漏掉源名会让第一个出问题的源喊完之后其余源永久静音,而告警也定位不到该查哪个网关。
 
 **保障的覆盖面必须说清楚**: 对账只在可观测路径上成立(推理若真的发生,流式路径会带出正文,翻成 `observed` 触发告警);M3 非流式那种两个信号双缺的路径,没有任何保障——本版让它可见,但不能让它可判。
 
@@ -80,10 +80,10 @@ issue 判定「M3 开启推理静默失效,模型不推理」。实测推翻了
 
 ### 其他
 
-- 缓存回放的 `thinking_observation` 是 `ThinkingObservation` 枚举实例而非裸字符串: JSON 复活出来的是 `str`,与字段注解分叉,`CacheMW._rehydrate` 现在显式转换。旧版本写入的缓存值若不在三态值域内,转换失败由既有降级路径吞成「按未命中回源」——方向正确,不会把坏缓存当好数据用。
+- 缓存回放的 `thinking_observation` 是 `ThinkingObservation` 枚举实例而非裸字符串: JSON 复活出来的是 `str`,与字段注解分叉,`CacheMW._rehydrate` 现在显式转换。取值不在本版三态值域内时(多个项目共用同一 Redis、先升级的那个写入了新态)**降级为 `unknown` 并单独告警,响应内容照常复活**——一个纯可观测性字段不该有能力作废内容完好的缓存,否则未升级的项目会在这些 key 上每次真打网关、随后覆写回旧值,两个版本互相打对方的缓存;「整条作废」只留给真正破坏内容完整性的失败。
 - M3 的推理能力 `evidence` 刷新到 2026-08-25 复测。`can_disable` **仍为 `True`**(`reasoning_effort=none` → prompt 194 = 基线、completion 3、无正文,声明依然成立),同时补记两条限制: 推理信号在非流式路径不可观测;`enable_thinking` 与 `thinking={"type":"enabled"}` 对该模型无效,只有 `reasoning_effort` 是真开关。
 - `TransportResult` 同步新增该字段并由 `RetryMW` 透传;裁定在 `openai_compat` 的流式与非流式**两条**组装路径各做一次。
-- 遥测的新列只经 `TelemetryEmitter._record` 这一个出口下沉给 recorder(单一 helper 铁律),且在那里由枚举归一化为裸 `str`——`StrEnum` 虽是 `str` 子类,asyncpg 的参数编码对 `str` 子类不保证接受,而遥测写失败只是一条 warning,这类问题不会当场炸,只会让 Postgres 那一路悄悄少一列数据。
+- 遥测的新列只经 `TelemetryEmitter._record` 这一个出口下沉给 recorder(单一 helper 铁律),且在那里由枚举归一化为裸 `str`——`StrEnum` 虽是 `str` 子类,asyncpg 的参数编码对 `str` 子类不保证接受,而遥测写失败只是一条 warning,这类问题不会当场炸,只会让 Postgres 那一路悄悄少一列数据。归一化按外部输入防御: `LLMResponse` 无运行时校验,下游填裸 `str` 完全自然,而直接取 `.value` 会抛异常并被降级路径吞成**丢掉整行**遥测;域外取值同样只降级记 `unknown` 并单独告警,不拿整行当代价。
 
 
 ## 1.3.0(2026-08-24)