mixdog 1.4.1 → 1.4.3

package/scripts/lib/gate-scheduler.mjs

@@ -1,42 +1,149 @@
+import { randomUUID } from 'node:crypto';
+import { readFileSync, renameSync, rmSync, writeFileSync } from 'node:fs';
 import { availableParallelism } from 'node:os';
-import { testConcurrencyArg } from './run-node-tests.mjs';
 
-// A local gate has one budget, not a full machine's worth of workers per lane.
-// Six is the interactive-desktop default; an explicit budget still wins.
-export function gateTestBudget(env = process.env, parallelism = availableParallelism()) {
-  const argument = testConcurrencyArg([], env);
-  return argument ? Number(argument.split('=')[1]) : Math.min(6, parallelism);
+// The local gate runs every lane from one shared pool of workers: each lane's
+// files are cut into shards, and whichever worker is free takes the longest
+// shard left, whatever its lane. Shard sizes come from the previous run's
+// measured per-file times, so the split stays balanced without hand-kept
+// weights, and no worker idles while another lane still has work.
+
+// The gate runs its whole process tree at below-normal priority, so a
+// foreground app always wins the CPU. Priority does not bound what the
+// desktop shell shares with every process (handles, desktop heap, work
+// queues; explorer.exe has failed with "no work queue available" under many
+// concurrent test processes), so the default also caps the worker count:
+// half the cores, leaving the other half for what the budget does not count
+// (each shard's runner, the processes tests spawn, the desktop build's own
+// threads), and never more than DEFAULT_MAX_WORKERS. Each worker is three
+// node processes (shard runner, its node --test batch, the test file) plus
+// whatever the test spawns. Measured on a 20-core Windows PC with other work
+// running: at 10 workers, test-file processes began dying within ~20 ms with
+// no output (every remaining file of a shard, in 2 of 2 runs); at 8 and 6
+// they did not.
+// An explicit budget still wins: --test-concurrency (gate argument), then
+// MIXDOG_TEST_CONCURRENCY.
+const DEFAULT_MAX_WORKERS = 8;
+export function gateTestBudget(argv = [], env = process.env, parallelism = availableParallelism()) {
+  const index = argv.findIndex((arg) => arg === '--test-concurrency' || arg.startsWith('--test-concurrency='));
+  let explicit = env.MIXDOG_TEST_CONCURRENCY || undefined;
+  // A flag without a value is an error, not "unspecified".
+  if (index >= 0) explicit = argv[index].includes('=') ? argv[index].split('=')[1] : (argv[index + 1] ?? '');
+  if (explicit === undefined) return Math.max(1, Math.min(DEFAULT_MAX_WORKERS, Math.floor(parallelism / 2)));
+  const budget = Number(explicit);
+  if (!Number.isInteger(budget) || budget < 1)
+    throw new Error(`gate test budget must be a positive integer (got "${explicit}")`);
+  return budget;
+}
+
+// Estimates for work with no recorded time yet: the next run records it.
+export const DEFAULT_ESTIMATE_MS = { fast: 2_000, slow: 25_000, command: 60_000 };
+
+// Previous runs' measured times ({ version, ms: { key: ms } }). A missing,
+// unreadable or malformed file counts as no history, and an entry that is
+// not a finite positive number is dropped (its work gets the default).
+export function readTimingHistory(file) {
+  let ms;
+  try {
+    ms = JSON.parse(readFileSync(file, 'utf8'))?.ms;
+  } catch {
+    return {};
+  }
+  if (!ms || typeof ms !== 'object' || Array.isArray(ms)) return {};
+  return Object.fromEntries(
+    Object.entries(ms).filter(([, value]) => typeof value === 'number' && Number.isFinite(value) && value > 0)
+  );
 }
 
