cross-tab-worker-databus 0.21.20 → 0.21.22

package/docs/roadmap.md

@@ -1,6 +1,22 @@
 # Roadmap
 
-0.21.20 was released on September 24, 2026. The project is intentionally continuing through reliability-focused releases before a 1.0.0 stability freeze.
+0.21.22 was released on September 24, 2026. The project is intentionally continuing through reliability-focused releases before a 1.0.0 stability freeze.
+
+## 0.21.22 delivered scope
+
+- **One method, applied to every multi-term guard the previous scan ranked**: delete a single operand, run the whole suite, and classify which of three answers that mutant gave — pinned, incapable of deciding, or live and unnamed. Nine behaviors are now named that no test had named (three socket-staleness terms in `websocket.ts`, two of `trace.ts`'s four dedup/activity counters, two operands of the `storage-event` envelope, two option legs of `DedupManager.start()`'s sweep gate), and the suite went 923 → 929.
+- **The ledger did not move, and that is the result.** `src/` holds the same **47** zero-count branch arms at the end of six phases as at the start, with only the denominator changing (1962 → 1958, from two dominated operands being deleted). Every operand probed is evaluated on every call, so no coverage total distinguishes "runs" from "decides" — the class is invisible to the arm ledger and visible only to a whole-suite mutant.
+- **A joint assertion was the recurring cause, not an absent test.** `trace.ts`'s activity check looked pinned because one existing case raised *both* dedup counters in the same window, so whichever operand survived still emitted; the storage envelope's `seq` and `message` tests were both tripped by its single malformed write. Splitting the inputs, not adding coverage, is what made each leg die on its own.
+- **Five operands closed as incapable of deciding anything, by differential.** Two in the envelope and all three `typeof value !== 'number'` tests in `validation.ts` were proven rather than argued: a fixed input vector (23 stored payloads; 23 values × 3 validators × 4 variants) produced byte-identical accept/reject-and-message results for the guard as written and for each mutant. The kept `typeof` leg then had its own justification tested both ways — construct the coercing-global swap it exists to catch, and observe that it kills nothing while the leg is present and turns `heartbeatIntervalMs: '3000'` into an accepted option once the leg is gone.
+- **Two dominated operands deleted, and a clean bill plus a thin pin recorded instead of averaged away.** `getMetrics()` and `flush()` each re-read a field the same expression had just computed, which cannot decide anything and was silently absorbing the mutation that would have proved the getter's contract; deleted, with both kill sets in the comment. `ReplayManager.start()`'s four-term gate turned out fully pinned all along — the parity check between two same-shaped managers, which is exactly where `DedupManager`'s gap was. And five of `validation.ts`'s six surviving-leg deaths come from several cases while the sixth comes from one, so the site says so instead of reporting "six of nine".
+
+## 0.21.21 delivered scope
+
+- **One more branch arm closed by a call no test had made** — and this one needed a caller that does not exist in a browser. `queueStartAfterStop()`'s `stopPromise ?? Promise.resolve()` had never taken its fallback, which is the state a *failed initial open* lives in (its teardown is owned by `pendingStop`, so no gate exists); the only synchronous application-code seam inside that window is the caller's own `ClusterEnvironment` port, so the case wraps the port and calls back into `start()` from a storage write. `src/` goes from 48 zero arms to 47.
+- **Which half of that expression has teeth.** Not the leg the test was written for. The fallback operand is consequence-free — `start()` chains the reopen behind `pendingStop` independently, so the measured `start` → `stop` → `start` order survives deleting the read entirely. What the case does pin is the routing one line above it: `if (this.stopping)` → `if (this.stopPromise)` reddens it, and 192 of the file's other tests stay green doing it.
+- **A candidate test written, measured, and deleted.** A second case queued a restart behind a transport whose `stop()` was held open, which is the obvious pin for the gate operand — and it passed under all four single-leg mutations tried. A bounded number of awaited microtasks cannot see a wait whose wakeup needs the timer queue, so the test was removed and its zero kill rate recorded at the site, which is the finding; the number 194/194 under deletion is the statement, not a gap waiting for a better test.
+- **The guard whose deletion is a crash rather than a failed assertion.** Removing that read turns the restart's "wake, discover we are still stopping, re-queue" into a microtask busy-wait that starves the macrotask the transport stop needs: `tests/lifecycle-invariants.test.ts` aborts its worker ~40 s later inside `Builtins_RunMicrotasks` with an out-of-memory, and every test in `data-bus` still passes. The site comment now says the fuzz is that line's only witness, which is also a warning to anyone who trims its stop/restart interleavings for speed.
+- **Three shipped claims retracted against the code, in the two ways claims go stale.** A count that had drifted (the `?? this.currentRecord` fallbacks are three, not "the identical pair", and the third carried no site text at all), a measurement that no longer reproduces (the reopen absorb's "fails exactly one test" is 194/194, 194/194, 51-of-194 and 194/194 across the four experiments now recorded in its place), and eleven line-number citations into our own files of which six pointed at the wrong place — two of them into another comment. Every position is now a symbol or a condition, the surviving numbers are the dependency's (`centrifuge@5.7.4`, re-verified to its last `emit('error')`), and the whole prose change is proven code-free by comparing esbuild's emitted JavaScript against the previous revision.
 
