Skip to content

[PR] fix: keep dsh presets bootable when bili-native is disabled (#3662) - #3729

Merged
ranxianglei merged 2 commits into
masterfrom
wt-3662-ework-daemon
Oct 21, 2026
Merged

ranxianglei merged 2 commits into
masterfrom
wt-3662-ework-daemon

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Oct 18, 2026 •

Copy link
Copy Markdown
Owner

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

What this fixes

Disabling the plugin in dsh (plugin manager, or disabled: true in any patch layer) permanently bricked session creation: all four web-profile presets stayed pending (waiting for service: biliBasicCompactionPolicy), because the ONLY provider of that service was the very bili-native entry being disabled — while the static bundle patch kept gating them on it. The env kill-switch (BILLION_CONTEXT_PLUGIN=0) hit the same path. New/resume then failed with Unknown agent preset: standard; only removing the package from dsh.profile.bundles recovered.

Approach

Split the gate out of bili-native into its own always-on cordis entry bili-compaction-gate (new tsup entry + ./dsh-compaction-gate export), mounted alongside bili-native by BOTH patch carriers (dsh.bundle.patch.yml, launcher .bili-acp.patch.yml; the client shim gains the subpath export for the bare lane).

Behavior changes (old → new)

Path Before After
bili-native disabled / kill-switched 4 presets pending forever; new/resume session creation bricked presets boot with STOCK native compaction; dsh fully usable, compression off
bili-native module fails to load deliberate fail-safe brick (same pending state) stock fallback; #1729 wire-level refusal guard is the backstop
normal boot (plugin enabled) preset compaction rows disabled at load time unchanged (verdict active ⇒ same transform)
visible surface 1 cordis entry (bili-native) 2 entries; bili-compaction-gate must NOT be individually disabled (re-creates the brick) — documented in bundle patch header + CLIENTS/CONFIGURATION en+zh

Residual footgun (accepted): disabling bili-compaction-gate itself re-creates the old pending state; docs point at uninstall via dsh.profile.bundles removal instead. Pre-fix bricked installs self-heal on upgrade (the patch layer mounts from whatever package is listed in bundles, regardless of bili-native's state); manual un-brick recipe documented ($DSH_HOME/cordis.patch.yml, four inject: null rows — $DSH_HOME per issue comment, defaulting to ~/.dsh).

Tests

  • new tests/issue3662-dsh-gate-fallback.test.ts: gate surface; BRICK scenario (absent verdict ⇒ service still provided, ordering intact, transform passes rows through stock); normal boot (verdict active ⇒ rows disabled); verdict channel semantics; both patch carriers mount the gate (bundle yml + bare lane + URL lane).
  • tests/issue3354-dsh-compaction-disable.test.ts: transform test now runs under an explicit verdict; carrier agreement pins the gate insert block in both layers.
  • tests/dsh-native.test.ts: apply() asserts the verdict instead of overlay registration (active plan sets it; kill switch neither registers nor declares it).
  • tests/launcher.test.ts: overlay assertions pin the gate row in both lanes; shim tests cover the subpath export + symlink and the missing-gate-bundle degradation (no dangling symlink).
  • Targeted local run: 5 suites, 384 pass / 0 fail / 0 skip; typecheck clean; check:unused-exports + check:test-import-graph clean; full suite on CI.

Cluster update (2026-10-21): rebased + encompasses #4147 / #3962 / #4002

(ai: zhipuai-lb/glm-5.3) — this section; owner-directed one-PR treatment of the dsh.bundle.patch.yml cluster (discussion #3916, group ①).

Rebased onto current master (was 594 commits behind; conflicts in src/agent/dsh-native.ts, tests/issue3354-dsh-compaction-disable.test.ts, tsup.config.ts resolved by union — master's #3806 overlay-gap alarm and conversation-recall imports kept, the gate entry + verdict wiring kept).

#4147 (p1) — structural matching. The transform matched only the factory @deepseek-ai/dsh-agent-preset entry by name, so custom/profile-local presets' nested compaction-basic/command-compact rows stayed ARMED and a landed dsh checkpoint destroyed the compression substrate (reporter session a391: 128 refused calls, native-compaction-inferred: substrate destroyed). The middleware now walks ANY entry's resolved config.plugins; rows are disabled wherever they nest; entries without compaction rows keep their resolved config object identity. The four patch gate-row ids remain as ordering insurance for the factory roster, not the coverage boundary.

