[ranxianglei]

Closes #5500

Summary

The reported failure loop (502 context ~204011 tokens exceeds the model window 150000 ... with 0 compressible range(s) still visible) was TWO independent defects compounding:

Defect A — window lookup miss. CONTEXT_LIMIT_TABLE had no step-family row, so step-5-preview-free (real window 1048576 per the reporter's deployment config + models registry) fell to the base.modelContextLimit env default 2e5 → ~150K effective window. Payloads between their true size and that false wall tripped the preflight trigger and died in the fail-fast loop although upstream would have accepted them. Per the #852 doctrine too-large guesses self-heal on the first upstream overflow; too-small ones strand — this was the stranding side.

Defect B — char-bound estimate poisoning. The #604/#1498-F1 failure arm persists an ESTIMATE-grade lastInputTokens at char-bound scale (estimateCoreMessagesUpper ≈ 4x the token count) into stats_json, replayed into every later request. Every reader of that value is already grade-disciplined (#1839 kill-at-reader: effectiveTokenCount branch 1, statusInputBaseline #2029, baselineFloorRaw, noOverflowEvidence, the preflight per-round floor #1492) — EXCEPT the preflight ENTRY reads: the currentTokens init and the entryTokens floor fed the #1933-F3 overshoot budgets from the poisoned arm on every request after restart. This applies the per-round floor's exact discipline to those two entry reads.

Changes

• src/config.ts: { match: /^step-/i, limit: 1_000_000 } row (field evidence cited inline).
• src/preflight.ts: anchorGradeOk gate on the two entry reads, mirroring baselineFloor at preflight.ts:2144. usage/overflow-arm anchors still floor; the unknown-baseline (fork 会话 lastInputTokens=0 时 preflight 退化为字符估算,大载荷被低估后不压缩直接裸转上游 #553) regime is untouched; the arm emergency shrink on upstream failure to break relay-5xx deadlock #604 kernel truncate-rescue consumer (providerInputTokens in prepare-anthropic) is deliberately untouched.
• tests/issue5500-char-bound-poison.test.ts: table row + negative controls; entry-budget discipline asserted BOTH directions (Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617); e2e reproducing the issue shape through the REAL window-resolution chain (pre-fix: lookup miss → ~150K wall → deterministic exhaustion → fail-fast ≥400; post-fix: 1M native window → clean forward, zero summarization calls); control pinning that a MEASURED usage-grade overflow is still rejected.

Behavior changes (old → new)

1. Effective window for ^step- models without an explicit per-route/per-model override: ~150000 → ≈786432 (the new table row resolves native 1_048_576 — the field-enforced wall cited in localInputEstimate persists the char-count upper bound (~4x char/token) and poisons every later request; with a small window preflight then reports "exceeds the model window" on a 2.4x-overestimate #5500, same shape as the gemini-3 row; the per-request output-headroom reservation subtracts min(maxOutput, 25%×window) exactly as before, so ≈786432 when the 25% cap binds and closer to 1M for small maxOutput requests). Explicit configs still win (precedence chain unchanged).
2. Preflight entry sizing under an ESTIMATE-grade/unprovenanced baseline: the poisoned arm no longer floors entryTokens/currentTokens init — entry sizes on the per-request measurement instead. usage/overflow-arm anchors unchanged; the arm emergency shrink on upstream failure to break relay-5xx deadlock #604 kernel truncate rescue unchanged; fitting payloads still forward even when the summary lane 429s (fix: honor anthropic-beta context-1m to size the context window #301/The symptom is that switching models causes a hang, suspected to be triggered by dsh native compression. | 症状是切模型卡住疑似触发 dsh 原生压缩 #4288 preserved, now explicitly pinned).
3. No wire-format, persistence-format, config-surface, or kernel changes.

Notes

Round 2 — remaining ungraded readers (folded in from review of this PR)

A class sweep after review found three more stats.lastInputTokens reads without the same grade discipline — the same mechanism family as Defect B, fixed here rather than split out (one mechanism, one PR):

• src/request-membership.ts (Bypass requests overwrite the main session's usage anchor, causing long sessions to permanently stop compressing. | 旁路请求覆盖主会话 usage 锚点，导致长会话永久停止压缩 #4805 membership probe): the needle’s main-scale bar now refuses estimate-grade/unprovenanced scalars and falls back to the needle’s own per-request scale. Honest note: the incoming payloadUpper rides the SAME char-bound caliber (localInputEstimate) today, so the ratio self-normalizes — this gate is drift-robustness, not a fix of a live inflation.
• src/server/prepare-responses.ts (codex compact-intercept minted usage): the minted input_tokens no longer carries a char-bound arm — est is chars/4 while a poisoned scalar is char-count scale, so the scales do NOT cancel here; codex recomputes its ledger from the minted body, so an inflated mint would skew it until the next real report.
• src/session.ts noteAbortedTurn (In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293/In 0.1.210, sessions with frequent interruptions barely trigger compression reminders — tokens accumulate to 330k before | 中断频繁的会话在 0.1.210 里压缩提醒几乎不触发,积到 33 万 token 才压一次 #5415): untrustworthy meters are skipped when stamping the abort re-anchor flag. A char-bound arm would delay the In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293 re-anchor ~4x and make In 0.1.210, sessions with frequent interruptions barely trigger compression reminders — tokens accumulate to 330k before | 中断频繁的会话在 0.1.210 里压缩提醒几乎不触发,积到 33 万 token 才压一次 #5415 bank PHANTOM growth debt (anchor − reference at char-bound scale) into the nudge cadence.

Additional behavior changes (old → new):

4. Bypass requests overwrite the main session's usage anchor, causing long sessions to permanently stop compressing. | 旁路请求覆盖主会话 usage 锚点，导致长会话永久停止压缩 #4805 membership needle bar: reference scale = the graded scalar (usage/overflow-arm) when present, else the needle’s own per-request scale (was: any positive scalar wins). Legacy unprovenanced sessions get the needle-scale fallback — needle movement can be slightly LESS conservative on those sessions only.
5. Codex compact-intercept minted usage: input_tokens = max(graded scalar, raw-body est, 1) (was max(any scalar, est, 1)). Narrow path: codex host + BILI_CODEX_COMPACT=intercept + compact-form request + healthy ACP state.
6. Abort re-anchor stamp: stamped only from usage/overflow-arm meters (was: any positive finite meter). Sessions whose meter is estimate-grade-only lose the In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293 absorb on aborts — the cadence falls back to plain growth-based nudging there (the arm was false information anyway).

