fix(klaud): wait for capacity instead of discarding passing smoke runs / fix(klaud): 等待容量而不是丢弃已通过的冒烟测试

[adibarra]

Since #4075, the largest loss is late capacity deferral. #4077, #4097, #4099 and #4107 all passed their smoke run on the new image. Each then failed one capacity snapshot at the final label step, so finish closed the PR and deleted the branch. The next wave redoes resolve and smoke from scratch, about $4 of agent time plus a GPU smoke run each time. #4078, #4083 and #4087 were thrown away the same way at a targeted dispatch after the PR existed.

Once the PR exists, a failed check now waits instead of deferring. check-capacity --wait polls every 60s for up to 9 minutes per call, under Claude Code's Bash limit. Each poll is charged to a per-candidate budget, KLAUD_CAPACITY_WAIT_MINUTES (120), stored in $KLAUD_EVIDENCE. The charge is written before each sleep, so a killed tool call still counts and the wait can't loop forever. Exit codes: 0 proceed, 1 call again, 2 budget spent, so defer. A failed dashboard read counts as unavailable for that poll rather than aborting.

Before the PR exists, deferral is still immediate, since nothing is lost. The 80% gate itself is unchanged: the CI scheduler doesn't rank Klaud sweeps below human sweeps, so the gate is what keeps Klaud on idle capacity.

Risk: waiting uses up the 6h agent job. If the final sweep is dispatched late and the job times out, recovery already validates a green exact-head PR sweep afterward (lifecycle.reconcile). A job that times out while still waiting ends as unexpected-error, which closes the PR just as a deferral does today.

中文

自 #4075 以来，最大的损失来自后期的容量推迟。#4077、#4097、#4099 和 #4107 都已在新镜像上通过冒烟测试，但随后在最终标签步骤中有一次容量快照未通过，于是 finish 关闭了 PR 并删除了分支。下一轮会从头重新解析并重新做冒烟测试，每次约花费 4 美元的代理时间外加一次 GPU 冒烟运行。#4078、#4083 和 #4087 也以同样方式在 PR 创建后的定向调度阶段被丢弃。

PR 创建后，容量检查失败会等待而不是推迟。check-capacity --wait 每 60 秒轮询一次，每次调用最多 9 分钟，低于 Claude Code 的 Bash 时限。每次轮询计入每个候选的预算 KLAUD_CAPACITY_WAIT_MINUTES（120 分钟），该预算保存在 $KLAUD_EVIDENCE 中。预算在每次休眠前写入，因此被终止的工具调用也会计入，等待不会无限循环。退出码：0 继续，1 再次调用，2 预算已用完，需要推迟。仪表盘读取失败只视为该次轮询不可用，而不会中止。

PR 创建之前仍然立即推迟，因为此时没有损失。80% 阈值本身不变：CI 调度器不会把 Klaud 的 sweep 排在人工 sweep 之后，因此该阈值是让 Klaud 只使用空闲容量的机制。

风险： 等待会消耗 6 小时的代理任务时间。如果最终 sweep 调度较晚且任务超时，恢复流程已经会在之后验证当前 head 上通过的 PR sweep（lifecycle.reconcile）。若任务在等待期间超时，结果为 unexpected-error，与现在的推迟一样会关闭 PR。

AI model disclosure

Prepared with anthropic/claude-opus-5-5, which analyzed the Klaud outcomes, wrote the code and tests, and drafted this description. No delegated agents were used.