 ## 0.21.20 delivered scope
 

package/docs/zh/benchmarks.md

@@ -2,20 +2,20 @@
 
 # 浏览器基准趋势
 
-> 数据截至 2026-09-23，基于 87 份归档的 `bench-results/browser-*.json` 报告（运行 `pnpm bench:browser` 追加一份；用 `node scripts/bench-trend.mjs` 重新生成本文档）。
+> 数据截至 2026-09-24，基于 91 份归档的 `bench-results/browser-*.json` 报告（运行 `pnpm bench:browser` 追加一份；用 `node scripts/bench-trend.mjs` 重新生成本文档）。
 
 发布门禁的对比基线是 `pnpm bench:compare --fail-above-pct 50`：将最新报告与其之前若干报告中同一指标的中位数比较，*并且*要求新值超过这些样本中的最大值（50% 上限用于吸收共享 runner 的噪声；此前只看"上一份报告"并不稳定，因为页内热路径指标在同一份代码上会在快/慢两档之间来回跳；而在两档交替的样本上，中位数只有在这两档混合分布时才稳定，因此单看百分比仍可能纯粹因为档位切换而失败）。被第二个条件放过的指标会显式列在 “within-baseline, not gated” 里，而不是悄悄算作通过。本文记录长期趋势：数值为逐指标延迟，越低越好；最后一列是最近 5 份报告内的最优值，而不是历史纪录——页内基准矩阵在 2026 年 9 月初变更过测量语义，更早的报告不可比。
 
 <!-- BENCH-TREND:BEGIN (machine-generated table) -->
 | 指标 | 上次 (ms) | 本次 (ms) | Δ | 近 5 次最优 (ms) |
 |---|---|---|---|---|
-| publish per-message (ms, lower is better) — dedicated | 38.4709 | 36.5875 | -1.88 | 36.5875 |
-| publish per-message (ms, lower is better) — shared | 33.9845 | 33.7984 | -0.19 | 33.7984 |
-| wildcard dispatch ×1000 (ms, lower is better) | 5.5 | 5.7 | +0.20 | 5.5 |
-| publishBatch ×1000 (ms, lower is better) | 3.5 | 3.6 | +0.10 | 3.5 |
-| dedup ×1000 (ms, lower is better) | 29.2 | 11 | -18.20 | 11 |
-| trace + publish ×1000 (ms, lower is better) | 7 | 4.6 | -2.40 | 4.6 |
-| first-packet cold dispatch (ms, lower is better) | 0.1 | 0 | -0.10 | 0 |
+| publish per-message (ms, lower is better) — dedicated | 40.1172 | 41.3311 | +1.21 | 36.5875 |
+| publish per-message (ms, lower is better) — shared | 33.9868 | 34.1115 | +0.12 | 33.7984 |
+| wildcard dispatch ×1000 (ms, lower is better) | 6.5 | 6.7 | +0.20 | 5.7 |
+| publishBatch ×1000 (ms, lower is better) | 4.6 | 4.4 | -0.20 | 3.6 |
+| dedup ×1000 (ms, lower is better) | 15.6 | 19 | +3.40 | 11 |
+| trace + publish ×1000 (ms, lower is better) | 4.6 | 6.6 | +2.00 | 4.6 |
+| first-packet cold dispatch (ms, lower is better) | 0 | 0 | +0.00 | 0 |
 <!-- BENCH-TREND:END -->
 
