bug(runtime): "nothing to fold" disables compaction for the turn

[Colafornia]

What happened

A single mid-turn compaction attempt that finds nothing safe to fold disables all compaction for the rest of the turn — including the reactive path that is supposed to recover from a real provider context_overflow. The turn can then die on input capacity even though a later fold would have succeeded.

Two different outcomes currently share one flag, state.summarizerFailure:

1. The summarizer ran and failed (provider error, malformed output, truncation, context overflow). Marking it here is correct — it exists to stop re-issuing a doomed call, per the History compaction always fails open on kimi-coding-plan/k3-256k: summarizer instruction as system prompt is ignored and breaks prefix cache #4634 incident (~15 repeated failing calls over ~47 minutes).
2. No summarizer call was ever made — selectSafeCompactionPrefix found no completed span that preserves tool call/result pairs, honors the head anchor, and leaves the reserved tail (no_safe_completed_span).

Outcome 2 is a property of the event pool at that step, not a failure. As the turn proceeds, tool results keep landing and a safe prefix can appear — but the flag short-circuits every later attempt before the ledger is even re-read:

• packages/runtime/src/ai-sdk-compaction.ts:977-983 — entry short-circuit on state.summarizerFailure.
• packages/runtime/src/ai-sdk-compaction.ts:1111-1125 — all fail-open reasons, including no_safe_completed_span, are written into the flag.
• packages/runtime/src/ai-sdk-compaction.ts:1205-1287 — recoverFromOverflowError routes through the same function, so a real provider rejection after the flag is set cannot be rescued either.

Expected: "nothing foldable right now" is re-evaluated as the ledger grows; only outcomes that actually invoked the summarizer suppress further attempts.

How to reproduce

In a live session this needs one unlucky timing: the proactive trigger fires on a step where the uncovered frontier cannot yield a safe span — e.g. fewer foldable events than the headAnchor + reserveTailEvents: 1 headroom allows, or an open tool pair at the tail. Verified deterministically by stubbing the ledger, summarizer, and projection seams (no provider involved):

1. Event pool = user anchor + one completed tool call/result pair. Under the mid-turn spec this leaves nothing foldable.
2. Call compactActiveRequestHistory: returns no_safe_completed_span; summarizer call count = 0; state.summarizerFailure is set.
3. Append a second completed tool pair — a safe prefix now exists — and call again: returns early on the flag; listRuntimeEvents is not even re-read; summarizer count stays 0.
4. Control: clear state.summarizerFailure, call again with the same pool — fold succeeds, summarizer count = 1.

Environment

• Repo HEAD: f58f30203 (upstream/main)
• Surface: runtime (@maka/runtime), all model routes

Suggested fix

Mark only summarizer-attributable failures; let structural outcomes lapse:

• no_safe_completed_span (and equivalent insufficient-headroom outcomes) must not write state.summarizerFailure — zero summarizer calls were made, and the outcome becomes obsolete as the pool grows.
• provider_error, malformed_*, output_length, empty_summary, input_too_large stay marked as today — the bounded-retry behavior from History compaction always fails open on kimi-coding-plan/k3-256k: summarizer instruction as system prompt is ignored and breaks prefix cache #4634 is unchanged.
• Ledger/durability read failures are a third category if a different retry policy is wanted; they are not summarizer failures either.

Acceptance tests (seams exist in mid-turn-capacity-backend.test.ts / overflow-reactive-recovery.test.ts):

• First attempt yields no_safe_completed_span with zero summarizer calls; after the pool gains a completed pair, the next attempt folds successfully.
• Same setup, then a real provider context_overflow: reactive recovery folds instead of being blocked.
• provider_error / malformed-summary outcomes still mark the turn as failed; summarizer call counts stay capped as existing tests assert.

For reference, DeepSeek-Reasonix keys its compaction "stuck" state on an input hash and resets it when new input arrives (internal/agent/context_manager.go) — the same principle: a grown pool may present a new foldable boundary.

