[ranxianglei]

Problem

A session with a persisted matching-origin calibration factor k-hat > 1 can hit a non-retryable local 502 even though folding would have brought it back under the window. Two symptoms, one root cause — a unit mismatch between two token scales:

1. The raw-range futility gate compared an uncalibrated foldable pool (kernel range.tokens) against a calibrated payload deficit (result.payloadEstimate - limit). With k-hat>1 the deficit is inflated x-k while the pool is not, so the gate declared a recoverable round "futile" and skipped every fold -> the over-window payload was never forwarded and tier recovery never got a chance.
2. The post-fold fail-fast message quoted a raw rebuilt size while the fit decision used a calibrated size, so it claimed a sub-window number (~550759) "exceeds the model window 872000".

Root cause

Once a matching-origin k-hat is set (#1933 F1), applyEstimateCalibration scales the host's own estimate (payloadEstimate, the fit check, the pre-fold tokenCount) by k-hat, but the kernel's range.tokens pool stays on the raw count. The gate's two operands were on different scales.

Fix

• src/preflight.ts: route the foldable pool through applyEstimateCalibration before comparing it to the deficit -- pre-walk gate, relaxed probe, mid-walk bail -- and in the log/failure text, so both sides share one scale. Folding a raw-mass range removes its mass from the text, so its calibrated saving is bounded by mass*k-hat; comparing mass*k-hat * slack < deficit is the unit-unified bound. Identity for k-hat=1 / absent / mismatched-origin keeps the 模型压缩输出是输入的数倍 #1841 uncalibrated path byte-identical.
• src/server.ts: hoist the calibrated post-fold size so the fail-fast message quotes the same scale as the fit decision.

Behavior change (intended)

Before	After
Calibrated (k-hat>1) session whose raw pool x slack < calibrated deficit -> skipped as "futile", bricked behind a non-retryable 502	Same session now proceeds to fold; recovers when the fold closes the gap
Fail-fast message quoted the raw post-fold size (could claim a sub-window number exceeds the window)	Quotes the calibrated post-fold size, same scale as the fit decision

The uncalibrated path (k-hat=1 / absent / mismatched origin) is unchanged -- applyEstimateCalibration returns the raw value there.

Config surface

None. No new config field, env var, CLI flag, wire shape, or persistence format.

Tests

New tests/issue3645-calibrated-futility.test.ts:

• calibrated case (k-hat=2, matching origin): the walk now proceeds and forwards instead of skipping as futile;
• uncalibrated control: futility skip unchanged (模型压缩输出是输入的数倍 #1841 preserved).

Existing issue1841-preflight-futility and preflight-round-budget suites stay green locally. Full-suite verdict = remote CI.

Fixes #3645

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-18_preflight-calibrated-futility-gate (8ede209)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.4299-pr.3650.36383f7

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr3650.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

查重: 不重复。 按失败签名(预检 futility gate 的未校准池 vs 校准 deficit 单位错配)检索 open/closed issue 与近期 PR:唯一同主题条目即本 PR 所修的 #3645 本身;相邻的 #3649(tier-recovery candidate search)机制不同(tier 恢复的候选搜索 vs raw-range gate 的单位),不构成重复。

开始独立审核(不照抄 issue 结论):已拉取分支 pr3650-review @ 99ab9c4,diff = src/preflight.ts(+16/-9)、src/server.ts(+7/-5)、新增 tests/issue3645-calibrated-futility.test.ts(222 行)。master 在 PR 之后前进了 4 个提交(#3618 dsh attach origin、#2950 persona-lineage ref guard),其中 #2950 触碰过 src/preflight.ts / src/server.ts 区域——正在核对热文件语义冲突、既有行为回归清单与测试有效性,随后按八问给结论。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Review 结论:可合(人工合并:https://github.com/ranxianglei/billion-context/pull/3650)——方向、机制、回归面已独立复核,并在 PR 分支补了两个 follow-up commit(见下)。同一结论也已回帖到原 bug 报告线程(#3645)。