Round-2 tests: T6a/b/c (membership: poisoned / graded-wins / legacy-fallback), T7 + T7-control (forged usage e2e through the real intercept path: poisoned → raw-body estimate; usage-grade 12K → exactly 12K), T8 (re-anchor: skip-on-poison + stamp-on-usage with the #5415 debt arithmetic intact). issue5415/issue5293 fixtures now declare their usage-grade provenance.

Notes

• Coordinates with open PR [PR] fix: auxiliary requests must not arm or clobber the session meter (#4974) #5005 (side-lane baseline arming): same meter-arming domain, complementary direction — [PR] fix: auxiliary requests must not arm or clobber the session meter (#4974) #5005 gates WHO may arm; this change gates what an armed estimate-grade value may INFLUENCE at the preflight entry. This PR touches only config.ts + preflight.ts (+ test); no overlap with [PR] fix: auxiliary requests must not arm or clobber the session meter (#4974) #5005's files.
• Local verification: targeted suites green (issue5500 11/11; issue5415 + issue4805 + codex-compact + codex-compact-e2e 51/51; issue5293 2/2 incl. its real-pipe e2e), typecheck clean. Full-suite verdict = CI on this head.
• The deeper question (lookup-miss fallback: treat unknown as unknown instead of the 2e5 default) is a product decision — filed separately as a decision-request issue.

Round 3 — v0.1.212 field follow-up: cycle-stale × destructive emergency-truncate band (commit 8222bc4)

A field report on v0.1.212 (Windows / Pi / Responses lane) showed the CYCLE-STALE demotion (#4383 — last send's usage report never settled) handing the kernel meter the char-bound estimate (min(localInputEstimate, raw upper)), inflating it ~2.4–4x over the measured size. On a session whose measured usage sat safely under the window, that inflation crossed the kernel's destructive emergency-truncate band (truncate.threshold = 0.95 × window) — false clips then LATCH into the delivered view until the next successful fold (#3939 ③). The freshness demotion itself is intended (#4383); the defect was the SCALE of the ride-on scalar. Verified line-by-line against both v0.1.212 and master before fixing.

Fix C6 (src/server.ts effectiveTokenCount): cycle-stale now rides contextEstimateTokens — the billing-caliber per-send estimate stamped at the SAME relay chokepoint (k̂-scaled where route+model provenance matches), i.e. the same replacement reading staleSmallAnchorFallback already returns for its own demotion class — capped by min(…, current raw upper) so it can never exceed today's fail-closed worst case. This preserves #4383's intent (per-send evidence outranks a frozen anchor: 150K ≠ the frozen 4733) while removing the 2–4x inflation.

Additional behavior change (old → new):

7. Kernel meter feed during no-report cycles: min(contextEstimateTokens, current raw upper) / "estimate" (was: min(localInputEstimate, raw upper) / "estimate" — char-bound scale). Affects the nudge tiers, the EMERGENCY truncate band, the terminal escape and the absorb view — all of which read this one value. UNCHANGED: usage/overflow-arm anchor branches, fresh-settle ([Bug] statusInputBaseline falls back to lastUsageGradeTokens instead of the calibrated local estimate — so one 4.7k side-request sample blocks folding for 4 hours | #1595 承诺的"改用本地估算"兜底在 statusInputBaseline 里没有实现，一次 4.7k 侧请求样本即可挡住折叠 4 小时 #4383 regression guard), the [Bug] 上游从不报告 usage 时 nudge 永不触发——上下文无界增长直到超窗（ChatGPT 登录态后端） #728 anchor-less never-reporting lane (shared fall-through untouched), the host preflight trigger (does not consume this value), and statusInputBaseline display reads. Three [Bug] statusInputBaseline falls back to lastUsageGradeTokens instead of the calibrated local estimate — so one 4.7k side-request sample blocks folding for 4 hours | #1595 承诺的"改用本地估算"兜底在 statusInputBaseline 里没有实现，一次 4.7k 侧请求样本即可挡住折叠 4 小时 #4383/[Bug] Pre-deduction count after T1 compression is overwritten by a stale usage rollback, immediately followed by a false | [Bug] T1 压缩后的预扣计数被旧 usage 回退覆盖，紧接着误触发 T2 #5134 regression assertions re-pinned from the char bound to the calibrated scalar (same chokepoint freshness, corrected scale — one test's own message already claimed "its calibrated estimate wins"; it is now literally true).

Kernel-side residuals — NOT in this PR (require owner approval per Kernel Boundary rules 1+3, filed as #5693, p1): (a) ProcessTurnInput carries no evidence grade, so the destructive node could still act on a mis-calibrated estimate (proposed: optional tokenCountSource, refuse clipping when an unmeasured reading is contradicted by a sub-threshold usage-grade anchor); (b) truncateLargeToolOutputs ignores protected tools entirely — no isMessageProtected, no ALWAYS_PROTECTED_TOOLS (compress/acp_rule), no config.protectedTools — a Kernel Boundary rule 6 ("protected-tool hard exclusion") violation even at genuine overflow. Same node, same failure domain as the field damage → one checklist in that request. Cross-links: #5217/#5255 (same-node siblings, distinct mechanism), #3691 (latch design origin — why false positives are expensive).

Round-3 tests: C6a/C6b (ride-on both directions + four unchanged calibers: fresh-settle / #728 anchor-less / legacy absent-field / k̂ overshoot cap), C7 e2e reproducing the exact field chain (usage settle → no-report resend → repeat prepare; teeth-proven: with the fix reverted the clips latch into the delivered body, with it applied the session stays unclipped), C7-control (the kernel band itself still fires on a direct over-limit reading through the real pipeline node).