#3962 — coverage telemetry. Every rewrite is recorded on the shared globalThis channel (dshCompactionGateTelemetry(): covered entries × row counts, factory flag). The durable [dsh-client] boot notes (web + non-web branches) append actual coverage, so a patch-row miss on a renamed roster is visible without the host loader log. Waiver added in check-unused-exports for the channel payload type (same rationale as #3754).

#4002 (p3, direction 2) — docs. CLIENTS.md en+zh now state preset sessions have no /compact menu BY DESIGN (bili owns compression; no re-enable switch per #3651/#3666) and describe the structural coverage + the coverage note.

Tests: new tests/issue4147-dsh-overlay-structural.test.ts (5 cases: custom preset + group recursion, negative identity/absent-verdict, telemetry dedup/max/factory/reset, name truncation); both existing negative-assertion rows flipped to the structural semantics (entries without compaction rows still pass verbatim). Targeted: typecheck clean, 5 suites 414 pass / 0 fail, unused-exports + test-import-graph + build clean. Full-suite verdict: CI test lanes on the pushed head.

Closes #3662
Closes #3889
Closes #4147
Closes #3962
Closes #4002

@ranxianglei ranxianglei changed the title fix: keep dsh presets bootable when bili-native is disabled (#3662) [PR] fix: keep dsh presets bootable when bili-native is disabled (#3662) Oct 18, 2026
@ranxianglei
ranxianglei force-pushed the wt-3662-ework-daemon branch 2 times, most recently from 691f9e0 to e6a87e4 Compare October 18, 2026 15:53
@github-actions

github-actions Bot commented Oct 18, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: wt-3662-ework-daemon (69cf200)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.4970-pr.3729.b7117e5

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-pr3729.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

上游 PR 冲突已解决:分支 wt-3662-ework-daemon 已 rebase 到当前 master(落后 41 个提交)并 force-push,PR 现在 mergeable: true(状态 blocked = 仅等新 head 的 CI 检查)。

八问

  1. 是什么问题:本 PR(禁用 bili-native 后 dsh preset 仍可启动的修复)与 master 产生合并冲突,无法合入。
  2. 什么原因导致的:两条线都改了同一批文件——master 侧 dsh attach mode test intermittently fails on Windows CI: the async recovery flow reverts the registered origin back to i | dsh attach 模式测试在 Windows CI 偶发红:异步恢复流程把注册 origin 改回旧值(测试隔离竞态) #3616 扩展了 dsh-native.ts 的 incremental-client import 并给测试文件加了新 import(once/startServer/defaultConfig)、新增 origin-generation 逻辑;PR 侧把 registerCompactionDisable 改名为 declareDshCompactionActive。另外 rebase 验证时发现 PR 自身一个遗漏:它把 src/agent/dsh-compaction-gate.ts 加进了 tsup entry,却漏了同步两个 gate 脚本(scripts/check-test-import-graph.mjs / scripts/check-unused-exports.mjs)里硬编码的 ENTRY_FILES 清单(注释明确要求「Keep in sync with tsup.config.ts entry」),导致 check:test-import-graph 实际会红——PR 描述里「clean」的说法不成立。
  3. 影响范围与严重程度:只影响本 PR 的可合并性,无用户面影响;冲突解除前 PR 完全不可合(blocking),属开发流程问题。
  4. 怎么修复的:rebase + 两处冲突按「语义并集」解——import 块同时保留 master 的新增符号和 PR 的改名;测试文件取 PR 的 verdict 版 import(PR 有意废弃了 BILI_COMPACTION_DISABLE_SERVICE 断言,合并后全文确认零引用)+ master 的三个新 import。gate 脚本缺漏用独立 commit 修掉(两个清单各加一行 src/agent/dsh-compaction-gate.ts)。
  5. 是否有回归问题:无。逐项核对:master 的 dsh attach mode test intermittently fails on Windows CI: the async recovery flow reverts the registered origin back to i | dsh attach 模式测试在 Windows CI 偶发红:异步恢复流程把注册 origin 改回旧值(测试隔离竞态) #3616 origin-generation 逻辑与新测试原样保留;PR 的 verdict 机制(declareDshCompactionActive/isDshCompactionActive)原样保留;package.json version 未动;typecheck 干净;dsh-native + issue3354 + issue3662 + launcher 四个定向套件 397 pass / 0 fail / 1 skip;两个 gate 脚本均 clean。全量套件按惯例交给远程 CI 在新 head 上判定。
  6. 是否破坏缓存:不涉及(未触碰缓存/重试/游标逻辑)。
  7. 是否改了设计方向:不涉及,纯 rebase 对齐,两边意图均为并集保留。
  8. 是否引入了新的配置:无。

