[ranxianglei]

Continuation of #2725 — same branch (2026-10-11_cross-model-baseline), same single commit, rebased onto current master (d95799d86, 2026-10-24).

#2725 was closed unmerged by stale-ticket cleanup after #2683 was closed "completed" — but verification on 2026-10-24 showed the fix never landed in master (both producers of #2683 are still live there; evidence in the #2725 thread). GitHub refuses to reopen a closed PR whose head branch was force-pushed, so this PR carries the identical fix forward from the same branch.

Rebase notes: text conflicts resolved in src/server/handle.ts (#4205 window-caliber stamps — retirement block sits after both stamps, before the lastModel stamp) and src/server/relay.ts (#3129 v1 framer insertion; #3701 close-1009 byte-shed branch kept deliberately un-alerted, only the calibrated arm call applied); the calibrated-provenance flag was extended to the two armFailureShrink call sites master added after this branch cut (#3148 circuit-open fail-fast, #3701 oversize resend). Verified on the rebased head: typecheck clean + 49/49 across the targeted baseline/meter suites (issue2683 G1/G2/G3, #1498-F2, context-provenance-outbound and adjacent anchor/floor suites).

Closes #2683

What

Two producers of ONE mechanism family — session input baselines that outlive the context that measured them — each verified against the issue's live logs and reproduced with e2e fixtures:

1. The failure arm adopts a bound, not a reading. armFailureShrink (no-usage relay 5xx / network failure) persisted the char-count UPPER BOUND of the outbound payload (estimateCoreMessagesUpper class, 1.7–4× real) as session.stats.lastInputTokens. The fix: arm emergency shrink on outbound payload size, not raw body (#1493) #1498-F1 bound floor was justified by the kernel's tool-result truncation loop consuming the arm as remaining depth (arm emergency shrink on upstream failure to break relay-5xx deadlock #604 rescue) — but since the emergency shrink / preflight 的本地估算用错了视图：超长会话（客户端历史 127 万，压缩后 outbound 仅 ~16 万）在上游失败无 usage report 时，bili 按原始历史估算 #1492/任何一轮没有拿到 usage 报告的请求，都会把上下文分母放大 12–20×（nudge 读到 >1000% usage、假 EMERGENCY、基线冻结数分钟） #1839 reader quarantine NO processTurn feed passes an estimate-grade arm into any band (preflight runs noEmergencyTruncate; every decision floor gates on source), so the bound's only live effect was phantom mass: inflated preflight entry budgets, responses-lane output headroom, displays — surviving model/window changes until the next usage report. (Issue's live case: 441,193 armed against a 202,885-token payload.)
   → The arm now adopts the same billing-caliber reading the preflight trigger uses (k̂-scaled chars/4 + wire overhead + image reserve), and its log names the provenance: (k̂-calibrated outbound reading) / (raw outbound reading) (issue proposal feat: OpenAI Responses API support (/v1/responses) #4).

2. Trusted baselines survive model switches. A usage report measured under model X kept deciding requests for model Y — only ORIGIN changes demoted it (DSH: 切换模型后，preflight 在窗口远未满时提前触发压缩 #1933 F2). After a mid-session switch the foreign bill pinned the preflight loop's decisionTokens above the fold target forever (the floor survives folding — folding shrinks the payload, not the bill) → folds until summary-budget exhaustion although the payload fit. (Issue's live case: 566,322 billed under a 1M-window model judged against a 400k window → two ~3-minute emergency folds at 62% of new-window occupancy.)
   → A real model change (both sides known; unknown→known never counts) now retires the trusted-grade baseline at the per-request model-stamp site — before prepare sizes the nudge and before the trigger runs — logged with both models and the retired value. The next usage report re-anchors. Precedents: [Bug] 上下文展示估算 206K，usage 返回后为 98K：希望校准估算并区分安全上界 #2117 B already gates k̂ by model ("a factor learned on another model acts as absent"); DSH 分叉会话后触发大量 preflight overflow compress #2383 names foreign-model baselines untrusted for regime selection.

Behavior changes (old → new)

• Armed value on a no-usage failure: char-count upper bound → billing-caliber estimate. It may now land BELOW a future bill on dense JSON/CJK routes where chars/4 undercounts (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) — acceptable because estimate-grade values are quarantined from every decision floor and k̂ closes most of the gap where route+model provenance matches.
• Usage/overflow-arm baseline + usage anchor across a model switch: trusted verbatim → retired until the next usage report. Forensic state preserved: cumulative inputTokens meter, last-usage timestamp (contextTokensAt), calibration ring, cache-ledger history. Same-model window moves do NOT retire (the bill stays valid → genuine overflow still folds); estimate-grade arms survive (byte-scale property of the payload).