Local verification (this head): targeted suites green — issue5500 15/15; effectiveTokenCount consumer batch (issue4383×2, silent-backend-nudge, fork-nudge-deadzone/trigger, issue1839, issue1492, issue1820, issue2033, issue2029, issue5134, first-sight-drain-wire) 50/50; neighbor batch (issue4288, preflight-fail-fast, issue1933, issue5415, issue5293, issue4805, codex-compact + codex-compact-e2e, wire-contract) 96/96; typecheck clean. Full-suite verdict = CI on this head.

[github-actions]

📦 Built Package Artifact

Branch: 2026-11-02_step-window-char-bound-poison (56a7fe2)

Latest build — install this exact version

npm install -g billion-context@0.0.9153-pr.5501.224f824

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

Recent builds (commit → version)

commit	version	build
56a7fe2	billion-context@0.0.9153-pr.5501.224f824	build
8dede7d	billion-context@0.0.9024-pr.5501.5c3ea23	build
8222bc4	billion-context@0.0.8174-pr.5501.9d91d4a	build
b263268	billion-context@0.0.7814-pr.5501.1efbc31	build
ac098ee	billion-context@0.0.7812-pr.5501.2cccdad	build
b5e1ee7	billion-context@0.0.7811-pr.5501.fb89109	build

Download artifact

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

tar xzf billion-context-pr5501.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。 本 PR 关闭 #5500（canonical 缺陷报告）。相关但不同：#5005（同域互补——#5005 管「谁能 arm」，本 PR 管「已 arm 的 estimate-grade 值能在 preflight entry 影响什么」，无文件重叠）；#5502（lookup-miss fallback 产品决策，已正确另立 decision-request）。前序同类均已闭环：#1839/#1846、#2029、#1492/#1503。无其他在途 PR 加 step- 表行或修 char-bound entry poisoning。

标签： bug · omp · scope/wiring · p1 —— p1 可用性缺陷（会话硬 fail-fast 死循环）、omp ACP 车道症状、宿主侧请求管道/窗口解析域。源 bug #5500 已带同样四标；本镜像线程的逐-issue 标签端点从当前环境不可达（404），为避免误打未强制写入，仅在此报备分类。

CI： 镜像缓存 30 check，failed=0，pending=8——全在 test 分片（ubuntu/windows × node22/24）跑；static-gates/kernel-guard/docs-sync/build-artifact/gate/mysql-contract 已绿。基座 = 当前 master head（0 behind），无 stale-base 风险。全量判定等 test 分片落定。

结构化八问

1. 是什么问题：#5500 失败循环——omp/step-5 会话报 502 context ~204011 tokens exceeds the model window 150000 ... with 0 compressible range(s)，非可重试、跨会话反复。两个独立缺陷叠加：(A) 窗口解析漏 step-* 行 → 落 base 默认 2e5 → 有效窗口被钉死 150K；(B) char-bound 的 ESTIMATE-grade lastInputTokens（≈4× token 数）被持久化进 stats_json 并回放进每个后续请求。

2. 什么原因导致的：(A) CONTEXT_LIMIT_TABLE 缺 step-* 行，step-5-preview-free（真实窗口 1048576）落到 env 默认 2e5 × .75 = 150K 假墙；按 #852 教义猜太大自愈、猜太小搁浅，这是搁浅侧。(B) 该值的所有读取方早已被 grade 纪律约束（#1839 kill-at-reader），唯独 preflight 的两个 ENTRY 读（currentTokens 初始化 + entryTokens floor）没约束——重启后每次请求都用中毒臂喂 #1933-F3 超支预算。根因定位准确：不是持久化层的问题，是这两个 entry 读漏了 grade gate。

3. 影响范围与严重程度：任何把 step-* 模型路由过来且未显式配 per-route/per-model 窗口的宿主（被 150K 假墙搁浅）＋ 任何 stats_json 带 estimate-grade 中毒臂的会话（报告者 omp/step-5；fleet 53/2420 会话 estimate>2× real）。p1：会话完全不可用——确定性耗尽 → 硬 fail-fast ≥400，无法恢复。

4. 怎么修复的：(A) config.ts 加 { match: /^step-/i, limit: 1_000_000 } 表行（字段证据 inline 引用）；(B) preflight.ts 给两个 entry 读加 anchorGradeOk gate（lastInputTokensSource ∈ {usage, overflow-arm}），逐字复刻 baselineFloor（preflight.ts:2081-2083）已有纪律——usage/overflow-arm 锚照常 floor，estimate-grade 臂退到 per-request 实测。走 kill-at-reader 而非改持久化，方向对。

5. 是否有回归问题：逐项过触碰路径既有行为——

• 空消息分支：旧 unknownBaseline→0，新→lastInputTokens；但 baselineKnown=false ⟺ server.ts:3466 lastInputTokens≤0，实际为 null，无漂移。
• usage/overflow-arm 锚：不变（anchorGradeOk=true 路径原样）。
• unknown-baseline 制度（fork 会话 lastInputTokens=0 时 preflight 退化为字符估算,大载荷被低估后不压缩直接裸转上游 #553）：不动。
• arm emergency shrink on upstream failure to break relay-5xx deadlock #604 kernel truncate-rescue（prepare-anthropic providerInputTokens）：刻意不动。
• fitting payload 仍转发（fix: honor anthropic-beta context-1m to size the context window #301/The symptom is that switching models causes a hang, suspected to be triggered by dsh native compression. | 症状是切模型卡住疑似触发 dsh 原生压缩 #4288）：保留并被 T2a/T3 显式钉住。
• provider-measured 墙仍拒：T4 control 钉住（source=usage 60K>40K → ≥400）。
• 未发现破坏既有行为。

6. 是否破坏缓存：不涉及。不碰 prefix-cache 接缝、重试、去重键、增量游标；preflight sizing 只影响「何时触发压缩」，不改出站字节/cache 前缀。

7. 是否改了设计方向：无新架构。是既有 #1839 kill-at-reader grade 纪律向剩余两个未约束 entry 读的补全；更深的 lookup-miss-fallback 明确留给 #5502 决策，不在本 PR。