 说明：

package/docs/zh/roadmap.md

@@ -1,6 +1,22 @@
 # 路线图
 
-0.21.20 已于 2026 年 9 月 24 日发布。项目会先持续完成可靠性发布，再进入 1.0.0 稳定性冻结。
+0.21.22 已于 2026 年 9 月 24 日发布。项目会先持续完成可靠性发布，再进入 1.0.0 稳定性冻结。
+
+## 0.21.22 已完成范围
+
+- **一套方法，用上一轮扫描排出的每一个多操作数守卫**：只删一个操作数，跑整个测试套件，然后判断这个变异给出了三种答案里的哪一种——被钉住了、根本没有裁决能力、活着却没有任何测试命名它。于是九处从未被测试命名的行为有了名字（`websocket.ts` 里三个判断连接是否过期的项、`trace.ts` 四个活跃/去重计数器中的两个、`storage-event` 信封的两个操作数、`DedupManager.start()` 清扫闸门的两条选项腿），套件从 923 增至 929。
+- **账本一次都没有动，而这正是结论。** 六个阶段结束时 `src/` 仍然是 **47** 条零计数分支，与开始时相同，变的只有分母（1962 → 1958，来自删掉的两条被支配操作数）。被探测的每一个操作数都在每次调用里求值，所以没有任何覆盖率总量能区分“执行了”和“裁决了”——这一类问题对分支账本是隐形的，只有整套件变异能看见。
+- **反复出现的原因是联合断言，不是缺测试。** `trace.ts` 的活跃性检查看上去早已被钉住，因为那个既有用例在同一个窗口里把*两个*去重计数器都抬了起来，于是留下哪一条腿都照样产出；信封的 `seq` 与 `message` 两条测试也同时被它唯一那条畸形写入触发。让每条腿各自致死的，是把输入拆开，而不是补覆盖率。
+- **五条操作数被判定为“无法裁决任何东西”，而且是差分证明的。** 信封里的两条，加上 `validation.ts` 三条 `typeof value !== 'number'`，全部以固定的输入向量证明而非推论（23 个存储载荷；23 个值 × 3 个校验器 × 4 个变体），原写法与每个变异体得到的“接受/拒绝及其消息”结果逐字节相同。被保留的那条 `typeof` 腿随后也被两头检验了自己的保留理由——把它存在的意义就是防住的那个“换成会强制转换的全局函数”的改写构造出来，结果带着这条腿时该改写杀不死任何东西，而把腿也删掉后 `heartbeatIntervalMs: '3000'` 就成了被接受的选项。
+- **删掉两条被支配操作数，另外记下一次的满分和一次的薄。** `getMetrics()` 与 `flush()` 各自把同一个表达式刚算过的字段又读了一遍，这种读取不可能裁决任何东西，却正在悄悄吞掉那个本可以证明 getter 契约的变异；已删除，两组击杀集都写在注释里。`ReplayManager.start()` 的四项闸门则被证实每一条腿早已被钉住——这是两个同形管理器之间的对等检查，也正是 `DedupManager` 缺口的所在。而 `validation.ts` 六条致死里有五条来自多个用例、第六条只来自一个，所以代码位置上如实写明，而不是汇成一句“九之六”。
+
+## 0.21.21 已完成范围
+
+- **又一条未覆盖分支被“从来没有测试发出过的调用”关闭了**——而这一次需要的调用方在浏览器里并不存在。`queueStartAfterStop()` 的 `stopPromise ?? Promise.resolve()` 从未走过它的兜底分支，而那正是*首次打开失败*时所处的状态（它的清理由 `pendingStop` 掌管，所以根本不存在那道闩）；那个窗口内唯一会同步回调进应用代码的缝隙，是调用方自己提供的 `ClusterEnvironment` 端口，于是用例把端口包了一层，从一次存储写入里回调进 `start()`。`src/` 的零计数分支从 48 条降到 47 条。
+- **那个表达式里真正有后果的是哪一半。** 不是这条测试当初为它而写的那一半。兜底操作数是没有后果的——`start()` 自己就会把重开挂在 `pendingStop` 之后，所以把整个读取删掉，实测的 `start` → `stop` → `start` 顺序依然成立。这条用例真正钉住的是它上面一行的路由：把 `if (this.stopping)` 改成 `if (this.stopPromise)` 会让它变红，而文件里其余 192 个测试对此毫无反应。
+- **一个候选测试被写出、被测量、然后被删除。** 第二个用例把重开排在一个 `stop()` 被挂住的传输后面，看起来正是那个闩的显然钉法——可它在四种单腿变异下全部通过。有限次数的 await 微任务看不见一个必须等定时器队列的等待，于是这个测试被删掉，而它的零击杀率写在了代码位置上；“删掉它 194/194 依旧全绿”这个数字本身就是结论，而不是一个等着更好测试来填的空洞。
+- **有一条守卫，删除它的表现是崩溃，不是断言失败。** 把那个读取删掉，重开的“醒来—发现还在 stopping—重新排队”就变成一个微任务忙等，饿死传输停止所依赖的那个宏任务：`tests/lifecycle-invariants.test.ts` 在约 40 秒后于 `Builtins_RunMicrotasks` 内因内存耗尽中止自己的 worker，而 `data-bus` 里的每个测试仍然全绿。位置注释现在写明：那条 fuzz 就是这一行的唯一见证——这也是对任何为了提速而删掉它 stop/restart 交错序列的人的警告。
+- **三处随包发布的声明被按代码撤回，正好对应声明腐坏的两种方式。** 一个计数漂了（`?? this.currentRecord` 兜底有三处，不是“那相同的一对”，而且第三处压根没有任何位置说明）；一个测量不再复现（重开兜底那句“只失败一个测试”，被现在写在它原地的四个实验取代：194/194、194/194、51/194、194/194）；以及十一处指向本仓库自己文件的 `:NNN` 行号引用里有六处指向了别处——其中两处落进了*另一段注释*。所有位置引用现在都改成符号或条件；保留下来的编号是依赖的（`centrifuge@5.7.4`，连它最后一个 `emit('error')` 都重新核对过）；而整场纯文字改动是靠比对 esbuild 产出的 JavaScript 与上一个 revision 来证明没有碰到代码的。
 
 ## 0.21.20 已完成范围
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cross-tab-worker-databus",
-  "version": "0.21.20",
+  "version": "0.21.22",
   "description": "Framework-agnostic cross-tab data bus with Dedicated/Shared Worker clustering and Centrifuge support.",
   "type": "module",
   "license": "MIT",