备注:本地 wt-3662-ework-daemon 分支被 #3662 的工作区占用,故经临时分支完成 rebase 后推送到同一远端 ref,已清理。收工判定:冲突已清、定向验证全绿,等 CI 转绿后可由人工合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

标签: bug / dsh / scope/wiring / p1 —— 禁用 dsh 插件后新建/恢复会话完全不可用(用户不能用),修复点在客户端接线层。
查重: 不重复 —— open+closed 全量检索,#3662 仅被本 PR 引用;同家族的 #3354/#3451 是功能侧(引入 disable transform)、#3568 是文档、#1777 是启动告警,症状与机制均不同。

以下是对 diff 的独立重审(未照抄 PR 描述),本地已复现全部关键路径:npm ci + build、typecheck、4 个定向测试套件(398 pass / 0 fail)、两条 gate 脚本本地实跑。

① 方向关(原 issue)

根因定位正确:静态补丁层(看不到运行时状态)把 4 个 web preset 的启动锁在 biliBasicCompactionPolicy 服务上,而该服务的唯一提供者就是被禁用的 bili-native 本体——自引用 latch,禁用即断供。本 PR 修的是生产者(把 gate 拆成 always-on 独立入口),不是治标(没有从 preset 行上摘掉 inject 依赖——那才是堵)。env kill-switch 走同一代码路径(apply() 根本不执行),一并覆盖。方向成立。

② 回归审查(逐项过既有行为)

既有行为 结论
正常启动(插件启用) 保持。verdict 在 apply() 的 plan.mode === "off" 早退之后声明(src/agent/dsh-native.ts),启用 ⇒ transform 生效 ⇒ preset 行照旧被禁用,行为等价(已核代码位置)
kill-switch / 禁用 有意变更(PR 正文表已披露):fail-safe brick → stock fallback。#1729 wire 层拒绝守卫不动,是兜底
host 平面 compaction-basic auto: false 行 + #2605 retryPolicy 行 保持,两个载体(bundle yml + launcher overlay)逐字节未动(已核 src/launcher.ts:3192)
client shim 保持兼容。writeDshClientShimFiles 新增可选第 5 参(:3112);subpath export 与 symlink 仅在 bundle 存在时写(hasGate, :3117)— 不产生 dangling symlink,dshPluginEntry 的 realpath 探测路径不受影响
package.json exports 纯增量(./dsh-compaction-gate);无 version bump,内容分支正确
kernel/ pi-subagents/ wire 格式 / 持久化 零触碰
可见面 1 个 cordis 入口 → 2 个;单独禁用 bili-compaction-gate 会重建旧 brick——已披露并写进 bundle 头注释 + CLIENTS/CONFIGURATION 中英文档,卸载路径指向 dsh.profile.bundles 移除

一个我独立发现的竞态窗口(非阻塞):dsh 对兄弟入口并行 import,gate 若先于 bili-native 的 apply() 落地,preset 会在无 verdict 下解析 ⇒ 即使 bili 启用,行也保持 stock 直到进程重启(dsh 原生自动压缩与 bili 并存)。旧设计输掉这个竞态的表现是 brick,新设计是有界的功能退化,且有 #1729 守卫兜底——可用性优先于 fail-safe,符合 #3128 原则,PR 正文已披露(absent verdict 语义)。可接受。

另核了一处不对称:URL lane 无条件写 gate 的 file:// URL(src/launcher.ts:3182),而 shim lane 有存在性守卫。病态分裂态(dist 缺 gate bundle)下两者结果等价(服务都不可提供 ⇒ 旧式 pending),且 file:// 指向缺失文件是响亮失败而非静默跳过,该状态几乎不可达(同一构建产物)。不改,不立项,记录理由于此。