8. 是否引入新配置：无新 env/config/CLI。仅一条内置表行（运行时数据，非 config-schema 字段）＋内部逻辑。显式 per-route/per-model override 仍优先（precedence chain 不变）。

行为变更披露确认（§Deliverables #4）

PR 正文已披露两条，复核一致：

1. ^step- 且无显式覆盖的模型有效窗口：150000 → ≈750000–786432（见下方 nit fix: keep system prompt byte-stable for prefix cache #1）。
2. estimate-grade/unprovenanced baseline 下 preflight entry sizing：中毒臂不再 floor entryTokens/currentTokens init，改按 per-request 实测。

非阻塞发现（建议，不拦合）

1. 常量 vs 正文数字不一致：代码设 limit: 1_000_000（×.75=有效 750000），但正文写「≈786432」（=1048576×.75）。二者都安全（自愈合方向，都 >>150K 墙）。建议二选一：把常量改成 1_048_576（同时对齐字段实测真实窗口、gemini-3 同表行 config.ts:608、正文数字三者一致，需同步改 T1 三处断言），或把正文 786432 改成 750000。我倾向前者——让代码精确等于字段证据的真实窗口。纯精度/一致性、零功能差异；因属你 PR 里声明窗口的取值，我不擅自改，你点头即可顺手 push。
2. 「所有其它 reader 都已约束」表述过强：实际还有两处读 lastInputTokens 未查 source——src/request-membership.ts:67（Bypass requests overwrite the main session's usage anchor, causing long sessions to permanently stop compressing. | 旁路请求覆盖主会话 usage 锚点，导致长会话永久停止压缩 #4805 membership needle 启发式，中毒时 referenceScale 放大约 4× → mainScale 阈值偏高，软降级，不同子系统）和 src/server/prepare-responses.ts:585(codex compact-intercept 伪造 usage body，中毒时伪造 inputTokens 偏大，窄 codex 路径)。两处都不喂 preflight entry sizing、都不引发 localInputEstimate persists the char-count upper bound (~4x char/token) and poisons every later request; with a small window preflight then reports "exceeds the model window" on a 2.4x-overestimate #5500 循环，故不在本 PR 范围；但属同一机制家族，建议并入 localInputEstimate persists the char-count upper bound (~4x char/token) and poisons every later request; with a small window preflight then reports "exceeds the model window" on a 2.4x-overestimate #5500 跟踪清单，不单开 issue。

测试有效性

高质量、非假测：T1 单元钉死窗口解析（basename-first 匹配 + 三负控）；T2a/T2b 双向断言 entry 预算纪律（estimate 拒 / usage 放行，#3617）；T3 e2e 走真实窗口解析链复现 issue 形状（pre-fix 确定耗尽→fail-fast≥400 / post-fix 干净转发 0 summary 调用）；T4 e2e control 钉住 measured 墙仍拒。断言有牙，bug 回归能抓住。

判定

可合（需人工 merge：step- 默认窗口属 fallback/default-value 变更，产品决策面，§7.4 必须 human）。前置条件：① CI test 分片转绿（当前 0 failed / 8 pending）；② 发现 #1 二选一定论（建议 1_048_576）。发现 #2 为非阻塞 follow-up（并入 #5500）。方向正确、根因定位准、回归面干净、diff 整洁（3 文件每行 on-purpose）、测试扎实。

[ranxianglei]

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

判定：APPROVE（可合；需人工 merge）
（GitHub 禁止同账号 approve 自己的 PR，故以 COMMENT 事件提交本审核，判定等同 APPROVE。）

本 PR 修 #5500 的两个叠加缺陷：(A) 窗口查找表缺 step-* 行 → 落 env 默认 2e5 → 150K 假墙；(B) #604 失败臂持久化的 char-bound ESTIMATE-grade lastInputTokens（≈4× token 数）被回放进每个后续请求，而 preflight 的两个 ENTRY 读恰好没做 grade gate。两处修法方向都对：补表行（在 miss 处止血）＋给 entry 读补上与 baselineFloor 相同的 anchorGradeOk 纪律（kill-at-reader）。本轮为独立二审（mirror 线程已有一轮 review 报告；我重读了 diff、独立过了回归清单，并补做了变异验证）：

变异验证（本地定向，非全量）

• revert src/config.ts 到 master → T1+T3 失败（表行缺失 → e2e 复现 issue 形状：确定性耗尽 → fail-fast ≥400）；
• revert src/preflight.ts 到 master → C2a 失败（estimate 臂退回 floor entry 预算）；
• 均恢复后原样；新 head 上定向套件 5/5 绿。

结构化八问

1. 是什么问题：omp/step-5 会话反复命中不可重试的 502 context ~204011 exceeds window 150000 ... 0 compressible range(s)——会话硬失败、无法恢复。

2. 什么原因导致的：两个独立缺陷叠加。(A) CONTEXT_LIMIT_TABLE 无 step-* 行，step-5-preview-free（真实墙 1048576）落到 base 默认 2e5 × .75 = 150K 有效窗口；按 #852 教义猜太大自愈、猜太小搁浅，这是搁浅侧。(B) 该 estimate-grade 值的所有读取方早已被 grade 纪律约束（#1839 kill-at-reader），唯独 preflight 的两个 ENTRY 读（currentTokens 初始化 + entryTokens floor，喂 #1933-F3 超支预算）漏了 gate——重启后每次请求都用中毒臂定 entry 尺寸。根因定位准确：不是持久化层的问题，是这两个读漏了 gate。

3. 影响范围与严重程度：p1，两面——①任何把 step-* 模型路由过来且未显式配窗口的宿主（被 150K 假墙搁浅）；②任何 stats_json 带 estimate-grade 中毒臂的会话（报告者 omp/step-5；fleet 实测 53/2420 会话 estimate>2×real）。触发即完全不可用：确定性耗尽 → 硬 fail-fast ≥400。