-// Splits the budget across legs in proportion to leg.weight (default 1), at least
-// one worker each. Spare workers go one at a time to the leg with the highest
-// weight per worker, so longer legs get more. Needs budget >= legs.length.
-export function allocateWorkers(legs, budget) {
-  const shares = legs.map(() => 1);
-  for (let spare = budget - legs.length; spare > 0; spare--) {
-    let best = 0;
-    for (let i = 1; i < legs.length; i++) {
-      if ((legs[i].weight ?? 1) / shares[i] > (legs[best].weight ?? 1) / shares[best]) best = i;
+// Writes the history through a temp file of its own (concurrent gates in
+// other worktrees share the file) and renames it into place. Never throws:
+// returns the error, or null once saved; the timings only plan the next run.
+export function saveTimingHistory(file, ms) {
+  const temp = `${file}.${process.pid}.${randomUUID()}.tmp`;
+  try {
+    writeFileSync(temp, `${JSON.stringify({ version: 1, ms })}\n`);
+    renameSync(temp, file);
+    return null;
+  } catch (error) {
+    try {
+      rmSync(temp, { force: true });
+    } catch {
+      // The save error is the one reported.
     }
-    shares[best]++;
+    return error;
   }
-  return shares;
 }
+// Enough shards per worker that the pool's tail (the shards still running
+// when the queue is empty) stays short; the floor bounds the per-shard runner
+// start-up and process sweep (~1-2 s each).
+const SHARDS_PER_WORKER = 8;
+const MIN_SHARD_MS = 15_000;
 
-// The test runners read their worker count when spawned, so it cannot grow
-// mid-run: shares are static. With fewer workers than legs, each worker runs
-// legs one after another.
-export async function runGateLegs(legs, budget, runLeg) {
-  if (!Number.isInteger(budget) || budget < 1) throw new Error('gate test budget must be a positive integer');
-  if (budget < legs.length) {
-    let next = 0;
-    await Promise.all(
-      Array.from({ length: budget }, async () => {
-        while (next < legs.length) await runLeg(legs[next++], 1);
-      })
-    );
-    return;
+// test.mjs selects every lane file whose path contains a given filter, so a
+// file whose path contains another file's path has to share that file's shard
+// or it would run twice. Returns the indivisible groups.
+export function filterGroups(files) {
+  const parent = new Map(files.map((file) => [file, file]));
+  const find = (file) => (parent.get(file) === file ? file : find(parent.get(file)));
+  for (const inner of files) {
+    for (const outer of files) if (outer !== inner && outer.includes(inner)) parent.set(find(outer), find(inner));
   }
-  const shares = allocateWorkers(legs, budget);
-  await Promise.all(legs.map((leg, index) => runLeg(leg, shares[index])));
+  const groups = new Map();
+  for (const file of files) groups.set(find(file), [...(groups.get(find(file)) ?? []), file]);
+  return [...groups.values()];
+}
+
+// Cuts each lane ({ name, files, estimateOf }) into shards of about one
+// target size, longest-first into the least-loaded shard, so a file longer
+// than the target runs alone. Returns [{ lane, files, estimate }].
+export function planShards(lanes, budget) {
+  const total = lanes.reduce((sum, lane) => sum + lane.files.reduce((s, file) => s + lane.estimateOf(file), 0), 0);
+  const target = Math.max(MIN_SHARD_MS, total / (budget * SHARDS_PER_WORKER));
+  const shards = [];
+  for (const lane of lanes) {
+    const items = filterGroups(lane.files)
+      .map((files) => ({ files, estimate: files.reduce((s, file) => s + lane.estimateOf(file), 0) }))
+      .sort((a, b) => b.estimate - a.estimate);
+    const laneTotal = items.reduce((sum, item) => sum + item.estimate, 0);
+    const bins = Array.from({ length: Math.min(items.length, Math.max(1, Math.round(laneTotal / target))) }, () => ({
+      lane: lane.name,
+      files: [],
+      estimate: 0,
+    }));
+    for (const item of items) {
+      const bin = bins.reduce((least, candidate) => (candidate.estimate < least.estimate ? candidate : least));
+      bin.files.push(...item.files);
+      bin.estimate += item.estimate;
+    }
+    shards.push(...bins);
+  }
+  return shards;
+}
+
+// Per-file time of a finished shard: what its own tests reported plus an
+// equal share of the rest of the shard's wall time (process start-up, module
+// load, the runner's own work). Returns { file: ms }.
+export function measuredFileTimes(files, reportedMs, wallMs) {
+  const reported = files.reduce((sum, file) => sum + (reportedMs[file] ?? 0), 0);
+  const overhead = Math.max(0, wallMs - reported) / files.length;
+  return Object.fromEntries(files.map((file) => [file, (reportedMs[file] ?? 0) + overhead]));
+}
+
+// Runs every unit on `budget` workers, longest estimate first. A failed unit
+// marked stopOnFail stops new starts (running units finish). Returns the
+// units never started.
+export async function runPool(units, budget, run) {
+  if (!Number.isInteger(budget) || budget < 1) throw new Error('gate test budget must be a positive integer');
+  const queue = [...units].sort((a, b) => b.estimate - a.estimate);
+  let stopped = false;
+  await Promise.all(
+    Array.from({ length: Math.min(budget, queue.length) }, async () => {
+      while (!stopped && queue.length > 0) {
+        const unit = queue.shift();
+        unit.ok = await run(unit);
+        if (!unit.ok && unit.stopOnFail) stopped = true;
+      }
+    })
+  );
+  return queue;
 }