Config surface

None — no new env/config/CLI keys.

Both compression modes

Retirement sits at the request-handling chokepoint before prepare/trigger, so it applies identically in plugin mode and proxy mode (the mode split happens downstream). The arm-caliber change affects both modes equally (both consume outboundPayloadBreakdown). No wire-shape change: message rebuild, system handling, tool injection, acp_summary, preflight protocol all untouched — only session-state bookkeeping values.

Tests

• New tests/issue2683-cross-model-baseline.test.ts: G1 cross-model retirement e2e (566K billed under claude-a → claude-b turn makes ZERO summarization calls, no phantom (566000/…) meter, re-anchored to the real 6K bill); G2 no-usage network failure arms inside [chars/8, char-count) with the provenance clause, retry forwards without folding; G3 control — a SAME-MODEL over-window bill STILL FOLDS via its trusted floor (floor semantics preserved for genuine overflow).
• Re-scaled #1498-F2 assertion in tests/issue1493-arm-outbound-payload.test.ts (raw-view measurement pinned at chars/4 scale; char-count adoption now explicitly forbidden).
• Updated pins in tests/context-provenance-outbound.test.ts (cross-model authority loss vs same-model retention; fixture growth sized so the billing-caliber arm still outgrows the fixture bill and the replace-stale-usage mechanic keeps firing).
• Targeted batch green on the rebased head: typecheck + 58 tests across 11 suites covering every touched consumer path (issue1493, issue1839, issue2313, issue2484, issue1595, issue1569, issue2033, silent-backend-nudge, issue857, context-provenance-outbound, issue2683). Full suite to CI.

Known side effect (disclosed): a host that mixes multiple models under one UN-FORKED session id will see freshly established baselines wiped at each switch — self-healing (the next usage report re-anchors within one round trip); persona-style hosts are forked per KDD#9 and unaffected.

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-11_cross-model-baseline (74bfaef)

Latest build — install this exact version

npm install -g billion-context@0.0.5533-pr.4583.db9caaf

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
74bfaef	billion-context@0.0.5533-pr.4583.db9caaf	build

Download artifact

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

tar xzf billion-context-pr4583.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(stale baseline / armFailureShrink / #2683)检索 open+closed issue/PR:该机制族内唯一 open 条目即本 PR;前序 #2725(同分支 2026-10-11_cross-model-baseline、同一单提交)已被 stale-ticket 清理关闭未合入,本 PR 按设计承载同一修复(GitHub 不允许重开 head 被 force-push 的已关闭 PR);其余历史 baseline 类 PR(#1503/#1940/#2316/#2032/#4443 等)均已合并且机制不同。

标签: bug, scope/kernel, p2(bugfix PR;改动面是 src 侧压缩决策输入——baseline 退役 + arm 口径,跨客户端核心管道、不涉及特定客户端;功能性缺陷、影响明确但有自愈绕行 → p2)

审核进行中,初步结论:

• 已拉取分支 pr-4583(单提交 74bfaef),通读全部 8 个文件 diff,并对照原 issue [Bug] armFailureShrink adopts the upper-bound estimate as the session baseline - it survives a model/window change and then forces an unnecessary emergency fold (2x ~3-minute stalls) #2683 的两个实测 case 逐条验证:两个 producer 都覆盖到,修复打在根因上——① arm 从 char-count 上界改为 billing-caliber reading(k̂ 缩放 chars/4 + wire overhead + image reserve);② 真实模型切换时在请求入口退役 trusted-grade baseline(位于 "Output reservation extended to all wires" (#4184) landed without test coverage: widespread red on master is blocking th | 「输出预留扩到所有 wire」(#4184) 落地没带测试面: master 大面积红卡住 v0.1.204 发版 (已 revert, 重做要求见内) #4205 window stamp 之后、lastModel stamp 之前,顺序正确;unknown→known 不触发,same-model window 移动不触发)。
• 回归清点:base 之后 master 前进的 18 个提交与本 PR 无文本冲突(mergeable=True)也无语义冲突(窗口内对 src/server.ts/src/session.ts 的改动不触碰本 PR 改动的区域);被清字段(lastUsageGradeTokens/overflowArmTokens/contextTokens 等)的所有读者都有 undefined/typeof 守卫;estimateCoreMessagesUpper 在 preflight/prepare-* 等 6 处仍有使用,从 arm 移除不留死代码;四条失败路径(circuit-open [bug] 上游在会话中途不可用时，bili 把一次故障放大成四重代价：固定节奏 20 连击全部冷发、输出预算被静默钳制 83 次（最狠 65536→1360）、紧急收缩与 nudge 停摆均不可见 #3148 / network / close-1009 byte-shed [Bug] Jimp is active, but image accumulation still triggers Responses WS 1009; duplicate request size not reduced (v0.1. | [Bug] Jimp 已生效，图片累积仍触发 Responses WS 1009；重复请求未缩小（v0.1.200） #3701 / upstream 5xx)全部带上 provenance flag;PR 声称「arm emergency shrink on upstream failure to break relay-5xx deadlock #604 truncation-loop 消费者已不存在」已核实(preflight 四处 feed 均走 noEmergencyTruncate)。
• 测试有效性:G1/G2 是真回归钉(回退修复会变红:G1 无修复时 trusted floor 钉住 decisionTokens≥target 必然产生 summarization call;G2 无修复时 arm≥char-count 违反新上界断言);G3 是负向控制(同模型真溢出仍 fold,#Gate Test Discipline 双向要求满足);fix: arm emergency shrink on outbound payload size, not raw body (#1493) #1498-F2 重标为双向区间断言 [chars/8, chars);context-provenance-outbound 在四 wire 矩阵里钉了 cross-model 行且 fixture 增长按新口径重新定尺寸。
• CI:head 74bfaef 目前 3 个 test shard 红(windows-22 shard1/shard3、ubuntu-24 shard1),其余 lane 仍在跑。正在排查这三个红是 master 既有 flake 还是本 PR 引入,确认后单独报告。