4. 怎么修复的：(A) 表加 { match: /^step-/i, limit: 1_048_576 }（字段证据 inline 引用；basename-first 匹配覆盖 relay 前缀 id）；(B) anchorGradeOk = source ∈ {usage, overflow-arm} 门住两个 entry 读——逐字复刻 preflight.ts:2145 baselineFloor 既有纪律；usage/overflow-arm 锚照常 floor，estimate 臂退到 per-request 实测。走 kill-at-reader 而非改持久化，方向对。

5. 是否有回归问题：逐项过触碰路径既有行为——usage/overflow-arm 锚不变；unknown-baseline 制度（#553）不动；#604 kernel truncate-rescue（prepare-anthropic providerInputTokens）刻意不动；fitting payload 转发保留（#301/#4288，T2a/T3 显式钉住）；provider-measured 墙仍拒（T4 control 钉 source=usage 60K>40K → ≥400）；唯一角落差异（!baselineKnown 且空消息：0→lastInputTokens）与 baselineFloor 同形分支逐字一致，出站 payload 实际恒 ≥1 条消息。未发现回归。

6. 是否破坏缓存：不涉及。不碰 prefix-cache 接缝、重试、去重键、增量游标；preflight sizing 只影响「何时触发压缩」，不改出站字节/cache 前缀。

7. 是否改了设计方向：无新架构。是既有 #1839 kill-at-reader grade 纪律向剩余两个未门控 entry 读的补全；更深的 lookup-miss-fallback（未知当未知处理而非 2e5 默认）已正确拆成独立 decision-request（#5502），不在本 PR。

8. 是否引入新配置：无新 env/config/CLI。仅一条内置表行（运行时数据，非 config-schema 字段）＋内部逻辑；显式 per-route/per-model override 仍优先（precedence chain 不变）。

行为变更披露确认（§Deliverables #4）

正文两条披露与代码一致，复核通过：

1. ^step- 且无显式覆盖的模型有效窗口：~150000 → ≈786432（native 1_048_576 − min(maxOutput, 25%×window) 预留；maxOutput 小时接近 1M）。
2. ESTIMATE-grade/unprovenanced baseline 下 preflight entry sizing：中毒臂不再 floor entryTokens/currentTokens init，改按 per-request 实测。

已顺手修（非阻塞发现，直接修在本分支）

• 常量与披露数字不一致：正文原写 ≈786432，代码常量却是 1_000_000（有效 750000）。把常量对齐到字段实测真实墙 1_048_576（同表 gemini-3 行同形；消除「猜略小」方向——750K–786K 之间无可压 payload 原先会被误 fail-fast），同步三处 T1 断言与正文披露 + commit ac098ee（定向套件 5/5 绿，CI 在新 head 重跑）。

非阻塞观察（不属本 PR，并入 #5500 跟踪，不单开 issue）

正文「除 entry 外所有 reader 都已约束」的表述略过强：还有两处读 lastInputTokens 未查 source——src/request-membership.ts:70（#4805 membership needle 启发式，中毒时 referenceScale 放大约 4× → mainScale 阈值偏高，软降级）与 src/server/prepare-responses.ts:597（codex compact-intercept 伪造 usage body，中毒时 inputTokens 偏大，窄 codex 路径）。二者都不喂 preflight entry sizing、都不引发 #5500 循环，属同一机制家族（缺 grade 纪律的 reader），已在 #5500 留评论记录。

测试有效性

T1 钉表行解析＋三负控；T2a/T2b 双向断言 entry 预算纪律（#3617：estimate 拒 / usage 放行）；T3 走真实窗口解析链复现 issue 形状（pre-fix 确定性耗尽→fail-fast≥400 / post-fix 干净转发、零 summary 调用）；T4 control 钉 measured 墙仍拒。断言有牙，本 bug 回归能抓住。

合并判定

可合。需人工 merge——step 家族默认窗口属 fallback/default-value 变更，§7.4 human-only 域。CI 将在新 head ac098ee 重跑（常量对齐，本地已定向验证）。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

PR #5501 二审完成：可合（APPROVE，需人工 merge）。GitHub review 已提交（同账号自批限制 → 以 COMMENT 事件带 APPROVE 判定）；独立重读 diff + 回归清单 + 补做变异验证（revert config.ts / preflight.ts 各自使对应测试转红）。已顺手修两处小问题：① PR 正文披露数字与代码常量不一致（正文 ≈786432 vs 代码 1_000_000）→ 常量对齐字段实测墙 1_048_576 + 三处断言 + 正文同步，commit ac098ee（定向套件 5/5，CI 新 head 重跑中）；② 两处未查 source 的 lastInputTokens reader（request-membership.ts:70、prepare-responses.ts:597）已在 #5500 留评论跟踪，未单开 issue。

[ranxianglei]

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

新 head ac098eeaa 复核：可合状态不变（APPROVE，需人工 merge）

本轮复审范围 = 上一审核 head 以来的唯一增量——commit ac098ee（上轮顺手修：常量对齐）：

• src/config.ts step 行 1_000_000 → 1_048_576 ＋ 三处 T1 断言同步（字段实测真实墙，同表 gemini-3 行同形；正文披露数字随之精确化）；
• 本地定向套件 5/5 绿；fetch 核对分支 tip = ac098eeaa、master+2 commits，此后无其他 push，全窗 diff 仍为 3 文件 +357/-2（上轮已逐段审过）；
• CI 本 head：镜像缓存 30 checks / failed=0（test 分片运行中），static-gates/kernel-guard/docs-sync/build-artifact 全绿。

无新发现。上轮两处非阻塞观察（request-membership.ts:70、prepare-responses.ts:597）继续在 #5500 跟踪。结论不变：可合，等 owner 人工 merge（§7.4 default-value 变更 human-only）。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

新 head ac098eeaa 复审完成：增量仅上轮顺手修的常量对齐 commit（config.ts + 三处断言，本地定向套件 5/5），分支无其他 push，CI 镜像缓存 failed=0（test 分片运行中）。结论不变：可合（APPROVE，需人工 merge），GitHub 复核评论已提交。注：GitHub 禁止同账号 approve 自己的 PR，审核只能以 COMMENT 事件带判定提交——若 lane 因缺少 APPROVE 事件反复重派，需 owner 侧调整门禁口径或人工 merge 收尾。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: qwen/qwen3.8-27b)

