fix(klaud): harden candidate dispatch and recovery / 加固 Klaud 候选调度与恢复

[adibarra]

Klaud was spending agent slots on candidates whose published baseline could not be reconstructed, repeating recent candidates, and leaving family claims behind when the Claude action returned success without a verifiable lifecycle outcome. Public image data also needs to remain benchmark evidence rather than authority for repository state or upstream compatibility.

This change makes current master configs define the pool and keeps upstream image research in the candidate agent. Planning now reviews a bounded batch, softly deprioritizes recent same-base candidates, resolves the public display model, and reconstructs the complete baseline before claiming or dispatching. Missing baselines and capacity losses can backfill from later reviewed candidates.

Every non-skipped agent exit now runs trusted reconciliation. No-PR/no-run sessions release immediately, active children remain owned, terminal sessions finish validation or cleanup, and sanitized diagnostics are uploaded before an unverifiable candidate fails. Public/private reads share bounded retries for transient network, 408, 429 and 5xx failures while semantic errors still fail immediately. Existing gzip decoding, image normalization and always-delete unsuccessful cleanup remain included.

Klaud performance changelog descriptions are now limited to one clear English sentence of at most 120 characters. They contain the engine version change and at most one essential compatibility adjustment; detailed evidence and results stay in attempt comments.

Validation:

• Ruff format and lint passed
• 21 focused tests passed
• local behavior checks passed for retry/gzip handling, non-retryable invalid JSON, review schema, cooldown ordering, baseline backfill, claim release, active-run preservation and sanitized diagnostics
• strict full-repository zizmor: no unsuppressed findings
• --no-ignores workflow scan: only the six documented existing exceptions
• actionlint: only its existing $/.github/... reusable-workflow parser limitation

中文

Klaud 曾将 agent 名额用于无法重建完整公开基线的候选，重复选择近期候选，并在 Claude action 返回成功但生命周期结果无法验证时遗留配置族认领。公开镜像数据仍只应作为 benchmark 证据，不能决定仓库当前状态或上游兼容性。

本变更改为由当前主配置定义候选池，并由候选 agent 负责上游镜像调查。规划阶段限制单次检查批次，对同一基准 SHA 的近期候选进行软降序，解析公开展示模型，并在认领或调度前重建完整基线。基线缺失或容量变化时，后续已检查候选可以补位。

每个未跳过的 agent 步骤结束后都会执行受信任的收尾。没有 PR/run 的会话立即释放，活动子运行继续保留归属，已结束会话完成验证或清理；无法验证的候选会先上传脱敏诊断，再使作业失败。公开和私有读取对临时网络故障及 HTTP 408、429、5xx 使用有界重试，语义错误仍立即失败。原有 gzip 解码、镜像规范化及失败后始终删除分支的清理逻辑继续保留。

Klaud 的性能变更日志描述现在限制为最多 120 个字符的一句简洁英文，只包含引擎版本变化和至多一项关键兼容性调整；详细证据和结果保留在尝试评论中。

验证：

• Ruff 格式和 lint 通过
• 21 项相关测试通过
• 本地行为检查覆盖重试/gzip、无效 JSON 不重试、检查 schema、冷却排序、基线补位、认领释放、活动运行保留及脱敏诊断
• 全仓库严格 zizmor：无未忽略发现
• 工作流 --no-ignores 扫描仅有六项已记录的现有例外
• actionlint 仅报告其对现有 $/.github/... 可复用工作流语法的解析限制