③④ 代码与测试

代码干净:verdict 通道(globalThis key __BILI_DSH_COMPACTION_VERDICT__)是两个独立 bundle/模块实例间的最小必要桥,注释讲清了为什么;入口拆分后 #3354 的顺序保证(transform 先于 provide)在新入口内原样保留。测试是真测试且双向(#3617 纪律):brick 场景同时断言「服务仍被提供 + provide 晚于 transform 注册 + 配置逐字段原样透传」,正常场景断言精确到行级变更,carrier agreement 把 GATE_INSERT 文本钉死在三个载体(bundle yml / bare lane / URL lane 正则),kill-switch 断言「既不注册也不声明」。本地 398 pass / 0 fail / 0 skip。

CI 状态与一处已修复的红

旧 head e6a87e464 唯一红:test-import-graph——根因是 scripts/check-test-import-graph.mjs 维护了一份 tsup entries 的硬编码镜像(ENTRY_FILES,注释自述 Keep in sync),漏了新文件;check-unused-exports.mjs 有同样的潜在漂移。我独立复现了根因、本地打上同样两行修复并验证两条 gate 通过;提交时发现远端分支已被推进到新 head 1fea506ca,其中已含完全相同的两行修复,故丢弃我的重复提交,未向分支推任何东西。新 head 当前:11 条 e2e lane 全 SUCCESS、PR Build Artifact SUCCESS、主 CI 的 guard 组(plan/lane-guard/kernel-guard/test-import-graph/docs-coverage/unused-exports)写作时在跑。16:05Z 的 ci-gate-plugins.yml failure 是 workflow_dispatch 手动车道(真模型×真 CLI,非必需门禁),不在合并判定内。

结构化八问

  1. 是什么问题:在 dsh 里禁用压缩插件(插件管理器或 env kill-switch)后,四个 web profile preset 永久 pending,新建/恢复会话直接失败,只能手动删包才能恢复。
  2. 什么原因导致的:preset 启动被静态补丁锁在一个服务上,该服务唯一提供者正是被禁用的入口本身——自引用 latch;禁用移除提供者却留住了锁。env kill-switch 因 apply() 不执行而走同一路径。
  3. 影响范围与严重程度:p1。凡禁用过插件(或设 BILLION_CONTEXT_PLUGIN=0 等)的 dsh web 用户,新建+恢复会话完全不可用,直到手动从 dsh.profile.bundles 移除包;影响版本 0.1.199–0.1.201(issue 侧已钉版本)。
  4. 怎么修复的:把 gate 从 bili-native 拆出为独立 always-on 入口 bili-compaction-gate(新 tsup 入口 + 新 subpath export),原子地「注册 transform + 提供服务」——锁永远可满足;transform 是否真正禁用 preset 行改由 bili-native apply() 按进程声明的 verdict 决定(跨 bundle 走 globalThis 通道),无 verdict = 全量 stock 透传。已砖的安装在升级后自愈(补丁层从 bundles 列出的包挂载,与 bili-native 状态无关)。
  5. 是否有回归问题:审过的既有行为见上表——正常启动等价、host 平面行与 一次坏工具参数就让整轮作废:希望把 MALFORMED_RESPONSE 纳入重试名单 #2605 行逐字节不变、shim 向后兼容、exports 纯增量、kernel/wire/持久化零触碰。唯一有意变更(brick→stock fallback)已按规范披露旧值→新值+原因。残余竞态(gate 先落地 ⇒ 进程生命周期内行保持 stock)有界、已披露、有 nudge 死锁无检测:同一 session 交替收到两种结构互斥的请求流(巨型单消息 ↔ 多轮)时不告警 #1729 兜底。
  6. 是否破坏缓存:不涉及——未触碰缓存/重试游标/去重键逻辑;dsh 补丁文件幂等重写,内容除新增 gate 行外不变。
  7. 是否改了设计方向:无转向——See how to integrate the plugin that automatically disables dsh compression. | 自动关闭 dsh 压缩的插件看如何集成 #3354 单入口设计扩展为同载体内的双入口拆分(always-on gate + 运行时 verdict),是既有 transform/服务注册机制的自然延伸;新增 1 个 tsup 入口 + 1 个 npm subpath export。
  8. 是否引入了新的配置:无新 env/config 字段/CLI flag。可见面新增 1 个 cordis 入口(bili-compaction-gate,文档注明不可单独禁用)与 1 个 npm subpath export;无需迁移,砖态安装升级自愈,手动解砖配方已写入文档。