Merge conflict resolved and rebased onto current master. Summary of this run:

• Live state at start: PR open, mergeable_state: dirty vs master @ 0d0c0b95; head branch 2026-11-02_step-window-char-bound-poison @ 8222bc4152.
• Rebase onto origin/master (58a87962) — 4 commits replayed, single conflict in tests/issue5134-post-fold-credit.test.ts (T5).
• Conflict nature (semantic, not textual): this PR refined T5's stale ride-on from the char bound (350_000/estimate) to the calibrated per-send estimate (190_000/estimate); master's 2026-11-08 burst fix rewrote T5 to the opposite doctrine — pending credit outranks cycle-staleness (180_353/usage). In the merged code the credit-pending override (src/server.ts:2956) executes before this PR's new anchorStale branch (src/server.ts:3012–3015), so for T5's exact state (credit = 89_096 pending) the merged code returns 180_353/usage — master's assertion is the correct one. Kept master's T5 verbatim. Nothing of this PR's intent is lost: its calibrated-ride-on still governs stale anchors without pending credit, which its own suite pins.
• Targeted verification (local, targeted files only): tests/issue5134-post-fold-credit.test.ts, tests/issue5500-char-bound-poison.test.ts, tests/issue4383-stale-usage-anchor.test.ts, tests/issue5293-abort-nudge.test.ts, tests/issue5415-absorb-debt.test.ts → 38/38 pass; npm run typecheck clean. Full-suite verdict is left to CI on the pushed head.
• Push discipline: git fetch origin immediately before push — remote head unchanged (8222bc4152, no concurrent agent activity), so lease-guarded force-push of the conflict-rebase was safe: 8222bc4152 → 8dede7d140.
• Live state after push: mergeable: true, mergeable_state: blocked — conflict gone; now waiting on CI + approval as usual. No red-CI fixing was needed this run (the blocker was purely the dirty merge state).
• Disclosure: my first attempt at this summary was misposted to a different thread (wrong number resolution) and did not reach any upstream GitHub thread; it has been deleted from the mirror. No residual artifact anywhere.

---

中文结构化结论（八问）

1. 是什么问题:PR 与 master 产生合并冲突(merge state dirty),无法合并。
2. 什么原因导致的:PR 分支从 0d0c0b95 切出后,master 侧的 2026-11-08 burst fix 改写了同一测试 T5(tests/issue5134-post-fold-credit.test.ts),与 PR 对同一断言的细化形成语义级冲突——不是文本撞行,而是两条分支对「credit 挂起 + 锚点周期过期」这一场景给出了相反的裁决教义。
3. 影响范围与严重程度:仅阻塞本 PR 合并,不影响任何已发布行为;冲突解决本身零行为变更。
4. 怎么修复的:rebase 到最新 master;T5 保留 master 版本——因为合并后代码里 credit 挂起覆盖(src/server.ts:2956)先于 PR 新增的过期锚点骑乘分支(src/server.ts:3012–3015)执行,T5 场景下真实返回值就是 master 断言的 180_353/usage;PR 的校准估计骑乘行为在 credit==0 路径完整保留并由其自有测试套件钉住。
5. 是否有回归问题:无。定向跑了 5 个相关测试文件共 38 个用例全绿,typecheck 干净;PR 自身行为与其 master 侧行为的交互(credit 挂起时 usage 锚点优先、无 credit 时校准估计优先)与两个提交各自声明的意图一致,属代码顺序的自然结果,非本次引入的漂移。
6. 是否破坏缓存:不涉及——纯冲突解决,未触碰缓存/重试/游标逻辑。
7. 是否改了设计方向:不涉及。
8. 是否引入了新的配置:不涉及。

判定:冲突已清除、定向验证通过,PR 现处于正常等 CI/审批状态,本轮收工。

[ranxianglei]

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

Round 3 review — head 8dede7d14 (rebased onto current master + two new commits)

Scope since round 2 (head ac098eeaa): the branch was rebased onto current master (58a879622) and gained two new commits. Real PR surface = 10 files, +901/-12 (config.ts, preflight.ts, request-membership.ts, server.ts, prepare-responses.ts, session.ts, 4 test files). No kernel/, no lockfile churn, no unrelated files.

Original fix survived the rebase intact (verified by grep at the new head): src/config.ts:727 step row { match: /^step-/i, limit: 1_048_576 }; src/preflight.ts:1832/1835/1928 anchorGradeOk gate on both entry reads; test pins intact.

New commit 22fe21cc5 — extend grade discipline to the remaining lastInputTokens readers. Closes my round-1 finding (two ungraded readers) plus a third I had missed:

• src/request-membership.ts:70 — the #4805 needle main-scale bar now refuses estimate-grade/unprovenanced scalars, falls back to the needle's own per-request scale. Honest in-body note: today the incoming payloadUpper rides the same char-bound caliber so the ratio self-normalizes; the gate is drift-robustness, not a live-inflation fix. Correctly scoped.
• src/server/prepare-responses.ts:609 — codex compact-intercept minted usage no longer carries a char-bound arm (max(gradedScalar, est, 1)); the scales do NOT cancel here (est is chars/4 vs a poisoned char-count scalar) and codex recomputes its ledger from the minted body. Correct.
• src/session.ts noteAbortedTurn — untrustworthy meters skip the abort re-anchor stamp (a char-bound arm would delay the #5293 re-anchor ~4x and make #5415 bank phantom growth debt). Correct; dependent fixtures (issue5293/issue5415) updated to declare usage-grade provenance.
  Tests T6a/b/c (membership: poisoned / graded-wins / legacy-fallback), T7 + T7-control (forged-usage e2e through the real intercept path), T8 (re-anchor skip-on-poison + stamp-on-usage with the #5415 debt arithmetic intact). Bidirectional (#3617). ✓

