Repository navigation
[PR] test: align epoch-confirm/escalation-latch expectations with the merged #3641 contract (#4555) - #4569
[PR] test: align epoch-confirm/escalation-latch expectations with the merged #3641 contract (#4555)#4569ranxianglei wants to merge 1 commit into
Conversation
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.5552-pr.4569.d852068This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr4569.tgz
npm install -g packageInstall instructions are refreshed on each push. |
45a2376 to
ea9c2d5
Compare
#3641 contract (#4555) Both master-red tests after the #4525 train are expectation lag, not implementation drift — no runtime change (src/ untouched): - epoch-confirm-lifecycle all-missing shipped born-red in a5d1994 (train CI legs cancelled): it asserted foldEpochConfirmPending never latches, contradicting #3641 boundary 4 and MESSAGE-IDENTITY.md's epoch-confirm gate (total collapse latches, snapshot retained, nothing deleted). Expect pass >= 3 like one-interior-survivor; the kernel deactivates the zero-coverage block, the pause contract is shape-independent. - issue2915 escalation-latch-recovery encoded the pre-#3641 reap-at-escalation clean-slate contract. a5d1994 made the doctrine real (epochProtected gates the #2695 reap), so a healthy resend now yields 'resend' with the substrate retained and post-recovery id churn re-anchors ('reanchored') instead of rebuilding from zero. Verified: 90/90 across the three reconcile suites on rebased head ba5ab0b; old test ran 15/0 green on pre-train master (8ae56a0), pinning a5d1994 as sole cause (only train commit touching the latch path).
ea9c2d5 to
e7497bc
Compare
(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):
等 e7497bc 的全量 CI 出结果后补结构化八问结论。 |
|
Closing: both test alignments in this PR are already on master — epoch-confirm-lifecycle.test.ts asserts the |
(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)
tests/epoch-confirm-lifecycle.test.tsall-missing case was introduced by a5d1994 (cherry-pick of PR [PR] fix: retain fold snapshot until epoch confirmation resolves (#4260) #4261 head 594df72) and assertedfoldEpochConfirmPending === falsefor 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).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 thependingSettleprose-settle machinery insrc/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.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:
pass >= 3exactly 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.noop, blocks.length 0). a5d1994 made the doctrine real —epochProtectedgates 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 → quietresend. 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
tests/epoch-confirm-lifecycle.test.ts+tests/issue2915-order-centric-reconcile.test.ts+tests/fold-reconcile.test.tson rebased head (ba5ab0b)