结构化八问

1. 是什么问题:带校准系数 k̂>1 的会话,负载超窗且明明还有可压内容时,bili 误判「压了也没用」整轮跳过压缩,返回不可重试的本地 502;报错还称「约 55 万 token 超过 87.2 万窗口」——一个低于窗口的数字被说成超窗,自相矛盾。
2. 什么原因导致的:DSH: 切换模型后，preflight 在窗口远未满时提前触发压缩 #1933 F1 起,matching-origin 的 k̂ 存在时,宿主侧估计(payloadEstimate、fit 判定、pre-fold tokenCount)都按 k̂ 放大,但 kernel 的可压池 range.tokens 仍是原始计数。futility 门的两操作数不同标度(原始池 vs 校准后缺口),k̂>1 时缺口被放大 x-k 而池没有 → 可恢复轮被判「徒劳」;post-fold 报错文案同样混用了原始与校准两个数。
3. 影响范围与严重程度:仅 matching-origin k̂>1、且「原始可压池 × 1.2 < 校准后缺口」的会话(高校准系数、贴窗/超窗场景,[Bug][preflight] Calibrated deficit vs uncalibrated foldable mass prematurely skips raw folds (0.1.199) #3645 的 DSH native 即此)。严重度:完全不可用(不可重试 502,只能手动缩上下文)。其余会话零影响。
4. 怎么修复的:统一量纲——可压池在比较前走同一个 applyEstimateCalibration(前置门、relaxed probe、mid-walk 中止、日志/报错文案四处决策点全过一遍);把校准后的 post-fold 尺寸提升出来,让报错文案与 fit 判定引用同一标度。折掉原始质量 M 至多省 M×k̂ 校准 token,故 M×k̂×slack < deficit 是单位统一的界。k̂=1/缺失/origin 不匹配时 applyEstimateCalibration 原值返回,模型压缩输出是输入的数倍 #1841 路径不变。
5. 是否有回归问题:无。触碰路径既有行为逐项清点:未校准 futility skip(issue1841 A/B 两例)绿;relaxed probe 与轮次预算(preflight-round-budget 2 例)绿;status 映射(429→503 / deterministic 超窗→413 / 其余 502)未动;meter anchor 赋值(preflight: one failed summarization call fails the whole turn with a non-retryable 502 and blocks the session #1987)未动;tier funnel 入口条件未动;wire 形状零变化(转发 payload/SSE/tool calls 均未动,只动本地决策与本地文案)。
6. 是否破坏缓存:不涉及——前缀缓存、重试策略、增量游标、去重键、持久化格式均未触碰。
7. 是否改了设计方向:不涉及——在既有校准设计内做单位统一,无架构转向。
8. 是否引入新配置:不涉及——无新 env/config/CLI/wire/持久化面(diff 核实)。

Review 发现并已直接修掉(两个 additive commit,已推到 PR 分支)

• 日志混标度(同一机制族):preflight compressed N range(s) info 日志原先混用校准前折叠数和原始 post-fold anchor,k̂>1 会话里 (A → B) 差值对不上报告的节省量 → 改引校准后 post-fold 尺寸(commit 7351fce8b)。行为变更披露:k̂≠1 会话该 info 日志第二个数从原始 anchor 变为校准值(纯日志,无功能影响)。
• 症状 2 缺回归测试:新增 e2e case(commit 8ede209a5)——k̂=2 + summary 比源段更重,使压缩轮确定性耗尽(post-fold 原始质量低于窗口、校准质量仍超窗),断言报错引用的是超窗的校准数。变异验证过:把表达式临时换回旧写法,测试立即失败,失败输出正是自相矛盾的 sub-window 引用(约 21556 < 窗口 30000)——测试有牙。

一处 PR 描述需修正的细枝:PR 称「未校准路径不变」,严格说 fail-fast 引用的数从 lastInputTokens(棘轮 floor、不含图片)换成了「重建文本+wire 开销+图片」,有两个角落差异——anchor 陈旧偏高时、含图会话(差 imageTokens)。两处都让引用值更接近实际发出的大小,属同族改进而非回归,但按披露义务写明:旧值→新值如上,原因=引用「实际要发出去的」而不是 meter 的历史 floor。

验证证据:定向套件本地 8/8 通过;typecheck 干净;远程 CI 在新 head 8ede209a5 上全绿(全部 static gates + ubuntu/windows × node22/24 test shards,零失败)。与当前 master dry-run merge 无冲突;PR 显示 behind 是因 master 期间合入了 #3618/#2950,不构成合并障碍。

判定:可合,请人工合并。

[github-actions]

📦 Merged Build Published

This merge to master was published as 0.0.639-master.80cbe28 (rolling master tag).

npm install -g billion-context@0.0.639-master.80cbe28  # this exact build
npm install -g billion-context@master  # always the newest merged build

npm view billion-context@master prints the current version at any time.