New commit 8dede7d14 — C6: cycle-stale meter rides the calibrated per-send estimate (v0.1.212 field follow-up). When the last cycle ended without a usage settle, effectiveTokenCount handed the kernel meter the CHAR-BOUND estimate, inflating it ~2–4x and crossing the kernel's DESTRUCTIVE emergency-truncate band (0.95×window) on measured-safe sessions, latching clips into the delivered view until the next fold (#3939 ③). Fix at src/server.ts:3012: cycle-stale now rides contextEstimateTokens (the billing-caliber k̂-scaled estimate stamped at the same relay chokepoint, cache-ledger.ts:1849), capped min(…, raw upper) so it can never exceed today's fail-closed worst case. I traced the control flow: the branch sits after the usage/grade branches (skipped while anchorStale) and before the #728 anchor-less fall-through; usageAnchorCycleStale (session.ts:951) requires lastInputTokensSource==="usage" so anchor-less sessions never enter it; an absent contextEstimateTokens degrades gracefully to the legacy reading. Preserves #4383's intent (per-send evidence outranks a frozen anchor) while removing the inflation.
Tests C6a (ride-on, both calibers), C6b (four UNCHANGED calibers: fresh-settle / #728 anchor-less / legacy absent-field / k̂-overshoot cap), C7 (e2e field chain: usage settle → no-report resend → repeat prepare; teeth-proven — reverted fix latches clips, applied fix stays unclipped), C7-control (the kernel band STILL fires on a genuine over-limit reading through the real node — C6 does not soften the band). Three #4383/#5134 assertions re-pinned from the char bound to the calibrated scalar (accurate). High quality. ✓

Behavior-change disclosures verified against code (body items 1–7): step window ~150000→≈786432 (native 1_048_576); preflight entry sizing under an ESTIMATE baseline; membership bar; codex minted usage; abort re-anchor; C6 meter feed min(contextEstimateTokens, raw)/"estimate" (was char-bound). All match. No wire-format / persistence / config-surface / kernel changes. No new config/env. No cache impact.

Kernel boundary respected: this PR touches NO kernel/ files and carries no kernel-change: declaration; the two kernel-side residuals (destructive action on unprovenanced evidence; truncate ignoring protected tools) are correctly routed to #5693 (p1, scope/kernel, needs-decision) instead of smuggled in. No 🔶 需要决策 trigger for this PR.

CI (mirror cache @ 8dede7d): 30 checks / failed=0 / pending=0 — fully green at this exact head.

House conventions: no as-any/@ts-ignore/console.error in the touched files; test-fixture cast style matches the existing suite; comments explain why; clean focused diff.

Verdict: APPROVE (可合). Human merge required — the step-family default window (150K→786K) is a fallback/default-value change, product-decision tier per §7.4. Note: GitHub hard-blocks same-account self-approval, so this lands as a COMMENT event carrying the APPROVE verdict (same constraint as rounds 1–2).

八问

1. 是什么问题:#5500 会话死循环——step-* 模型窗口查表缺失(落 2e5 默认→实际墙仅 ~150K,真实窗口 1048576)+ char-bound 估算毒化(ESTIMATE-grade lastInputTokens 按字符尺度持久化、重放进每个后续请求)。本轮新增第三层缺陷(v0.1.212 现场):cycle-stale 降级把 char-bound 估算喂给内核计量器,放大 2–4× 越过破坏性 emergency-truncate 带(0.95×窗口),把截断夹进已交付视图并 latch。
2. 什么原因导致的:(A) CONTEXT_LIMIT_TABLE 无 step-* 行→落 base.modelContextLimit 2e5 默认;(B) #604/#1498-F1 失败臂按字符尺度持久化 ESTIMATE-grade 锚点,除 preflight entry 两个读点外其余读点已有 grade 纪律;(C) effectiveTokenCount 的 cycle-stale 分支骑在 char-bound 标量上。本轮补齐了我上轮点名/遗漏的三个无 grade 读点(membership / codex forge / abort re-anchor)+ C6 计量器骑点。
3. 影响范围与严重程度:p1。step-* 无显式配置的宿主(payload 介于真实大小与假墙之间→fail-fast 死循环);带中毒臂的会话(preflight entry 被抬高);cycle-stale 且测量安全的会话(误触发破坏性截断、工具输出被 clip 且 latch)。均为特定条件触发,非全量。
4. 怎么修复的:补 step 表行到字段实测墙 1_048_576(gemini-3 行同形);preflight entry 两读点加 anchorGradeOk gate(镜像 baselineFloor);三个残余读点同样 kill-at-reader(#1839 纪律);C6 让 cycle-stale 改骑 billing-caliber 的 contextEstimateTokens(k̂ 缩放),并用 min(…,当前 raw 上界)封顶,保证永不劣于今日 fail-closed 最坏值。根因修复,不是遮症状。
5. 是否有回归问题:无。逐项过:usage/overflow-arm 锚点分支、fresh-settle(#4383 回归护栏)、#728 无锚 never-reporting 车道、host preflight trigger、statusInputBaseline 显示读——全部不变(C6b 四标定 + C7-control 钉死);真实 over-window 拒绝未软化(C2-control / C3-control / C7-control 负向钉);依赖 fixture(issue5293/issue5415)已声明 usage-grade 出处。高危模式(fail-fast 变吞错、阈值漂移、输出格式变化)逐项过,无命中。
6. 是否破坏缓存:不涉及。无缓存/重试/增量游标/去重键变更。
7. 是否改了设计方向:无。沿用既有 kill-at-reader 与 billing-caliber 估算机制,未转向新架构/边界/数据流。
8. 是否引入新配置:无。复用既有 stats 字段(contextEstimateTokens/localInputEstimate),无新 env/config/默认值。