最终结构化八问结论待 CI 分诊后给出。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

PR #4583 审核结论（独立复审：diff 逐行 + 测试全读 + master 漂移核对）

审原 issue（方向关）：#2683 两个 case 的复现数据完整（v0.1.191 实测日志），定位准确——修的是基线「测量语境失效后继续生效」这个根机制，不是症状。本 PR 与 #2683 提议的两条修法对应且更稳（不 clamp，直接换口径；退役只限 trusted-grade）。方向成立。

回归审查（逐路径清点改动前既有行为）：

• armFailureShrink 四条失败路径（circuit-open / network / close-1009 byte-shed / upstream 5xx）全部保留 arm 行为，只改 adopted 值口径——arm emergency shrink on upstream failure to break relay-5xx deadlock #604 救援臂仍然 load-bearing，G2 钉住 retry 恢复且零折叠；
• fix: arm emergency shrink on outbound payload size, not raw body (#1493) #1498-F2「raw 视图下仍要 arm」以新区间 [chars/8, chars) 双边钉住（旧断言 >= chars 改为 >= chars/8 && < chars），不是放松而是随口径变化的等价重标定；
• G3 负向对照钉住同模型真溢出仍走 trusted floor 折叠——退役逻辑不过度触发；
• 被清字段（lastInputTokens/lastUsageGradeTokens/contextTokens/overflowArmTokens/source）的全部读者均有 undefined 守卫（session.ts:800 ?? 0、:839 anchored 检查、server.ts:3248、handle.ts:2080、sessions-data num()>0、persist typeof 序列化）；
• 与 base 之后 master 新增 18 个提交逐一核对：server.ts 窗口 diff 与本 PR 符号零交集，session.ts 仅 After an interleaved metadata refresh, the entire history of the same session is resent in a different form, and already-collapsed content fails to re-anchor. | interleaved 元数据刷新后，同一会话整段历史换形态重发，已折叠内容 re-anchor 认不回来 #3301 加瞬态字段，无语义冲突；GitHub mergeable=True。

代码审查：退役点放在 handle.ts 中 #4205 窗口戳之后、lastModel 更新之前，顺序正确（先按新模型算窗口再退役旧账）；unknown→known 不触发、同模型仅变窗口不触发，条件收紧到位；arm 侧复用 preflight trigger 的同一 billing-caliber 公式（k̂ 缩放 + overhead + image reserve），armCalibrated 标志纯增量。无 as any、无吞错、日志措辞带来源子句可 grep。

测试有效性：三个 case 都是 e2e 级（真实 startServer + fake SSE relay），断言有牙且双向；G1 的 nonStreamCalls() === 0 在修复前必红（trusted floor 会钉死 decisionTokens ≥ target → 强制折叠），是真正的回归钉。小缺口（非阻塞）：circuit-open/5xx 两条 arm 路径没有各自的 calibrated-flag 直测（与已测 network 路径同形代码）；退役后无持久化 roundtrip 断言（序列化守卫已覆盖 undefined）。