package/scripts/lib/gate-timing-reporter.mjs

@@ -0,0 +1,16 @@
+// node:test reporter the local gate adds to each shard it runs: one JSON line
+// per test file, { file, ms }, the file's top-level test time (relative to the
+// package the run started in, like `scripts/test.mjs --list`). The gate turns
+// these into the per-file estimates its next run plans shards from.
+import { relative } from 'node:path';
+
+export default async function* gateTimingReporter(source) {
+  const byFile = new Map();
+  for await (const event of source) {
+    if (event.type !== 'test:pass' && event.type !== 'test:fail') continue;
+    if (event.data.nesting !== 0 || !event.data.file) continue;
+    const file = relative(process.cwd(), String(event.data.file)).replaceAll('\\', '/');
+    byFile.set(file, (byFile.get(file) ?? 0) + (Number(event.data.details?.duration_ms) || 0));
+  }
+  for (const [file, ms] of byFile) yield `${JSON.stringify({ file, ms })}\n`;
+}

package/scripts/lib/run-node-tests.mjs

@@ -29,6 +29,8 @@ const TEST_LOG_PREFIX = 'mixdog-test-output-';
 const TEST_FAILURES_PREFIX = 'mixdog-test-failures-';
 const TEST_SCRATCH_PREFIX = 'mixdog-test-scratch-';
 export const RUNNER_TEMP_PREFIXES = [TEST_LOG_PREFIX, TEST_FAILURES_PREFIX, TEST_SCRATCH_PREFIX];
+// Printed first by every run; the local gate finds a dead unit's log by it.
+export const FULL_LOG_LABEL = 'Full test log: ';
 
 // Batches keep the file order Node was given; a single path larger than the
 // budget still gets its own batch rather than being dropped. An empty list
@@ -82,7 +84,8 @@ const concurrencyFlag = (arg) => arg === '--test-concurrency' || arg.startsWith(
 function concurrencyOf(runArgs, fileArgs) {
   const args = [...runArgs, ...fileArgs];
   const index = args.findIndex(concurrencyFlag);
-  const value = index < 0 ? null : positiveInt(args[index].includes('=') ? args[index].split('=')[1] : args[index + 1]);
+  let value = null;
+  if (index >= 0) value = positiveInt(args[index].includes('=') ? args[index].split('=')[1] : args[index + 1]);
   return value ?? Math.max(1, availableParallelism() - 1);
 }
 
@@ -141,11 +144,19 @@ function capture(command, args) {
   });
 }
 
-// Kills a process and every descendant still linked to it. On POSIX a batch
-// leads its own process group (spawned detached), so the group goes with it.
+// Kills a live batch and every descendant still linked to it (on Windows the
+// shared creation-checked tree kill). On POSIX a batch leads its own process
+// group (spawned detached), so the group goes with it.
+// Loaded only when a Windows kill runs, so importers that need just this
+// module's constants (the local gate) do not pull in the runtime.
+async function loadKillProcessTrees() {
+  const { killProcessTrees } = await import('../../src/runtime/agent/orchestrator/mcp/child-tree.mjs');
+  return killProcessTrees;
+}
+
 async function killTree(pid) {
   if (process.platform === 'win32') {
-    await capture('taskkill', ['/T', '/F', '/PID', String(pid)]);
+    await (await loadKillProcessTrees())([pid]);
     return;
   }
   for (const target of [-pid, pid]) {
@@ -189,17 +200,29 @@ async function processSnapshot() {
     });
 }
 