判定:可合(APPROVE),需人工 merge(step 家族默认窗口属 default-value 变更,§7.4 human-only)。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[review-lane r1] PR #5501 head 8dede7d 复审完成 — 可合 (APPROVE, review id 5852658993)。本轮分支已 rebase 到当前 master(58a8796)并新增两个提交:22fe21cc50(把 grade 纪律补到我上轮点名的两个无 grade 读点 + 第三个我漏掉的 abort re-anchor 读点)、8dede7d140(C6:cycle-stale 计量器改骑 billing-caliber 校准估算,修 v0.1.212 Windows/Pi/Responses 现场误截断)。原修复(config.ts step 行 1_048_576 + preflight entry anchorGradeOk gate)rebase 后完整保留。真实面 10 文件 +901/-12,无 kernel/、无 lockfile churn;kernel 侧残留(truncate 无视 protected tools / ProcessTurnInput 无证据等级)已由作者正确路由到 #5693(p1/kernel/needs-decision,现已关),本 PR 不触碰 → 无 🔶 需要决策。CI 镜像缓存 @8dede7d14:30 checks / failed=0 / pending=0 全绿。行为变更披露(7 项)逐条与代码核对一致。合并需人工(step 家族默认窗口 150K→786K 属 default-value 变更,§7.4 human-only);同账号自审仍被 GitHub 硬拦,verdict 以 COMMENT 事件承载。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

PR #5501 head 8dede7d14 复审收工 —— 可合（APPROVE，需人工 merge），GitHub review 已提交（review id 5852658993，同账号自批被硬拦 → 判定以 COMMENT 事件承载）。

本轮增量：分支已 rebase 到当前 master（58a879622）并新增两提交——22fe21cc（把 grade 纪律补到我上轮点名的两个无 grade 读点 + 第三个我漏掉的 abort re-anchor 读点）、8dede7d1（C6：cycle-stale 计量器改骑 billing-caliber 校准估算，修 v0.1.212 Windows/Pi/Responses 现场误截断）。原修复（config.ts step 行 1_048_576 + preflight entry anchorGradeOk gate）rebase 后完整保留。真实面 10 文件 +901/-12，无 kernel/、无 lockfile churn；kernel 侧残留已由作者正确路由到 #5693（p1/kernel/needs-decision，现已关）→ 本 PR 无 🔶 需要决策。CI 镜像缓存 @8dede7d14：30 checks / failed=0 / pending=0 全绿；行为变更披露（7 项）逐条与代码核对一致。合并需人工（step 家族默认窗口 150K→786K 属 default-value 变更，§7.4 human-only tier）。

（详细八问见上条 FLOOR 19 与 GitHub review 正文。）

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: qwen/qwen3.8-27b)

This trigger's "dirty" premise was stale — verified live before touching anything: at run start the PR was already mergeable: true (state behind, not dirty) at head 8dede7d140 from last run's conflict-rebase, with all 30 mirror-cache checks success/skipped (fetched 2026-11-11T03:15Z). But master had moved ~10 commits past my previous rebase base (58a87962 → dc041dd1ab), and three of those touched this PR's own files (src/config.ts ±89, src/server.ts −26, src/server/prepare-responses.ts ±31), so I caught up proactively rather than leave a latent semantic wedge:

• Rebase onto dc041dd1ab: all 4 commits replayed cleanly, zero conflicts.
• Targeted verification: same 5 suites as last run (issue5134-post-fold-credit, issue5500-char-bound-poison, issue4383-stale-usage-anchor, issue5293-abort-nudge, issue5415-absorb-debt) → 38/38 pass; npm run typecheck clean. (Workdir deps had been wiped by the daemon restart — redid npm ci + kernel build first.)
• Push discipline: fetched before push, remote head unchanged (8dede7d140, no concurrent activity), lease-guarded force-push: 8dede7d140 → 56a7fe2aa3.
• Post-push state (live): mergeable: true, behind only — master ticked once more during the run (ca1e6c83c9, PR [PR] fix: await unified snapshot writes and reset each debounce batch #6186 snapshot-write-lifecycle) which touches zero files of this PR (affinity-persist.ts / order-persist.ts / one new test), so I deliberately stopped chasing it: another rebase+CI cycle would buy nothing. behind is cosmetic — the merge commit picks up master at merge time. New CI runs on 56a7fe2aa3 are queued; verdict comes from CI/mirror cache, not from me polling.

---

中文结构化结论（八问）

1. 是什么问题:触发事件称 PR 与 master 冲突(dirty);实为上一轮 force-push 落地前的陈旧检测——本轮开始时实测已可合并。但 master 在我上轮 rebase 后又前进了约 10 个提交,其中 3 个碰到本 PR 的文件,存在潜在语义楔子。
2. 什么原因导致的:master 高活性推进(Request two capabilities: compressed block "retirement" and summary "external rewrite". | 请求两个能力：压缩块「退役」与摘要「外部重写」 #5765 blockOps gate、[PR] remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132 #6150、[PR] thirdparty-scan: flag dsh-live-compact as a known conflict (#5984) #6166 等连续合入),PR 分支基座落后;"dirty" 检测早于我上一轮的 rebase-push 生效。
3. 影响范围与严重程度:不修则无即时影响(PR 仍可合并,behind 非阻塞);风险是 master 侧对 config.ts/server.ts/prepare-responses.ts 的改动与本 PR 在同一决策域(meter sizing / 窗口表),拖到合并时才发现语义冲突代价更高。属预防性收敛,非缺陷修复。
4. 怎么修复的:rebase 到当时最新 master(dc041dd1ab),4 个提交干净重放、零冲突;定向 5 套件 38/38 + typecheck 验证语义无碰撞;lease 保护 force-push 收敛分支。master 再次前进的那一个提交([PR] fix: await unified snapshot writes and reset each debounce batch #6186)与本 PR 零文件重叠,判定不值得再烧一轮 CI,停止追。
5. 是否有回归问题:无。定向套件全绿、typecheck 干净;本轮未改任何代码内容(纯 rebase),行为面与上轮推送的 head 一致,仅基座更新。
6. 是否破坏缓存:不涉及——纯 git 操作,未触碰缓存/重试/游标逻辑。
7. 是否改了设计方向:不涉及。
8. 是否引入了新的配置:不涉及。

判定:分支已收敛到当前 master(除零重叠的最新一笔)、定向验证通过、无冲突;等 CI 在 56a7fe2aa3 上落定即可,本轮收工。