判定:可合 —— 待新 head 1fea506ca 的主 CI guard 组转绿(写作时在跑,其余 lane 已全部绿);已知唯一红(test-import-graph)已在分支上修复并经我本地验证。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

CI 终判(最终 head cf332a20a):全绿。

审核期间分支被 rebase 到当前 master(新 head cf332a20a,基线 6f2b113)。我核过两个 head 的 diff:差异只有 CONFIGURATION.md/.zh-CN.md 里 BILI_RECLAIM_FETCH_PATCH 行的 #3528 共存提示——那是随 rebase 从 master(#3690)带入的文档,PR 自身内容逐字节未变,上面的审查结论原样适用。

最终 head 上的完整 CI 结果(30 success / 14 skipped / 0 failure):

  • 主 CI guard 组全绿:plan、lane-guard、kernel-guard、test-import-graph(此前唯一红,gate 脚本两行修复后转绿)、docs-coverage、unused-exports、config-surface-guard、docs-sync、lock-sync、pi-subagents-guard、version-guard;
  • 全量测试分片全绿:ubuntu/windows × node 22/24 所有 shard(含最慢的 ubuntu-24 shard 1);
  • 11 条 e2e lane + PR Build Artifact 全绿。

mergeable_state: blocked 是人工审批门禁(合并本就 human-only),不是检查项未过。

判定维持:可合。 八问结论与回归审查见上一条回复;无新增发现。

awork and others added 2 commits October 21, 2026 22:07
Disabling or kill-switching the dsh plugin removed the only provider of
biliBasicCompactionPolicy while the bundle patch kept gating all four
web-profile presets on it — every preset stayed pending forever and new/
resumed sessions failed with "Unknown agent preset: standard".

The load-time overlay + gate service now live in their own always-on
cordis entry (bili-compaction-gate), mounted alongside bili-native by
both patch carriers (dsh.bundle.patch.yml, launcher overlay). The entry
registers the transform and provides the service atomically (the #3354
ordering guarantee survives the split); whether it disables the preset
compaction rows is decided per process by a verdict bili-native sets
from apply() — an absent verdict (entry disabled, kill-switched, module
failed to load) leaves every row stock.

Behavior changes:
- disabling/kill-switching bili-native: presets used to pend forever
  (session creation bricked); they now boot with STOCK native compaction
  — compression off, dsh fully usable.
- bili-native module failing to load: previously a deliberate fail-safe
  brick; now falls back to stock with the #1729 wire-level refusal guard
  as backstop (availability over fail-safe-brick).
- new visible cordis entry bili-compaction-gate; disabling IT re-creates
  the pending state (documented in bundle patch header + CLIENTS /
  CONFIGURATION en+zh). Full uninstall: remove billion-context from
  dsh.profile.bundles.

Closes #3662
…rface real coverage (#4147/#3962/#4002)

#4147: the #3354 load-time overlay matched the factory
@deepseek-ai/dsh-agent-preset entry by name, leaving custom and
profile-local presets' nested compaction-basic/command-compact rows
ARMED — a landed dsh checkpoint then destroyed the compression
substrate (#4147 consequence A). The transform now matches
STRUCTURALLY: any entry whose resolved config.plugins nests a
compaction row gets those rows disabled; entries without them keep
their resolved config object identity untouched.

#3962: every rewrite is recorded on the shared globalThis telemetry
channel (covered entries + row counts + factory flag) and the durable
[dsh-client] boot notes append the actual coverage, so bundle-patch
gate rows missing a renamed roster are visible without the host
loader log.

#4002: CLIENTS.md en+zh state that preset sessions have no /compact
menu by design (bili owns compression; no re-enable switch, #3651).