-const CLOCK_SLACK_MS = 2000;
+// How far a snapshot's creation time may sit from Date.now() read around the
+// spawn. Windows reports creation to the microsecond from the same system
+// clock, so only a clock tick separates them; POSIX ps reports whole elapsed
+// seconds. No wider: Windows hands a freed pid to the next process within
+// milliseconds, and a 2 s window made a process another run had just started
+// pass for one of this run's exited spawns, so it was killed with them.
+const clockSlackMs = (platform) => (platform === 'win32' ? 50 : 2000);
 
 // The processes this run owns, out of a snapshot: an anchor (the batch, or a
 // child a test process recorded in its spawn ledger) still alive under its
-// pid, the orphans an exited anchor left (they still name it as their parent;
-// a later process that reuses its pid only has younger children), every
-// descendant of those, and anything executing from the run's scratch root
-// (e.g. the postgres the memory runtime downloads into MIXDOG_DATA_DIR).
+// pid, the orphans an exited anchor left (they still name it as their parent),
+// every descendant of those, and anything executing from the run's scratch
+// root (e.g. the postgres the memory runtime downloads into MIXDOG_DATA_DIR).
+// An anchor is { pid, before, since, exited? }: the clock just before and just
+// after its spawn, and when its exit was seen. A pid alone never identifies a
+// process: an anchor seen to exit no longer holds its pid, so whatever holds
+// it now is another process; a live holder was created within its spawn; an
+// orphan was created while its anchor lived (a later process that reuses the
+// pid only has younger children); a descendant is younger than its parent.
 // Command-line text never makes a process the run's: an editor or terminal
 // opened on the scratch root is not.
-export function ownedProcesses(snapshot, anchors, { batchStart, scratchDir, platform = process.platform }) {
+export function ownedProcesses(snapshot, anchors, { scratchDir, platform = process.platform }) {
+  const slack = clockSlackMs(platform);
   const byPid = new Map(snapshot.map((entry) => [entry.pid, entry]));
   const childrenOf = new Map();
   for (const entry of snapshot) childrenOf.set(entry.ppid, [...(childrenOf.get(entry.ppid) ?? []), entry]);
@@ -210,14 +233,19 @@ export function ownedProcesses(snapshot, anchors, { batchStart, scratchDir, plat
     owned.set(entry.pid, entry);
     queue.push(entry);
   };
-  for (const { pid, since } of anchors) {
+  for (const { pid, before, since, exited } of anchors) {
     const holder = byPid.get(pid);
-    if (holder && holder.created >= batchStart - CLOCK_SLACK_MS && holder.created <= since + CLOCK_SLACK_MS) {
+    if (holder && exited === undefined && holder.created >= before - slack && holder.created <= since + slack) {
       claim(holder);
       continue;
     }
     for (const orphan of childrenOf.get(pid) ?? []) {
-      if (orphan.created >= since - CLOCK_SLACK_MS && (!holder || orphan.created < holder.created)) claim(orphan);
+      if (
+        orphan.created >= before - slack &&
+        orphan.created <= (exited ?? Number.POSITIVE_INFINITY) + slack &&
+        (!holder || orphan.created < holder.created)
+      )
+        claim(orphan);
     }
   }
   const fold = (path) => (platform === 'win32' ? path.toLowerCase() : path);
@@ -229,26 +257,30 @@ export function ownedProcesses(snapshot, anchors, { batchStart, scratchDir, plat
   while (queue.length > 0) {
     const parent = queue.pop();
     for (const entry of childrenOf.get(parent.pid) ?? []) {
-      if (entry.created >= parent.created - CLOCK_SLACK_MS) claim(entry);
+      if (entry.created >= parent.created - slack) claim(entry);
     }
   }
   return [...owned.values()];
 }
 
+// One anchor per recorded spawn (scripts/lib/test-spawn-ledger.mjs): its
+// "<pid> <before> <since>" line, with "<exited>" from the line that repeats
+// it once the exit was seen.
 async function readSpawnLedger(path) {
   const text = await readFile(path, 'utf8').catch(() => '');
-  return text
-    .split('\n')
-    .filter(Boolean)
-    .map((line) => {
-      const [pid, since] = line.split(' ').map(Number);
-      return { pid, since };
-    });
+  const spawns = new Map();
+  for (const line of text.split('\n').filter(Boolean)) {
+    const [pid, before, since, exited] = line.split(' ').map(Number);
+    const key = `${pid} ${before} ${since}`;
+    spawns.set(key, { ...(spawns.get(key) ?? { pid, before, since }), ...(exited ? { exited } : {}) });
+  }
+  return [...spawns.values()];
 }
 
 // A process a test leaves running must not outlive its batch: it can hold a
 // port, a database, or files in the scratch root the run removes at the end.
-async function reapLeftovers(child, { batchStart, ledger, scratchDir }) {
+// `batch` is the batch's own anchor, its exit already seen.
+async function reapLeftovers(child, { batch, ledger, scratchDir }) {
   // The batch itself has exited: only its process group can remain.
   if (process.platform !== 'win32' && child.pid) {
     try {
@@ -257,17 +289,19 @@ async function reapLeftovers(child, { batchStart, ledger, scratchDir }) {
       /* the group is empty */
     }
   }
-  const anchors = [...(child.pid ? [{ pid: child.pid, since: batchStart }] : []), ...(await readSpawnLedger(ledger))];
-  await killOwned(anchors, { batchStart, scratchDir });
+  const anchors = [...(child.pid ? [batch] : []), ...(await readSpawnLedger(ledger))];
+  await killOwned(anchors, { scratchDir });
 }
 
-async function killOwned(anchors, { batchStart, scratchDir }) {
-  const leftovers = ownedProcesses(await processSnapshot(), anchors, { batchStart, scratchDir });
+async function killOwned(anchors, { scratchDir }) {
+  // Every owned process ran when the snapshot was taken.
+  const snapshotAt = Date.now();
+  const leftovers = ownedProcesses(await processSnapshot(), anchors, { scratchDir });
   if (leftovers.length === 0) return;
   const pids = leftovers.map((entry) => entry.pid);
   console.error(`Killing ${pids.length} process(es) a test left running: ${pids.join(' ')}`);
   if (process.platform === 'win32') {
-    await capture('taskkill', ['/T', '/F', ...pids.flatMap((pid) => ['/PID', String(pid)])]);
+    await (await loadKillProcessTrees())(pids, { aliveAt: snapshotAt });
     return;
   }
   for (const pid of pids) {
@@ -279,12 +313,13 @@ async function killOwned(anchors, { batchStart, scratchDir }) {
   }
 }
 
-// Resolves { code, hung } once the batch exited and its leftovers are gone.
+// Resolves { code, hung, pid } once the batch exited and its leftovers are gone.
 // `scratchDir` '' = leave the scratch-root sweep to the caller (side-by-side batches).
 function spawnBatch(args, env, { files, ledger, scratchDir, timeoutMs }) {
   return new Promise((resolve) => {
-    const batchStart = Date.now();
+    const before = Date.now();
     const child = spawn(process.execPath, args, { stdio: 'inherit', env, detached: process.platform !== 'win32' });
+    const since = Date.now();
     let hung = '';
     const watchdog = setTimeout(() => {
       hung = `hung after ${minutes(timeoutMs)} min: ${files.join(' ') || '(discovered files)'}`;
@@ -302,10 +337,11 @@ function spawnBatch(args, env, { files, ledger, scratchDir, timeoutMs }) {
     const signals = process.platform === 'win32' ? [] : ['SIGINT', 'SIGTERM'];
     for (const signal of signals) process.on(signal, forward);
     const settle = async (code) => {
+      const batch = { pid: child.pid, before, since, exited: Date.now() };
       clearTimeout(watchdog);
       for (const signal of signals) process.off(signal, forward);
-      await reapLeftovers(child, { batchStart, ledger, scratchDir });
-      resolve({ code: hung ? 1 : code, hung });
+      await reapLeftovers(child, { batch, ledger, scratchDir });
+      resolve({ code: hung ? 1 : code, hung, pid: child.pid });
     };
     child.on('error', (error) => {
       console.error(error);
@@ -363,7 +399,7 @@ export async function runNodeTests(nodeArgs, fileArgs, options = {}) {
   pruneStaleTempDirs(RUNNER_TEMP_PREFIXES);
   const outputDir = await mkdtemp(join(tmpdir(), TEST_LOG_PREFIX));
   const logPath = join(outputDir, 'full.log');
-  console.error(`Full test log: ${logPath}`);
+  console.error(`${FULL_LOG_LABEL}${logPath}`);
   // Failure records live outside the log directory: the log directory holds
   // exactly one merged log, and these records are an internal artifact.
   const failureDir = await mkdtemp(join(tmpdir(), TEST_FAILURES_PREFIX));
@@ -461,7 +497,7 @@ async function executeRun(
   );
   // Peers were still running while each batch exited, so the scratch root's
   // own processes (e.g. a postgres a test started) are swept once, here.
-  if (overlapped) await killOwned([], { batchStart: Date.now(), scratchDir });
+  if (overlapped) await killOwned([], { scratchDir });
   // Logs and records merge in batch order, whatever order the batches finished.
   for (const [index, outcome] of outcomes.entries()) {
     const batchLog = batchLogOf(index);
@@ -518,6 +554,7 @@ async function executeRun(
     await rm(coverageDir, { recursive: true, force: true });
   }
   // Last line of the run, after every reporter and the coverage report.
-  process.stdout.write(`\n${formatFailureSummary({ failed, flaky, rerunFailed, hung, exitCode: status })}`);
+  const pids = outcomes.filter((outcome) => outcome.code !== 0 && outcome.pid).map((outcome) => outcome.pid);
+  process.stdout.write(`\n${formatFailureSummary({ failed, flaky, rerunFailed, hung, exitCode: status, pids })}`);
   return status;
 }

package/scripts/lib/scoped-temp.mjs

@@ -71,6 +71,7 @@ export function registerOwnedTempDir(dir) {
   installExitHook();
   return resolved;
 }
+
 /**
  * For code under test that makes its own mixdog-* directories in the system
  * temp dir: points TEMP/TMP/TMPDIR at one scoped directory for the rest of the

package/scripts/lib/test-failure-records.mjs

@@ -36,7 +36,20 @@ export function failureRecord(event, stacks) {
   if (event.data.todo) return null;
   if (event.data.details?.error?.failureType === 'subtestsFailed') return null;
   const ancestors = (stacks.get(file) ?? []).slice(0, event.data.nesting);
-  return { file, name: [...ancestors, event.data.name].join(' > ') };
+  const exit = processExit(event.data.details?.error);
+  return { file, name: [...ancestors, event.data.name].join(' > '), ...(exit ? { exit } : {}) };
+}
+
+// A test file whose process ended badly fails with only "test failed"; node
+// keeps how it ended on the error. A Windows NTSTATUS code (e.g. 0xC0000142,
+// the process could not initialise) reads better in hex.
+export function processExit(error) {
+  if (error?.failureType !== 'testCodeFailure') return '';
+  if (error.signal) return `test process killed by ${error.signal}`;
+  const code = error.exitCode;
+  if (!Number.isInteger(code) || code === 0) return '';
+  const hex = code > 255 || code < 0 ? ` (0x${(code >>> 0).toString(16).toUpperCase().padStart(8, '0')})` : '';
+  return `test process exited with code ${code}${hex}`;
 }
 
 // The runner sets FAILURE_RECORDS_ENV to a per-spawn JSONL path. Without it

package/scripts/lib/test-failure-summary.mjs

@@ -46,19 +46,24 @@ export function policyLine(rerunFailed) {
 // `hung` holds one "hung after N min: <files>" message per batch the runner's
 // watchdog killed; such a batch fails the run whatever its tests reported.
 // A non-zero `exitCode` with no failed test, flaky test or hung batch behind it
-// (a bad node flag, a crashed runner) is its own failure, never "0 failed".
-export function formatFailureSummary({ failed = [], flaky = [], rerunFailed = 0, hung = [], exitCode = 0 }) {
+// (a bad node flag, a crashed runner, a kill from outside) is its own failure,
+// never "0 failed". `pids` names the node --test processes that exited
+// non-zero, so a kill can be matched to the "Killing ... <pid>" line of
+// whatever killed them.
+export function formatFailureSummary({ failed = [], flaky = [], rerunFailed = 0, hung = [], exitCode = 0, pids = [] }) {
   const hungCount = hung.length > 0 ? `, ${hung.length} hung batch(es)` : '';
   const unexplained = exitCode !== 0 && failed.length === 0 && hung.length === 0;
+  const runner = pids.length > 0 ? `node --test pid ${pids.join(' ')}: ` : '';
   const lines = [
     unexplained
-      ? `${SUMMARY_TAG} RUN FAILED: exit code ${exitCode} with no failed test recorded (bad argument or runner error; see output above)`
+      ? `${SUMMARY_TAG} RUN FAILED: exit code ${exitCode} with no failed test recorded (${runner}bad argument, runner error or killed from outside; see output above)`
       : `${SUMMARY_TAG} ${failed.length} failed, ${flaky.length} flaky${hungCount} (every failed test is named below)`,
   ];
   for (const message of hung) lines.push(`${SUMMARY_TAG} HUNG ${message}`);
   for (const record of failed) {
     const origin = record.firstSeenOnRerun ? ` (first seen on re-run ${record.firstSeenOnRerun})` : '';
-    lines.push(`${SUMMARY_TAG} FAILED ${record.file} > ${record.name}${origin}`);
+    const exit = record.exit ? ` (${record.exit})` : '';
+    lines.push(`${SUMMARY_TAG} FAILED ${record.file} > ${record.name}${exit}${origin}`);
   }
   for (const record of flaky) {
     lines.push(

package/scripts/lib/test-spawn-ledger.mjs

@@ -1,12 +1,14 @@
 // Preloaded (--import) into every test process the runner starts: records
-// "<pid> <spawn time ms>" for each child process, so a process a test leaves
+// each child process's pid, spawn time and exit, so a process a test leaves
 // running can be traced back to the run after its parent exited
 // (scripts/lib/run-node-tests.mjs). Without the runner's environment it does
 // nothing.
 import { ChildProcess } from 'node:child_process';
 import { appendFileSync, readFileSync } from 'node:fs';
+import { constants } from 'node:os';
 import { dirname, join } from 'node:path';
 import { pathToFileURL } from 'node:url';
+import { getSystemErrorMap } from 'node:util';
 import { FAILURE_RECORDS_ENV, spawnLedgerPath } from './test-failure-records.mjs';
 
 // --test-force-exit ends a test file's process once every test registered so
@@ -54,11 +56,49 @@ if (
 
 const ledger = spawnLedgerPath(process.env[FAILURE_RECORDS_ENV]);
 if (ledger) {
-  // spawn, exec, execFile and fork all start their process here.
+  // spawn, exec, execFile and fork all start their process here. A spawn is
+  // "<pid> <before> <since>", the clock just before and just after it; the
+  // same line plus "<exited>" follows once its exit is seen. The 'exit' event
+  // comes before the child's handle is closed, so the pid is still this
+  // child's when the exit is written: from then on another process may hold it.
   const spawn = ChildProcess.prototype.spawn;
+  const exited = new Set();
   ChildProcess.prototype.spawn = function recordedSpawn(...args) {
+    const before = Date.now();
     const result = spawn.apply(this, args);
-    if (this.pid) appendFileSync(ledger, `${this.pid} ${Date.now()}\n`);
+    if (this.pid) {
+      const { pid } = this;
+      exited.delete(pid);
+      const spawned = `${pid} ${before} ${Date.now()}`;
+      appendFileSync(ledger, `${spawned}\n`);
+      this.once('exit', () => {
+        exited.add(pid);
+        appendFileSync(ledger, `${spawned} ${Date.now()}\n`);
+      });
+    }
     return result;
   };
+  // Once a child's exit is seen its pid is free, and Windows hands it to the
+  // next process started within moments — on a loaded gate often a parallel
+  // unit's test runner, node --test batch or test file process. A cleanup that
+  // signals every pid it recorded (e.g. `after(() => process.kill(pid))`, or a
+  // kill(pid, 0) liveness poll) then terminates that process instead, and the
+  // whole unit dies without output. Such a signal fails ESRCH here, as for a
+  // gone process, and never reaches whatever holds the pid now. Arguments Node
+  // rejects (a non-integer pid, an unknown signal) still go to Node, which
+  // throws before any signal is sent; this rejects only what it would send.
+  const kill = process.kill;
+  const [ESRCH] = [...getSystemErrorMap()].find(([, [name]]) => name === 'ESRCH');
+  const sendable = (signal) => signal === (signal | 0) || Boolean(constants.signals[signal || 'SIGTERM']);
+  process.kill = function killLiveChildOnly(pid, signal) {
+    // Only a primitive pid is checked: coercing an object would run its
+    // valueOf before Node's own coercion does. Number(pid) === (pid | 0) is
+    // Node's `pid == (pid | 0)` integer check.
+    const primitive = typeof pid === 'number' || typeof pid === 'string';
+    if (primitive && Number(pid) === (pid | 0) && exited.has(pid | 0) && sendable(signal)) {
+      // The shape of Node's own failed kill (an ErrnoException).
+      throw Object.assign(new Error('kill ESRCH'), { errno: ESRCH, code: 'ESRCH', syscall: 'kill' });
+    }
+    return kill.call(process, pid, signal);
+  };
 }

package/scripts/post-release-check.mjs

@@ -87,9 +87,12 @@ export async function checkFeed({ text, label, version, assets, hashAsset, requi
   for (const name of requiredFiles) {
     if (!byUrl.has(name)) problems.push(`${label}: files[] does not list ${name}`);
   }
+  // Every listed asset downloads at once; the problems keep files[] order.
+  const fileAssets = feed.files.map((file) => assets.find((candidate) => candidate.name === file.url));
+  const digests = await Promise.all(fileAssets.map((asset) => (asset ? hashAsset(asset) : null)));
   const hashes = new Map();
-  for (const file of feed.files) {
-    const asset = assets.find((candidate) => candidate.name === file.url);
+  for (const [index, file] of feed.files.entries()) {
+    const asset = fileAssets[index];
     if (!asset) {
       problems.push(`${label}: ${file.url} is not an uploaded asset`);
       continue;
@@ -99,7 +102,7 @@ export async function checkFeed({ text, label, version, assets, hashAsset, requi
       problems.push(`${label}: ${file.url} size ${size} != uploaded ${asset.size}`);
     }
     const sha512 = field(file, 'sha512');
-    const actual = await hashAsset(asset);
+    const actual = digests[index];
     hashes.set(file.url, actual);
     if (sha512 !== actual) problems.push(`${label}: ${file.url} sha512 does not match the uploaded asset`);
   }
@@ -233,6 +236,25 @@ export async function detectPublished({ fetchRelease, sleep = defaultSleep, atte
   }
 }
 
+/**
+ * Run independent `[name, run]` sections concurrently. Results keep the
+ * section order; a thrown section becomes a "check failed" problem.
+ */
+export function runSections(sections) {
+  return Promise.all(
+    sections.map(async ([name, run]) => {
+      try {
+        const outcome = await run();
+        return Array.isArray(outcome)
+          ? { name, problems: outcome, warnings: [] }
+          : { name, problems: outcome.problems, warnings: outcome.warnings };
+      } catch (error) {
+        return { name, problems: [`check failed: ${error.message}`], warnings: [] };
+      }
+    })
+  );
+}
+
 async function sha512Base64(url) {
   const response = await fetch(url, { redirect: 'follow' });
   if (!response.ok) throw new Error(`download ${url} answered ${response.status}`);
@@ -290,19 +312,7 @@ async function main() {
     ],
     ['npm', () => waitForNpm({ name: 'mixdog', version })],
   ];
-  const results = [];
-  for (const [name, run] of sections) {
-    try {
-      const outcome = await run();
-      results.push(
-        Array.isArray(outcome)
-          ? { name, problems: outcome, warnings: [] }
-          : { name, problems: outcome.problems, warnings: outcome.warnings }
-      );
-    } catch (error) {
-      results.push({ name, problems: [`check failed: ${error.message}`], warnings: [] });
-    }
-  }
+  const results = await runSections(sections);
 
   const lines = [`## Post-release check for ${tag}`, ''];
   for (const { name, problems, warnings } of results) {

package/scripts/pptx-kit-index.mjs

@@ -21,12 +21,11 @@ const SOURCES = [
 // A statement this short says its whole meaning (a constant, a one-line arrow); a longer one gives its first line.
 const WHOLE_STATEMENT = 240;
 
-const propertyName = (property) =>
-  property.type === 'Property'
-    ? String(property.key.name ?? property.key.value ?? '')
-    : property.type === 'SpreadElement'
-      ? '…'
-      : '';
+function propertyName(property) {
+  if (property.type === 'Property') return String(property.key.name ?? property.key.value ?? '');
+  if (property.type === 'SpreadElement') return '…';
+  return '';
+}
 
 // The keys an object literal offers a script, one level into object-valued keys ("body: { x, y, w, h }"). A table of
 // like entries (STYLES: an id per style, each with the same fields) reads as its ids and the fields each carries.