fix: address M2.5 verifier findings before merge

AIMD ceiling now respects per-source max_concurrency and the pacer is
assembled explicitly in the client; MIN_CALLS parses as strict int;
acceptance doc corrects source-5 attempt count to 549; design and
migration notes aligned with implemented 429/stall/suppression
semantics and AIMD constants documented.

research-wiki/migrations/chsanalyzer.md

@@ -188,5 +188,7 @@ stack = ExtractionProviderStack(
 | 开路时长指数递增,`retry_after_s` 上限由 cooldown_s(60s)变为 max_cooldown_s(缺省 300s) | **G1 消费点注意**: tracking.py arq defer 延期时长最多放大 5 倍;属期望行为(反复坏的源就该等更久) |
 | 阈值自动抬升由全局并发改为源级并发 | CHS .env 注释约定的本意(防并发误熔)保留;全局并发抬升是 M2 错误移植,已废除 |
 | 缺省选源 round_robin → health_aware(EWMA×在途 P2C) | 显式配置 `LLM__SELECTOR=round_robin` 者行为不变;迁移时建议直接吃新缺省 |
-| 429 不消耗重试预算(pushback 语义),调用级时间上限 = stall_window_s | 饱和期调用延迟上限从 ~3×backoff 变为 stall_window(300s);CHS 快失败偏好者可调小 `LLM__BACKPRESSURE__STALL_WINDOW_S` |
+| 429 不消耗重试预算(pushback 语义),调用级时间兜底 = 双条件 stall(本地超窗且全局无进展) | 饱和期调用延迟大幅拉长(全局有进展时可超 stall_window);CHS 快失败偏好者可调小 `LLM__BACKPRESSURE__STALL_WINDOW_S` |
+| **AIMD 自适应并发**(库常量: 每源初始 8、429 削减 ×0.5、成功 +1/limit、下限 1、ceiling=max(64, 源级 max_concurrency)) | 冷启动每源并发从"无上限"变为 8 起步爬升——高吞吐下游首分钟吞吐低于旧行为;无 env 开关(韧性常量,同 jitter 系数先例),需要禁用者构造函数注入自定义 pacer |
+| 连败通道受健康证据抑制(窗口样本 ≥ min_calls 且失败率 < fail_rate 时 5 连败不开路) | CHS "连败必开"在高流量健康源上不再成立(防随机噪声误熔唯一好源);冷启动/低流量语义不变 |
 | 结构化重问缺省 1→2 | 解析失败时最多多一跳成本;`PGW_STRUCTURED_MAX_RETRIES=1` 可显式还原 |