CI 状态（诚实报告）：head 74bfaef 的 11 个 test shard 全部红。判定为既有 master 红风暴而非本 PR，证据链：① base d95799d 自己的 CI run 就是 8/11 红，含 windows-22 s1/s3（本 PR head 亦红的两个 shard）；② master 今日连续多 push 全红；③ 已有跟踪 issue #4591（今天 15:41Z 立项，open）实证了 3 个失败域并给出签名：A) error-dump #4297 字节上限（'0: 1111 bytes'，tests/error-dump.test.ts:200，ubuntu 22,2/24,2 确定性红）、B) generation-boundary 测试自身位置参数错位（tests/generation-boundary.test.ts:130，win 22,3 + ubuntu 22,3/24,3）、C) stale kernel dist mtime 守卫（tests/golden-storage-drive.ts:73）；④ 用仓库自己的 partition 脚本算出本 PR 的三个测试文件落在 shard 5 和 shard 2，不在 base 绿/head 红的 shard 1。残留缺口：check-runs logs API 对本 PAT 返回 404（已完成 job 也取不到），无法逐测试点名对比——若 owner 要求确定性证据，#4591 修复落地后重跑即可一锤定音。作者声称的定向本地绿（typecheck + 11 suite 58 tests @ rebased head）与上述静态核查一致。

---

结构化八问

1. 是什么问题：会话输入基线活过了它的测量语境——①无 usage 的 relay 失败把 char-count 上界（超读 1.7–4×）持久化成会话基线；②模型 X 下测得的 trusted usage bill 继续替模型 Y 做折叠决策。
2. 什么原因导致的：armFailureShrink 采用上界估计而非计费口径读数；模型切换时没有任何机制退役模型相关的 trusted 基线（此前只有 ORIGIN 变化会降级，DSH: 切换模型后，preflight 在窗口远未满时提前触发压缩 #1933 F2）。
3. 影响范围与严重程度：中会话切模型或遇无 usage relay 失败的会话——幻影紧急折叠、单次最长 ~180s 停顿（live case 两轮共 ~3.6min）、预算/headroom/展示虚高直到下次 usage 报告。p2：可用、自愈、无数据丢失。
4. 怎么修复的：arm 改采 billing-caliber reading（与 preflight trigger 同一公式，k̂ 缩放）；真实模型切换（双向已知 id）在请求入口退役 trusted-grade 基线；estimate-grade arm 保留（字节尺度属性跨模型有效）；取证数据保留；下次 usage 报告重锚。
5. 是否有回归问题：审过的既有行为均保持（见上回归审查五条）。已披露行为变更（旧→新+原因）：a) armed 值从 char-count 上界 → billing-caliber 读数——密集 JSON/CJK 路由上可能低于未来账单（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 chars/4 低估面），可接受，因 estimate-grade 本就隔离于所有决策地板；b) 切模型后首轮真溢出依赖 overflow-arm 救援而非 proactive fold——fail-soft 取舍，代价是一次性而非循环；c) 多模型未 fork 宿主每次切换清掉刚建立的基线——一轮内自愈，persona 宿主按 KDD#9 已 fork 不受影响。高危模式逐项过：无 fail-fast→吞错、无未披露阈值/时序漂移、日志仅追加来源后缀。
6. 是否破坏缓存：不涉及——无缓存键/格式/账本变更，prompt-cache routing 不动，calibration ring 与 cache-ledger 历史保留。
7. 是否改了设计方向：不变——扩展现有 provenance/quarantine 体系（emergency shrink / preflight 的本地估算用错了视图：超长会话（客户端历史 127 万，压缩后 outbound 仅 ~16 万）在上游失败无 usage report 时，bili 按原始历史估算 #1492/任何一轮没有拿到 usage 报告的请求，都会把上下文分母放大 12–20×（nudge 读到 >1000% usage、假 EMERGENCY、基线冻结数分钟） #1839/DSH: 切换模型后，preflight 在窗口远未满时提前触发压缩 #1933/[Bug] 上下文展示估算 206K，usage 返回后为 98K：希望校准估算并区分安全上界 #2117 B）补两个缺口，无新边界。
8. 是否引入了新的配置：无。

合并判定

代码审核通过，可合——但两点留给人工：① auto-merge gate 规定 fallback/默认语义变更属产品决策，必须人工合并；② CI 当前被 #4591 红风暴淹没（base 同红，非本 PR 引入），建议 #4591 三项修复落地或 test 车道转绿后由 owner 合并。另注：本 PR 由 #2725 延续而来（force-push 关闭后 GitHub 拒绝 reopen），分支/提交未变，已在首条回复说明查重结论。