[ranxianglei]

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

Closes #4555

What this is

Both master-red tests after the #4525 train are test-expectation lag, not implementation drift. No runtime change — src/ untouched, two test files only.

Verification chain (all reproduced locally)

1. Born red: tests/epoch-confirm-lifecycle.test.ts all-missing case was introduced by a5d1994 (cherry-pick of PR [PR] fix: retain fold snapshot until epoch confirmation resolves (#4260) #4261 head 594df72) and asserted foldEpochConfirmPending === false for all 5 passes — while its own commit's implementation latches at pass 3 (streak>=3 && unmatched>=10 → src/fold-reconcile.ts:2060). Red at the train tip because the train's test legs were cancelled (master red after #4525 train merge: sidScope rename straggler + #4241 judge double-fix collision + missing BILI_CODEX_STATE_MODE ledger row #4554).
2. Regression attribution: git log 8ae56a01..HEAD -- src/fold-reconcile.ts src/session.ts tests/issue2915-order-centric-reconcile.test.ts → a5d1994 is the only train commit touching the latch path. The issue's suspicion on 9a9db1f is disproven at file level: that commit drops the pendingSettle prose-settle machinery in src/loop/tag-echo-filter.ts — a different subsystem. Old test independently verified 15/0 green in a worktree at pre-train master (8ae56a0), 14/1 red at tip.
3. Which side is the intent: the owner already decided this in Snapshot as session reconstruction: identity claim order takes priority, message hash demoted to reference signal (root- | 快照即会话重构:身份认领顺序优先,消息哈希降级为参考信号(根治 #3588/#3640 一类误判) #3641 (OP boundary 4, 2026-10-18): full collapse → not default reset — error + banner, user confirms, passthrough until confirmed; Stage C acceptance: 「全塌+相似度塌才进『待确认』态(passthrough+banner)」. The in-repo doctrine states it flatly (MESSAGE-IDENTITY.md, Epoch-confirm gate): “total collapse latches foldEpochConfirmPending, emits one error, and preflight passes requests through verbatim — folding pauses; nothing is deleted, nothing is silently reset.” The code comment names the class explicitly: “wholesale replacement by foreign history … its terminal verdict is now an epoch confirmation, not 'destroyed'”.

So the implementation matches the owner-decided contract; both tests lag it:

• all-missing: now expects pass >= 3 exactly like one-interior-survivor. The kernel deactivates the zero-coverage block (that part of the old expectation was right), but the pause contract is shape-independent — the snapshot is retained either way. Full lifecycle re-verified green: retention through passes 4–5, decompress works on the retained inactive block, realignment reactivates + restores the summary.
• escalation-latch-recovery: encoded the pre-Snapshot as session reconstruction: identity claim order takes priority, message hash demoted to reference signal (root- | 快照即会话重构:身份认领顺序优先,消息哈希降级为参考信号(根治 #3588/#3640 一类误判) #3641 reap-at-escalation clean-slate contract (noop, blocks.length 0). a5d1994 made the doctrine real — epochProtected gates the [Bug][proxy] unframed host-side history deletion leaves zombie fold blocks rendering carriers forever — one carrier flaps in/out every turn (persistent prefix-cache churn) #2695 reap — so the receipt is now: healthy resend → resend (missing 0, reaped 0, substrate retained, latch+streak cleared); post-recovery id churn → reanchored (the retained block re-anchors onto churned bytes, unmatched 0, no re-escalation); repeat of the same history → quiet resend. That is the upgrade-in-place healing contract restated for Snapshot as session reconstruction: identity claim order takes priority, message hash demoted to reference signal (root- | 快照即会话重构:身份认领顺序优先,消息哈希降级为参考信号(根治 #3588/#3640 一类误判) #3641, with an added stability pin against dead-looping.

Behavior changes

None. Tests-only alignment to behavior that already shipped via the #4525 train. If the owner instead wants all-missing to NOT latch (clean-slate reap), that would be an implementation change contradicting #3641 boundary 4 / MESSAGE-IDENTITY.md — please say so and I will revert and file the mechanism change instead.

Local evidence

• 90/90 across tests/epoch-confirm-lifecycle.test.ts + tests/issue2915-order-centric-reconcile.test.ts + tests/fold-reconcile.test.ts on rebased head (ba5ab0b)
• typecheck clean on rebased head
• pre-train worktree control: old suite 15/0 at 8ae56a0

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-24_4555-epoch-confirm-test-alignment (e7497bc)

Latest build — install this exact version

npm install -g [email protected]

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
e7497bc	[email protected]	build
ea9c2d5	[email protected]	build
45a2376	[email protected]	build

Download artifact

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

tar xzf billion-context-pr4569.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 即 #4555 的执行项(Closes #4555)。相关但不同:#4554(#4525 train 合并机械损伤,不同机制)、#4260(a5d1994 已修复的 reap-before-pause 缺陷)、#3641(语义决策来源)。出站查重:按失败签名(epoch-confirm-lifecycle all-missing + issue2915 escalation-latch-recovery)检索近期 open/closed issue 与 PR,无并列的修复 PR。

标签: bug / scope/kernel / p2(一句话理由:两个 master-red 均为测试期望滞后——确定性 CI 噪音、零运行时影响;涉及面是 fold-reconcile 压缩决策面,归 kernel scope。)

初审进展(独立复核,非照抄 PR 描述):

1. 方向关 ✅:diff 只动两个测试文件(+23/−13),src/ 零改动;方向与 owner 在 Snapshot as session reconstruction: identity claim order takes priority, message hash demoted to reference signal (root- | 快照即会话重构:身份认领顺序优先,消息哈希降级为参考信号(根治 #3588/#3640 一类误判) #3641 boundary 4 拍板的契约一致(全塌 → error+banner+passthrough 待确认,不是静默 reset),MESSAGE-IDENTITY.md epoch-confirm gate 原文逐字核对无误。
2. 断言有牙(双向验证):新头本地定向跑三套件 90/90 绿;把改动前的旧测试文件换回仓库对新实现反跑——各恰好挂一条、且正是被改的那条断言(all-missing latch expected false / actual true;escalation receipt 'resend' !== 'noop')。不是恒真改写,是这个 diff 本身让红变绿。
3. master 已前进 → 已 rebase:旧头(ea9c2d5)上 8 条 test lane 红,时间与刚合入 master 的 gate 修复(4d34abf,「merge-ref casualties, master-red too」)吻合,判定为旧基线上的存量红而非本 PR 引入;当前 master(146b34f,[PR] merge: tier-5 train — dsh prepare/streaming domain (7 PRs: bili_final terminal marker, auto-fold storm gate, wire-only guard) #4590 train)又改了 src/fold-reconcile.ts(f4a21bb),所以我把分支 rebase 到最新 master(e7497bc,无冲突,diff 不变),rebased 头上重跑定向 90/90 绿 + typecheck clean,已 push 到 PR 分支。

等 e7497bc 的全量 CI 出结果后补结构化八问结论。