中文版本

发生了什么

只要有一次轮中压缩尝试发现当前没有可安全折叠的区间，本轮剩余的所有压缩都会被禁用——包括本应处理 provider 真实 context_overflow 的反应式恢复路径。即使之后折叠本来能够成功，这个 turn 仍会因输入超限而终止。

当前两类不同的结果共用同一个失败标记 state.summarizerFailure：

1. 摘要器确实被调用并且失败了（provider 错误、畸形输出、截断、上下文超限）。这里置标记是对的——它的目的就是阻止重复发起注定失败的调用，即 History compaction always fails open on kimi-coding-plan/k3-256k: summarizer instruction as system prompt is ignored and breaks prefix cache #4634 事故（约 47 分钟内重复约 15 次失败调用）的修复。
2. 根本没有发起摘要调用——selectSafeCompactionPrefix 找不到既完整、又保住工具调用/结果对、又满足 head anchor 和尾部保留的跨度（no_safe_completed_span）。

第 2 类结果是当前这一步事件池的属性，不是失败。随着 turn 推进、工具结果持续落账，安全前缀可能出现——但失败标记让之后的每次尝试在重读账本之前就被短路：

• packages/runtime/src/ai-sdk-compaction.ts:977-983 —— state.summarizerFailure 置位后入口直接短路；
• packages/runtime/src/ai-sdk-compaction.ts:1111-1125 —— 所有 fail-open 原因（含 no_safe_completed_span）都写入该标记；
• packages/runtime/src/ai-sdk-compaction.ts:1205-1287 —— recoverFromOverflowError 走同一个函数，标记置位后真实的 provider 拒绝也无法被抢救。

预期行为："此刻没有可折叠内容"应随账本增长被重新评估；只有真正调用过摘要器的结果才应抑制后续尝试。

如何复现

真实会话中需要一个不巧的时机：主动触发器在某一步 firing，而当时的未覆盖头部恰好选不出安全跨度——例如可折叠事件少于 headAnchor + reserveTailEvents: 1 所需余量，或尾部卡在未闭合的工具调用对中间。已通过注入账本/摘要器/投影桩做了确定性复现（不经过真实 provider）：

1. 事件池 = 用户锚点 + 一对完整的工具调用/结果。按轮中折叠规则没有可折叠内容。
2. 调用 compactActiveRequestHistory：返回 no_safe_completed_span；摘要调用次数 = 0；state.summarizerFailure 被置位。
3. 追加第二对完整工具调用——安全前缀此时已存在——再次调用：因失败标记提前返回；连 listRuntimeEvents 都未重读；摘要调用次数仍为 0。
4. 对照：清除 state.summarizerFailure 后用同一事件池再调——折叠成功，摘要调用次数 = 1。

环境

• 仓库 HEAD：f58f30203（upstream/main）
• 范围：runtime（@maka/runtime），与模型路由无关

建议修复

只对可归因于摘要器的失败置标记，结构性结果不置：

• no_safe_completed_span（及等价的头部空间不足类结果）不得写入 state.summarizerFailure——它没有发起任何摘要调用，且随事件池增长会失效；
• provider_error、malformed_*、output_length、empty_summary、input_too_large 维持现状——History compaction always fails open on kimi-coding-plan/k3-256k: summarizer instruction as system prompt is ignored and breaks prefix cache #4634 引入的有界重试行为不变；
• 账本/持久化读取失败是第三类，如需不同重试策略可单独处理——它们同样不是摘要器失败。

验收测试（mid-turn-capacity-backend.test.ts / overflow-reactive-recovery.test.ts 已有可用桩点）：

• 首次返回 no_safe_completed_span 且摘要调用为零；事件池新增一对完整工具调用后，下次尝试折叠成功；
• 同样前置下，provider 真实 context_overflow 能被反应式恢复折叠，而不是被失败标记挡下；
• provider_error / 畸形摘要结果仍在本轮置标记，摘要调用次数上限与现有测试断言一致。

参考：DeepSeek-Reasonix 将压缩 stuck 状态按输入哈希记录，新输入到来即重置（internal/agent/context_manager.go）——同一原则：增长后的事件池可能出现新的可折叠边界。

[Colafornia]

/take