Rebased onto current master (594 commits) — resolves the
dsh-native/issue3354/tsup conflicts by union.
@ranxianglei

Copy link
Copy Markdown
Owner Author

(ai: zhipuai-lb/glm-5.3)

Cluster update pushed to this PR's branch (cf332a20 → 69cf2003, force-with-lease after rebase onto current master 594 commits ahead):

  1. Rebase: union-resolved src/agent/dsh-native.ts (master's Compression conflict banner (bili v0.1.201, node v24.14.0, win32-x64) #3806 gap alarm + conversation-recall imports kept; verdict declaration kept), tests/issue3354-… (import union), tsup.config.ts (both claude-mcp-entry and the new dsh-compaction-gate entry). registerCompactionDisable import dropped from dsh-native — registration lives in the gate entry now.
  2. [Bug][dsh] #3354 Native compression disabled: overlay does not override custom preset → native auto-compression remains | [Bug][dsh] #3354 原生压缩关闭 overlay 不覆盖自定义 preset → 原生自动压缩仍武装 → 压缩基底被毁(被迫换新会话)/上下文溢出循环 #4147 (p1): overlay matches STRUCTURALLY — any entry's nested compaction rows get disabled (custom presets were left armed before; reporter's substrate-destruction path). Entries without compaction rows keep object identity.
  3. dsh.bundle.patch.yml gates four preset ids that may not exist — the miss is only visible in the loader log #3962: globalThis telemetry of actual rewrites; both durable boot notes append real coverage ("Overlay coverage so far: N preset entries rewritten (…)") — a renamed roster no longer hides a patch-row miss in the loader log.
  4. The compact menu item in dsh preset sessions disappears without explanation: disableBasicCompaction requires a user-visi | dsh preset 会话的 compact 菜单项消失无说明:disableBasicCompaction 需要用户可见的提示或开关 #4002: CLIENTS en+zh — no /compact menu in preset sessions is BY DESIGN, no re-enable switch (Model request is connecting directly (uncompressed) | 模型请求正在直连(不压缩) #3651/Remove dsh.allowDshCompaction escape hatch (unconditional wire refusal) | 移除 allowDshCompaction 配置:dsh 原生压缩一律拦截 #3666).

Local: typecheck clean; 5 targeted suites 414/0/0; check:unused-exports (one reasoned waiver, #3754-style) + check:test-import-graph + build clean. Full-suite verdict per AGENTS.md §7.2 = CI test lanes on the pushed head.

PR body updated with the cluster section; Closes #3662/#4147/#3962/#4002.

@github-actions

Copy link
Copy Markdown
Contributor

📦 Merged Build Published

This merge to master was published as 0.0.729-master.9284a88 (rolling master tag).

npm install -g billion-context@0.0.729-master.9284a88  # 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.

ranxianglei pushed a commit that referenced this pull request Oct 23, 2026
… rows (#3962)

dsh.bundle.patch.yml gates four Web presets (preset-standard / preset-ptc /
preset-minimal / preset-cordis) on ids that exist only in the FACTORY roster.
A profile that replaces them with its own ids matches none of the four rows
and dsh logs `patch: entry <id> not found` at startup — buried in the loader
logger, surfaced nowhere else. From the outside the rows read as "working"
when they are silent no-ops.

State the roster-conditionality in the patch file itself (the shipped artifact
users inspect): the miss is NOT harmless — an ungated custom preset resolves
by import order and may win the race against the overlay middleware register,
leaving its compaction-basic AND command-compact rows armed (timing window;
worst case a dsh checkpoint outside bili destroys the compression substrate,
#3354/#1772/#1729). Mitigation: self-gate each custom preset entry with
inject:[biliBasicCompactionPolicy]. Self-check: grep the dsh loader log for
`patch: entry ... not found`. Do-not-delete note kept: factory-roster
profiles get a real ordering fence from these rows.

The launcher overlay carrier (writeDshAcpPatch) mirrors the same clause so the
two carriers stay in sync (file's own discipline + #3354 carrier-agreement
test). Comments/string-literal only — zero behavior change: no functional
patch row is added, removed, or reworded.

Rebased onto master after #3729 landed (gate moved to its own always-on
bili-compaction-gate entry; structural matching + coverage telemetry shipped
there). The four gate rows are unchanged on master, and structural matching
widens WHICH entries the transform reaches but not WHEN the middleware
registers relative to an ungated preset's resolution — the timing window this
note documents persists, so the text needs no rewording.

Targeted suites green locally: issue3354-dsh-compaction-disable (incl.
two-carriers-agree), issue2605-dsh-retry-patch, issue4147-dsh-overlay-structural
(16/16). Full typecheck/test on CI.

References #3962
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment