session-orchestrator 5.1.0 → 5.3.0

package/scripts/lib/process-group.mjs

@@ -0,0 +1,899 @@
+/**
+ * process-group.mjs — process-GROUP spawn/kill, identity re-verification and the
+ * gate-process ledger (Epic #1425: A4, plus the primitives A1/A2 and B3 build on).
+ *
+ * Why a group and not a child: `spawn(cmd, { shell: true })` makes the SHELL the
+ * child. `child.kill()` signals that shell only, and its own children —
+ * `tsgo --noEmit`, vitest workers — are reparented to PID 1 and keep running.
+ * That is the 2026-09-20 incident: four orphaned `tsgo` processes, two at
+ * PPID 1, up to 8.0 GB RSS each, host at 13% free memory. `detached: true`
+ * makes the shell a process-group LEADER (`child.pid === pgid`, setsid), so
+ * `process.kill(-pgid, sig)` reaches every descendant that did not setsid away.
+ *
+ * Measured on this host (Darwin 25.6.0, 2026-09-21) and load-bearing here:
+ *  - `process.kill(-pgid, 'SIGTERM')` does NOT terminate a grandchild that
+ *    installs `trap "" TERM`; SIGKILL to the group does. Hence the ladder.
+ *  - `close` fires only once EVERY group member has closed the shared pipe, so
+ *    a surviving grandchild hangs the promise. Hence the hard deadline.
+ *  - `maxBuffer` does not exist on async `spawn` (22 MB ran through unbounded).
+ *    Hence the hand-rolled byte cap, which reproduces the spawnSync/ENOBUFS
+ *    contract the existing gate tests pin ("21 MiB → exitCode 1").
+ *  - A detached child SURVIVES its parent's exit with PPID 1 unless someone
+ *    kills the group. Hence {@link installExitHandler}.
+ */
+
+import { spawn } from 'node:child_process';
+import { createHash } from 'node:crypto';
+import {
+  appendFileSync,
+  existsSync,
+  mkdirSync,
+  readFileSync,
+  writeFileSync,
+} from 'node:fs';
+import path from 'node:path';
+
+import { DEFAULT_KILL_GRACE_MS } from './wave-executor/dispatch-common.mjs';
+
+/**
+ * Re-exported, never re-defined: the SIGTERM→SIGKILL grace is a repo
+ * convention with its own named ceiling in `dispatch-common.mjs`. A second
+ * literal here would drift from it silently.
+ */
+export { DEFAULT_KILL_GRACE_MS };
+
+/** Append-only ledger of every gate process THIS repo's orchestrator started.
+ * Gitignored via `.gitignore:125` (`.orchestrator/runtime/`), verified with
+ * `git check-ignore --no-index -v`. */
+export const GATE_PROCESS_LEDGER_RELPATH = '.orchestrator/runtime/gate-processes.jsonl';
+
+/**
+ * How long to wait after SIGKILL before reading liveness back.
+ *
+ * Named ceiling (BV-004): 500 ms. On 2026-09-20 the hand-run cleanup routine
+ * measured immediately after `kill -9` and reported "still alive" for processes
+ * that were already gone — a signal is asynchronous, the reap is not instant.
+ * Revisit if a survivor is ever observed clearing later than this; raising it
+ * costs only the timeout path.
+ */
+export const DEFAULT_VERIFY_WAIT_MS = 500;
+
+/** Byte cap on captured stdout+stderr. Mirrors the 16 MiB `maxBuffer` the
+ * synchronous gate used, because async `spawn` has no `maxBuffer` at all. */
+export const DEFAULT_MAX_OUTPUT_BYTES = 16 * 1024 * 1024;
+
+/** Ledger entries older than this are not candidates for anything. */
+export const DEFAULT_LEDGER_MAX_AGE_MS = 24 * 3600 * 1000;
+
+/** Wall-clock ceiling for one gate command. Same 15 min as the synchronous
+ * path's `GATE_TIMEOUT_MS`, per the PRD parameter table (`gate.timeout-path-b-ms`),
+ * so both gate paths are allowed exactly as long. */
+export const DEFAULT_GATE_TIMEOUT_MS = 900_000;
+
+/** Lines of the tail returned as `output` (the full capture is `fullOutput`). */
+const OUTPUT_TAIL_LINES = 50;
+
+/** Signals whose default disposition terminates the process; on these the exit
+ * handler must kill our groups before we go away. */
+const EXIT_SIGNALS = ['SIGINT', 'SIGTERM', 'SIGHUP'];
+
+/**
+ * @typedef {object} GateProcessRecord
+ * @property {number} pid           PID of the group leader (the shell).
+ * @property {number} pgid          Process-group id — equals `pid` under `detached: true`.
+ * @property {number} startTime     `Date.now()` at spawn. The identity anchor against PID recycling.
+ * @property {string} commandSignature  Stable short form of the command — see {@link buildCommandSignature}.
+ * @property {string|null} sessionId    Owning session, or null when unknown.
+ * @property {string} recordedAt    ISO-8601 of `startTime`.
+ */
+
+/**
+ * Groups this process started and has not yet reaped: pgid → its `killFn`.
+ * Keyed by pgid so a double-register is idempotent. The stored `killFn` is the
+ * one the caller passed, which is why the exit handler never reaches
+ * `process.kill` for a group a test spawned through an injected seam.
+ *
+ * @type {Map<number, {killFn: (target: number, signal: string) => unknown}>}
+ */
+const LIVE_GROUPS = new Map();
+
+let exitHandlerInstalled = false;
+
+/**
+ * Default liveness probe: signal 0 tells us whether a PID exists without
+ * touching it. `ESRCH` = gone. `EPERM` = alive but not ours — reported as ALIVE
+ * on purpose, because "I may not signal it" is not "it is dead", and a survivor
+ * misreported as dead is exactly the false green this module exists to prevent.
+ *
+ * @param {number} pid
+ * @returns {boolean}
+ */
+function defaultIsAlive(pid) {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch (err) {
+    return err?.code === 'EPERM';
+  }
+}
+
+/** @param {number} ms @returns {Promise<void>} */
+function defaultSleep(ms) {
+  return new Promise((resolve) => {
+    setTimeout(resolve, ms);
+  });
+}
+
+/** @param {number} target @param {string} signal @returns {unknown} */
+function defaultKill(target, signal) {
+  return process.kill(target, signal);
+}
+
+/**
+ * Is `pgid` a value that may be NEGATED and handed to `kill(2)`?
+ *
+ * The only guard between a malformed ledger line and a POSIX broadcast. Three
+ * values are catastrophic rather than merely wrong, and all three are ordinary
+ * JSON numbers that reach here from a file on disk:
+ *  - `1` → `kill(-1, sig)` signals EVERY process the user may signal.
+ *  - `0` / `-0` → `kill(-0, sig)` signals the CALLER's own group (the session).
+ *  - a negative pgid → the negation turns it POSITIVE, so the group-wide kill
+ *    silently degrades into a single-PID kill of a stranger.
+ * `1` is excluded and not merely `<= 0` because pgid 1 is launchd's group on
+ * Darwin and init's on Linux — a real id, and never one of ours.
+ *
+ * @param {unknown} pgid
+ * @returns {boolean}
+ */
+function isSignalablePgid(pgid) {
+  return Number.isInteger(pgid) && /** @type {number} */ (pgid) > 1;
+}
+
+/**
+ * Is a parsed ledger line a usable {@link GateProcessRecord}?
+ *
+ * Shared by {@link readGateProcessLedger} and {@link pruneGateProcessLedger} so
+ * the reader and the pruner cannot disagree about what "usable" means — a line
+ * the reader rejects but the pruner keeps would sit in the ledger forever.
+ *
+ * `typeof === 'number'` is not enough for a value that gets NEGATED and handed
+ * to `kill(2)`: `1`, `0`, `-0`, `NaN` and a negative pgid are all numbers, and
+ * the first two are a POSIX broadcast and a self-kill respectively
+ * ({@link isSignalablePgid}). `commandSignature` is required for the same class
+ * of reason: without it the reaper's identity check has nothing to compare a
+ * `ps` row against but the row itself.
+ *
+ * @param {unknown} parsed
+ * @returns {boolean}
+ */
+function isValidLedgerRecord(parsed) {
+  if (!parsed || typeof parsed !== 'object') return false;
+  const r = /** @type {Record<string, unknown>} */ (parsed);
+  return isSignalablePgid(r.pid)
+    && isSignalablePgid(r.pgid)
+    && typeof r.commandSignature === 'string' && r.commandSignature.length > 0
+    && typeof r.startTime === 'number' && Number.isFinite(r.startTime);
+}
+
+/**
+ * Stable, short, non-reversible signature of a command line.
+ *
+ * Shape: `<first token>:<16 hex of sha256(full command)>`. The leading token
+ * stays readable so a `ps` row can be matched against it by name
+ * ({@link verifyProcessIdentity}); the hash discriminates two invocations of the
+ * same binary with different arguments without copying a command line — which
+ * may carry paths or tokens — into a ledger.
+ *
+ * @param {string} cmd  Full command line.
+ * @returns {string} e.g. `npm:6f1c4e0a9b2d7e35`. Empty input yields `:<hash of "">`.
+ */
+export function buildCommandSignature(cmd) {
+  const full = String(cmd ?? '');
+  const firstToken = full.trim().split(/\s+/)[0] ?? '';
+  const hash = createHash('sha256').update(full, 'utf8').digest('hex').slice(0, 16);
+  return `${firstToken}:${hash}`;
+}
+
+/**
+ * The command-name half of a signature (everything before the final `:`).
+ *
+ * @param {string} signature
+ * @returns {string}
+ */
+function signatureToken(signature) {
+  const s = String(signature ?? '');
+  const i = s.lastIndexOf(':');
+  return i === -1 ? s : s.slice(0, i);
+}
+
+/**
+ * Install the process-wide exit handler that kills every still-registered group
+ * with SIGKILL. Idempotent — repeated calls install nothing further.
+ *
+ * `detached: true` buys a process group at the price of survival: measured
+ * 2026-09-21, a detached child outlives its parent's exit with PPID 1. This
+ * handler is what pays that price back.
+ *
+ * On `exit` the cleanup is synchronous (no async work is possible there).
+ * On a terminating SIGNAL the cleanup runs, and the signal's default effect is
+ * restored ONLY when this handler is the sole listener for it — named ceiling
+ * (BV-004): with another listener present we clean up and leave the exit
+ * decision to it, rather than exiting out from under it. Revisit if a host ever
+ * needs us to terminate despite a foreign listener.
+ *
+ * @returns {void}
+ */
+export function installExitHandler() {
+  if (exitHandlerInstalled) return;
+  exitHandlerInstalled = true;
+
+  process.once('exit', () => {
+    killAllLiveGroupsSync();
+  });
+
+  for (const signal of EXIT_SIGNALS) {
+    const handler = () => {
+      killAllLiveGroupsSync();
+      if (process.listenerCount(signal) === 0) {
+        // We were the only listener: restore the default disposition so the
+        // process still dies from the signal it was sent.
+        process.kill(process.pid, signal);
+      }
+    };
+    process.once(signal, handler);
+  }
+}
+
+/**
+ * SIGKILL every registered group, synchronously, swallowing every error.
+ * Deliberately not exported: it is the exit path, not part of the interface.
+ *
+ * @returns {void}
+ */
+function killAllLiveGroupsSync() {
+  for (const [pgid, entry] of LIVE_GROUPS) {
+    if (!isSignalablePgid(pgid)) continue;
+    try {
+      entry.killFn(-pgid, 'SIGKILL');
+    } catch {
+      /* already gone, or not ours — nothing to do at exit time */
+    }
+  }
+  LIVE_GROUPS.clear();
+}
+
+/**
+ * Send SIGTERM → grace → SIGKILL to a whole process GROUP, then VERIFY.
+ *
+ * The verification is the point: `ok` is true only when nothing answers a
+ * liveness probe after `verifyWaitMs`. An exit code or a sent signal proves
+ * nothing (PRD B6), and a process that survives SIGKILL — e.g. one that
+ * `setsid`-ed out of the group — is reported as a survivor, never booked as
+ * success.
+ *
+ * Error semantics:
+ *  - `invalid-pgid` — the id is not a signalable process group
+ *    ({@link isSignalablePgid}). Checked FIRST, before `beforeSignal` and before
+ *    any signal: `ok: false`, `signalsSent: []`, nothing is sent. The ledger is
+ *    a file on disk, and `pgid: 1` in one line means `kill(-1, …)` — a POSIX
+ *    broadcast to every process this user may signal.
+ *  - `ESRCH` on the FIRST signal — the group was already gone. `ok: true`,
+ *    `error: 'ESRCH'`, no escalation.
+ *  - `ESRCH` on the escalation — proof of death, not a failure: `error` stays
+ *    null and the ladder completes normally.
+ *  - `EPERM` — a foreign process sits in the group and refused our signal. Stop
+ *    signal, never a retry: `ok: false`, `error: 'EPERM'`, `survivors: [pgid]`.
+ *
+ * ## `beforeSignal` — the ladder is ABORTABLE, and that is load-bearing
+ *
+ * The ladder sleeps `killGraceMs` (10 s by default) between the two signals, and
+ * a promise cannot be un-awaited. Without a gate consulted IMMEDIATELY BEFORE
+ * each signal, the escalation fires ~10 s after the caller has moved on — at a
+ * pgid the caller has already deregistered and the kernel may have recycled onto
+ * a stranger. Two callers need exactly that gate, for the same reason:
+ *  - {@link spawnInGroup} passes `() => !settled`, so a child that closed during
+ *    the grace window never draws a late group-wide SIGKILL.
+ *  - the orphan-reaper passes a fresh identity re-check (PRD B3), so a PID
+ *    recycled between SIGTERM and SIGKILL is never escalated against.
+ *
+ * It may be async; a throw counts as REFUSAL (fail-closed — an unmeasurable gate
+ * must never read as permission). An abort returns `aborted: '<signal>'`,
+ * `ok: false`, and `signalsSent` carries only the signals actually attempted.
+ *
+ * @param {number} pgid  Process-group id (positive; the negation happens here).
+ * @param {object} [opts]
+ * @param {(target: number, signal: string) => unknown} [opts.killFn]  Signal seam. Tests ALWAYS inject.
+ * @param {number} [opts.killGraceMs]
+ * @param {number} [opts.verifyWaitMs]
+ * @param {(pid: number) => boolean} [opts.isAliveFn]  Liveness probe on the group leader.
+ * @param {(ms: number) => Promise<void>} [opts.sleepFn]
+ * @param {((signal: string) => boolean|Promise<boolean>)|null} [opts.beforeSignal]  Consulted
+ *   immediately before EVERY signal. Anything but `true` aborts the rest of the ladder.
+ * @returns {Promise<{ok: boolean, signalsSent: string[], survivors: number[],
+ *   error: 'ESRCH'|'EPERM'|'invalid-pgid'|null, aborted: string|null}>}
+ *   `signalsSent` records ATTEMPTS in order, including one that threw.
+ */
+export async function killProcessGroup(pgid, {
+  killFn = defaultKill,
+  killGraceMs = DEFAULT_KILL_GRACE_MS,
+  verifyWaitMs = DEFAULT_VERIFY_WAIT_MS,
+  isAliveFn = defaultIsAlive,
+  sleepFn = defaultSleep,
+  beforeSignal = null,
+} = {}) {
+  /** @type {string[]} */
+  const signalsSent = [];
+
+  // Fail-closed BEFORE anything else: an unsignalable id never reaches `killFn`,
+  // not even through an injected seam, and `beforeSignal` is not consulted —
+  // there is nothing to permit.
+  if (!isSignalablePgid(pgid)) {
+    return {
+      ok: false, signalsSent, survivors: [], error: 'invalid-pgid', aborted: null,
+    };
+  }
+
+  /** Fail-closed gate: only an explicit `true` permits the signal. */
+  const permitted = async (signal) => {
+    if (typeof beforeSignal !== 'function') return true;
+    try {
+      return (await beforeSignal(signal)) === true;
+    } catch {
+      return false;
+    }
+  };
+  const aborted = (signal) => ({
+    ok: false, signalsSent, survivors: [], error: null, aborted: signal,
+  });
+
+  if (!(await permitted('SIGTERM'))) return aborted('SIGTERM');
+
+  signalsSent.push('SIGTERM');
+  try {
+    killFn(-pgid, 'SIGTERM');
+  } catch (err) {
+    const code = err?.code ?? null;
+    if (code === 'ESRCH') {
+      return { ok: true, signalsSent, survivors: [], error: 'ESRCH', aborted: null };
+    }
+    return {
+      ok: false, signalsSent, survivors: [pgid], error: code === 'EPERM' ? 'EPERM' : null, aborted: null,
+    };
+  }
+
+  await sleepFn(killGraceMs);
+
+  // Escalate unconditionally: SIGTERM is a REQUEST and a grandchild with a
+  // TERM trap ignores it (measured). SIGKILL on an already-dead group throws
+  // ESRCH, which is the cheapest possible proof that the ladder worked.
+  // "Unconditionally" means "regardless of what SIGTERM appeared to achieve" —
+  // never "regardless of whether this pgid is still the process we targeted",
+  // which is what `beforeSignal` re-decides here.
+  if (!(await permitted('SIGKILL'))) return aborted('SIGKILL');
+
+  signalsSent.push('SIGKILL');
+  let escalationEsrch = false;
+  try {
+    killFn(-pgid, 'SIGKILL');
+  } catch (err) {
+    const code = err?.code ?? null;
+    if (code === 'ESRCH') {
+      escalationEsrch = true;
+    } else {
+      return {
+        ok: false, signalsSent, survivors: [pgid], error: code === 'EPERM' ? 'EPERM' : null, aborted: null,
+      };
+    }
+  }
+
+  if (escalationEsrch) {
+    return { ok: true, signalsSent, survivors: [], error: null, aborted: null };
+  }
+
+  await sleepFn(verifyWaitMs);
+  const alive = isAliveFn(pgid) === true;
+  return {
+    ok: !alive,
+    signalsSent,
+    survivors: alive ? [pgid] : [],
+    error: null,
+    aborted: null,
+  };
+}
+
+/**
+ * Re-verify, immediately before signalling, that PID `pid` is still the process
+ * the ledger recorded — the guard against PID recycling (PRD B3, Feature Area 3).
+ *
+ * PURE: `snapshotLine` is an ALREADY-PARSED `ps` row, so this function does no
+ * I/O and is fully testable from text fixtures. `null` means the process is gone.
+ *
+ * Both checks must hold:
+ *  1. Start time. macOS has no `etimes` and its `lstart` is locale-dependent, so
+ *     the comparison runs over ELAPSED seconds: `nowMs - etimeSeconds*1000`
+ *     against the recorded `startTime`, within `toleranceMs`.
+ *  2. Command signature. The leading token of `expected.commandSignature` must
+ *     prefix the observed `args` (an exact full-signature match also passes).
+ *
+ * Fail-closed: a row with no measurable `etimeSeconds` yields
+ * `start-time-mismatch` with `observed.etimeSeconds === null`, so a caller can
+ * tell "measured, and it differs" from "could not be measured" — both refuse
+ * the signal, and only the second is an instrument gap.
+ *
+ * @param {number} pid
+ * @param {{startTime: number, commandSignature: string}} expected
+ * @param {object} [opts]
+ * @param {{pid?: number, ppid?: number, rssKb?: number, etimeSeconds?: number|null, cpuPct?: number, args?: string|null}|null} [opts.snapshotLine]
+ * @param {number} [opts.nowMs]
+ * @param {number} [opts.toleranceMs=2000]
+ * @returns {{match: boolean, reason: 'ok'|'gone'|'start-time-mismatch'|'signature-mismatch', observed: {etimeSeconds: number|null, args: string|null}}}
+ */
+export function verifyProcessIdentity(pid, expected, {
+  snapshotLine = null,
+  nowMs = Date.now(),
+  toleranceMs = 2000,
+} = {}) {
+  const gone = { match: false, reason: /** @type {const} */ ('gone'), observed: { etimeSeconds: null, args: null } };
+  if (!snapshotLine || typeof snapshotLine !== 'object') return gone;
+  // A row for a different PID describes a different process, not this one.
+  if (snapshotLine.pid !== undefined && snapshotLine.pid !== null
+    && Number(snapshotLine.pid) !== Number(pid)) return gone;
+
+  const etimeSeconds = typeof snapshotLine.etimeSeconds === 'number' && Number.isFinite(snapshotLine.etimeSeconds)
+    ? snapshotLine.etimeSeconds
+    : null;
+  const args = typeof snapshotLine.args === 'string' ? snapshotLine.args : null;
+  const observed = { etimeSeconds, args };
+
+  if (etimeSeconds === null) {
+    return { match: false, reason: 'start-time-mismatch', observed };
+  }
+  const observedStart = nowMs - etimeSeconds * 1000;
+  if (Math.abs(observedStart - Number(expected?.startTime)) > toleranceMs) {
+    return { match: false, reason: 'start-time-mismatch', observed };
+  }
+
+  const wanted = String(expected?.commandSignature ?? '');
+  const token = signatureToken(wanted);
+  const argsFirstToken = (args ?? '').trim().split(/\s+/)[0] ?? '';
+  const signatureOk = args !== null
+    && token.length > 0
+    && (argsFirstToken === token || buildCommandSignature(args) === wanted);
+  if (!signatureOk) {
+    return { match: false, reason: 'signature-mismatch', observed };
+  }
+
+  return { match: true, reason: 'ok', observed };
+}
+
+/**
+ * Absolute path of the gate-process ledger for a repo.
+ * The relpath lives in exactly one constant so no second site spells
+ * `.orchestrator` by hand.
+ *
+ * @param {string} repoRoot
+ * @returns {string}
+ */
+function ledgerPathFor(repoRoot) {
+  return path.join(repoRoot, ...GATE_PROCESS_LEDGER_RELPATH.split('/'));
+}
+
+/**
+ * Append one {@link GateProcessRecord} to the gate-process ledger.
+ *
+ * Append-only JSONL: concurrent gate runs from parallel sessions each add their
+ * own line and never rewrite a foreign one. Best-effort — a ledger write must
+ * never fail a gate — but a failure prints one WARN line to stderr rather than
+ * vanishing (a ledger that silently stops being written is indistinguishable
+ * from a host with no gate processes).
+ *
+ * @param {string} repoRoot
+ * @param {GateProcessRecord} record
+ * @param {object} [opts]
+ * @param {(filePath: string, line: string) => void} [opts.appendFn]  Sink seam. When
+ *   supplied, the parent directory is NOT created — the caller owns its sink, and
+ *   mkdir on a synthetic repoRoot would materialise a directory in a test.
+ * @returns {void}
+ */
+export function recordGateProcess(repoRoot, record, { appendFn } = {}) {
+  const target = ledgerPathFor(repoRoot);
+  try {
+    if (!appendFn) mkdirSync(path.dirname(target), { recursive: true });
+    const line = `${JSON.stringify(record)}\n`;
+    if (appendFn) appendFn(target, line);
+    else appendFileSync(target, line, 'utf8');
+  } catch (err) {
+    process.stderr.write(
+      `process-group: could not record gate process in ${GATE_PROCESS_LEDGER_RELPATH}: ${err?.message ?? String(err)}\n`,
+    );
+  }
+}
+
+/**
+ * Read the gate-process ledger, tolerantly.
+ *
+ * Malformed lines are skipped AND COUNTED: a silently skipping JSONL parser
+ * turns a partial read into a clean verdict, which is the exact failure mode a
+ * reaper must not have. "Malformed" includes a line whose `pid`/`pgid` is not a
+ * SIGNALABLE process-group id and one carrying no `commandSignature` — see the
+ * inline note at the check. Entries older than `maxAgeMs` are filtered out and
+ * counted separately — they are not candidates for anything, and their PIDs are
+ * the most likely to have been recycled.
+ *
+ * @param {string} repoRoot
+ * @param {object} [opts]
+ * @param {(filePath: string) => string} [opts.readFn]  Reader seam; must throw or
+ *   return '' for a missing file.
+ * @param {number} [opts.nowMs]
+ * @param {number} [opts.maxAgeMs]
+ * @returns {{records: GateProcessRecord[], malformedLines: number, expired: number}}
+ *   `records` are the fresh, parseable entries in source order.
+ */
+export function readGateProcessLedger(repoRoot, {
+  readFn,
+  nowMs = Date.now(),
+  maxAgeMs = DEFAULT_LEDGER_MAX_AGE_MS,
+} = {}) {
+  const target = ledgerPathFor(repoRoot);
+  let raw = '';
+  try {
+    if (readFn) raw = readFn(target) ?? '';
+    else if (existsSync(target)) raw = readFileSync(target, 'utf8');
+  } catch {
+    return { records: [], malformedLines: 0, expired: 0 };
+  }
+
+  /** @type {GateProcessRecord[]} */
+  const records = [];
+  let malformedLines = 0;
+  let expired = 0;
+
+  for (const line of String(raw).split('\n')) {
+    if (line.trim().length === 0) continue;
+    let parsed;
+    try {
+      parsed = JSON.parse(line);
+    } catch {
+      malformedLines += 1;
+      continue;
+    }
+    if (!isValidLedgerRecord(parsed)) {
+      malformedLines += 1;
+      continue;
+    }
+    if (nowMs - parsed.startTime > maxAgeMs) {
+      expired += 1;
+      continue;
+    }
+    records.push(parsed);
+  }
+
+  return { records, malformedLines, expired };
+}
+
+/**
+ * Drop expired and unusable lines from the ledger, rewriting it in place.
+ * "Unusable" is {@link isValidLedgerRecord}'s verdict — the same one
+ * {@link readGateProcessLedger} applies, so nothing the reader skips survives
+ * the pruner.
+ *
+ * Called once per gate command from {@link spawnInGroup}'s register path and
+ * once per scan from the orphan reaper: an append-only ledger with no pruner
+ * grows monotonically, and it had none until 2026-09-22.
+ *
+ * In-process fs only — never a shell `rm`/`mv` (PSA-003, and
+ * `.orchestrator/metrics/**` deletions are a blocked-command rule for a reason).
+ *
+ * Named ceiling (BV-004): read-filter-write is not atomic against a concurrent
+ * append, so a line appended between the read and the write is lost. Acceptable
+ * while the ledger is one line per gate command (tens per session) and its only
+ * consumer is a best-effort reaper. Revisit — with an exclusive lock or an
+ * append-only compaction sidecar — if the ledger ever gains a consumer that
+ * must not miss an entry.
+ *
+ * @param {string} repoRoot
+ * @param {object} [opts]
+ * @param {number} [opts.nowMs]
+ * @param {number} [opts.maxAgeMs]
+ * @param {{existsSync: Function, readFileSync: Function, writeFileSync: Function}} [opts.fs]  fs seam.
+ * @returns {number} Number of lines removed (expired + malformed). 0 when the
+ *   ledger is absent, empty, or unreadable.
+ */
+export function pruneGateProcessLedger(repoRoot, {
+  nowMs = Date.now(),
+  maxAgeMs = DEFAULT_LEDGER_MAX_AGE_MS,
+  fs: fsSeam,
+} = {}) {
+  const io = fsSeam ?? { existsSync, readFileSync, writeFileSync };
+  const target = ledgerPathFor(repoRoot);
+  let raw;
+  try {
+    if (!io.existsSync(target)) return 0;
+    raw = io.readFileSync(target, 'utf8');
+  } catch {
+    return 0;
+  }
+
+  const kept = [];
+  let removed = 0;
+  for (const line of String(raw).split('\n')) {
+    if (line.trim().length === 0) continue;
+    let parsed;
+    try {
+      parsed = JSON.parse(line);
+    } catch {
+      removed += 1;
+      continue;
+    }
+    if (!isValidLedgerRecord(parsed) || nowMs - parsed.startTime > maxAgeMs) {
+      removed += 1;
+      continue;
+    }
+    kept.push(line);
+  }
+
+  if (removed === 0) return 0;
+  try {
+    io.writeFileSync(target, kept.length > 0 ? `${kept.join('\n')}\n` : '', 'utf8');
+  } catch (err) {
+    process.stderr.write(
+      `process-group: could not prune ${GATE_PROCESS_LEDGER_RELPATH}: ${err?.message ?? String(err)}\n`,
+    );
+    return 0;
+  }
+  return removed;
+}
+
+/**
+ * Spawn a command as the leader of its OWN process group and guarantee that the
+ * whole group is gone when this promise settles.
+ *
+ * Timeout path: the group gets SIGTERM → `killGraceMs` → SIGKILL → verify
+ * (`timedOut: true`, `exitCode: 124`).
+ * Overflow path: the same ladder, triggered by the byte cap instead of the
+ * clock (`overflow: true`, `exitCode: 1` — the contract the synchronous gate's
+ * ENOBUFS behaviour already has).
+ *
+ * The promise resolves on `close` — which fires only when EVERY group member
+ * has closed the shared pipe — or, if a member survives SIGKILL, on a hard
+ * deadline of `killGraceMs + verifyWaitMs + 1000` measured from the start of the
+ * kill ladder. Without that deadline one `setsid`-escaped grandchild hangs the
+ * gate forever, which is the failure the whole module exists to prevent.
+ *
+ * The converse holds too: once this promise HAS settled, the ladder is
+ * cancelled (`beforeSignal` in {@link killProcessGroup}). A cooperative child
+ * that closes on SIGTERM would otherwise still draw a SIGKILL `killGraceMs`
+ * later, at a pgid `finish()` already deregistered — the exact late signal at a
+ * possibly-recycled group id this module exists to prevent.
+ *
+ * @param {string} cmd  Full shell command. Executable configuration, not data —
+ *   see `.claude/rules/security.md` § Session Config Command Trust.
+ * @param {object} [opts]
+ * @param {string} [opts.cwd]
+ * @param {NodeJS.ProcessEnv} [opts.env]
+ * @param {number|null} [opts.timeoutMs]  null disables the clock (the byte cap still applies).
+ * @param {number} [opts.killGraceMs]
+ * @param {number} [opts.verifyWaitMs]
+ * @param {number} [opts.maxOutputBytes]
+ * @param {Function} [opts.spawnFn]  `(cmd, options) => ChildProcess`. Tests inject a fake.
+ * @param {(target: number, signal: string) => unknown} [opts.killFn]  Tests ALWAYS inject.
+ * @param {() => number} [opts.now]
+ * @param {(pid: number) => boolean} [opts.isAliveFn]
+ * @param {(ms: number) => Promise<void>} [opts.sleepFn]
+ * @param {(record: GateProcessRecord) => void} [opts.onRegister]  Defaults to
+ *   {@link recordGateProcess} when — and only when — `repoRoot` is given.
+ * @param {string} [opts.repoRoot]
+ * @param {string} [opts.commandSignature]  Defaults to {@link buildCommandSignature}.
+ * @param {string|null} [opts.sessionId]
+ * @returns {Promise<{exitCode: number, output: string, fullOutput: string, timedOut: boolean,
+ *   overflow: boolean, killSignals: string[], survivors: number[], pid: number, pgid: number,
+ *   durationMs: number}>}
+ *   `exitCode` is 124 on timeout, 1 on overflow or spawn failure, else the child's own code.
+ *   `output` is the last ~50 lines of `fullOutput`. `survivors` are PIDs still alive after
+ *   SIGKILL + `verifyWaitMs` — a non-empty list is a REPORTED failure, never a silent one.
+ */
+export function spawnInGroup(cmd, {
+  cwd,
+  env,
+  timeoutMs = DEFAULT_GATE_TIMEOUT_MS,
+  killGraceMs = DEFAULT_KILL_GRACE_MS,
+  verifyWaitMs = DEFAULT_VERIFY_WAIT_MS,
+  maxOutputBytes = DEFAULT_MAX_OUTPUT_BYTES,
+  spawnFn = spawn,
+  killFn = defaultKill,
+  now = Date.now,
+  isAliveFn = defaultIsAlive,
+  sleepFn = defaultSleep,
+  onRegister,
+  repoRoot,
+  commandSignature,
+  sessionId = null,
+} = {}) {
+  installExitHandler();
+
+  return new Promise((resolve) => {
+    const startedAt = now();
+    /** @type {number|null} */
+    let pid = null;
+    /** @type {number|null} */
+    let pgid = null;
+    /** @type {string[]} */
+    const killSignals = [];
+    let fullOutput = '';
+    let capturedBytes = 0;
+    let timedOut = false;
+    let overflow = false;
+    let settled = false;
+    let ladderStarted = false;
+    /** @type {number[]} */
+    let survivors = [];
+    /** @type {ReturnType<typeof setTimeout>|null} */
+    let timeoutTimer = null;
+    /** @type {ReturnType<typeof setTimeout>|null} */
+    let deadlineTimer = null;
+
+    const finish = (childExitCode) => {
+      if (settled) return;
+      settled = true;
+      if (timeoutTimer) clearTimeout(timeoutTimer);
+      if (deadlineTimer) clearTimeout(deadlineTimer);
+      if (pgid !== null) LIVE_GROUPS.delete(pgid);
+      let exitCode;
+      if (timedOut) exitCode = 124;
+      else if (overflow) exitCode = 1;
+      else if (typeof childExitCode === 'number') exitCode = childExitCode;
+      else exitCode = 1;
+      resolve({
+        exitCode,
+        output: fullOutput.split('\n').slice(-OUTPUT_TAIL_LINES).join('\n').trim(),
+        fullOutput,
+        timedOut,
+        overflow,
+        killSignals,
+        survivors,
+        pid: pid ?? -1,
+        pgid: pgid ?? -1,
+        durationMs: now() - startedAt,
+      });
+    };
+
+    let child;
+    try {
+      // The command is trusted configuration; `shell: true` is intentional.
+      // `detached: true` is what makes this a GROUP leader — the whole point.
+      // nosemgrep: unsafe-shell-spawn
+      child = spawnFn(cmd, {
+        cwd,
+        env,
+        shell: true,
+        detached: true,
+        stdio: ['ignore', 'pipe', 'pipe'],
+      });
+    } catch (err) {
+      fullOutput = `process-group: failed to spawn "${cmd}": ${err?.message ?? String(err)}`;
+      finish(1);
+      return;
+    }
+
+    pid = typeof child?.pid === 'number' ? child.pid : null;
+    // Under `detached: true` the child calls setsid, so it IS its own group
+    // leader and pgid === pid (measured 2026-09-21, Darwin 25.6.0).
+    pgid = pid;
+
+    if (pgid !== null) {
+      LIVE_GROUPS.set(pgid, { killFn });
+      /** @type {GateProcessRecord} */
+      const record = {
+        pid,
+        pgid,
+        startTime: startedAt,
+        commandSignature: commandSignature ?? buildCommandSignature(cmd),
+        sessionId: sessionId ?? null,
+        recordedAt: new Date(startedAt).toISOString(),
+      };
+      const register = onRegister ?? (repoRoot ? (r) => recordGateProcess(repoRoot, r) : null);
+      if (register) {
+        try {
+          register(record);
+        } catch {
+          /* the ledger is an aid to the reaper, never a gate precondition */
+        }
+      }
+      // Prune AFTER the append, once per gate command — the ledger is
+      // append-only and had NO production pruner, so it grew monotonically
+      // (measured 2026-09-22: 369 lines/day on this host). Only on the real
+      // sink: with an injected `onRegister` the caller owns its storage and a
+      // prune here would write a file the test never asked for. Cheap by
+      // construction (the file holds one line per gate command) and wrapped,
+      // because a pruner that throws must not fail the gate it is housekeeping
+      // for.
+      if (!onRegister && repoRoot) {
+        try {
+          pruneGateProcessLedger(repoRoot);
+        } catch {
+          /* housekeeping, never a gate precondition */
+        }
+      }
+    }
+
+    // Signals are recorded AT SEND TIME, not merged from the ladder's return
+    // value: a cooperative child closes before `killProcessGroup` resolves, and
+    // a `killSignals` filled after `resolve()` reports an empty ladder for the
+    // very case the ladder worked.
+    const trackingKill = (target, signal) => {
+      killSignals.push(signal);
+      return killFn(target, signal);
+    };
+
+    const runLadder = () => {
+      if (ladderStarted || pgid === null) return;
+      ladderStarted = true;
+      // Hard deadline: `close` waits for EVERY group member to close the shared
+      // pipe, so one survivor would hang this promise forever.
+      deadlineTimer = setTimeout(() => {
+        if (survivors.length === 0 && pgid !== null) survivors = [pgid];
+        finish(null);
+      }, killGraceMs + verifyWaitMs + 1000);
+      deadlineTimer.unref?.();
+
+      killProcessGroup(pgid, {
+        killFn: trackingKill,
+        killGraceMs,
+        verifyWaitMs,
+        isAliveFn,
+        sleepFn,
+        // Cancel the ladder the moment this promise settles. `finish()` has by
+        // then deleted the pgid from LIVE_GROUPS, so a SIGKILL arriving
+        // `killGraceMs` later would be aimed at an id nobody here owns any more
+        // — and the kernel may have recycled it onto a foreign group.
+        beforeSignal: () => !settled,
+      })
+        .then((res) => {
+          survivors = res.survivors;
+          if (!res.ok) finish(null);
+        })
+        .catch(() => {
+          finish(null);
+        });
+    };
+
+    const onChunk = (chunk) => {
+      if (overflow) return;
+      const text = chunk.toString();
+      capturedBytes += Buffer.byteLength(text, 'utf8');
+      fullOutput += text;
+      if (capturedBytes > maxOutputBytes) {
+        overflow = true;
+        fullOutput += `\nprocess-group: output exceeded ${maxOutputBytes} bytes — killing process group.\n`;
+        runLadder();
+      }
+    };
+    child.stdout?.on('data', onChunk);
+    child.stderr?.on('data', onChunk);
+
+    if (typeof timeoutMs === 'number' && Number.isFinite(timeoutMs) && timeoutMs > 0) {
+      timeoutTimer = setTimeout(() => {
+        timedOut = true;
+        runLadder();
+      }, timeoutMs);
+      timeoutTimer.unref?.();
+    }
+
+    child.on('error', (err) => {
+      fullOutput += `\nprocess-group: child error: ${err?.message ?? String(err)}`;
+      finish(1);
+    });
+    child.on('close', (code) => finish(typeof code === 'number' ? code : null));
+  });
+}
+
+/**
+ * Test-only: the pgids this process currently has registered for exit-time
+ * cleanup. Exported so a test can prove a group is deregistered after `close`
+ * instead of asserting on module internals.
+ *
+ * @returns {number[]}
+ */
+export function _liveGroupPgids() {
+  return [...LIVE_GROUPS.keys()];
+}