gitnexus 1.6.8-rc.8 → 1.6.8

package/hooks/claude/hook-db-lock-probe.cjs

@@ -3,14 +3,57 @@
  * with a command line that looks like a GitNexus MCP/serve server?
  *
  * Backends (no user-installed Sysinternals):
- * - Linux: scan procfs under /proc (per-PID fd entries) via stat(2) (dev+inode); works without lsof;
- *   optional lsof fallback when proc scan finds nothing.
+ * - Linux: cmdline-first procfs scan under /proc, no lsof at all (#2180). Three
+ *   phases, cheapest first: (0) read /proc/<pid>/comm — a tiny task->comm read
+ *   that never touches the target's mm — and keep only PIDs whose comm is a
+ *   plausible node/gitnexus server; (1) read up to GITNEXUS_HOOK_PROC_CMDLINE_MAX
+ *   bytes of /proc/<pid>/cmdline via openSync+readSync (bounded, so a D-state
+ *   holder stuck on mmap_lock or a giant argv can't wedge the hook) and prefilter
+ *   with isGitNexusServerCommand; (2) only for the 0..N survivors, stat their
+ *   /proc/<pid>/fd/* and compare dev+inode against the target lbug. The lbug
+ *   handle is fd-visible (a @ladybugdb/core property), so this finds every real
+ *   owner without scanning every fd of every process.
  * - macOS / *BSD / etc.: trusted lsof + ps (absolute paths first).
  * - Windows: Restart Manager (rstrtmgr) via bundled PowerShell script +
  *   Win32_Process for command lines; trusted powershell.exe under %SystemRoot%.
  *
- * Fail-open on most errors; fail-closed only on lsof ETIMEDOUT (Unix) or
- * PowerShell ETIMEDOUT (Windows), matching the hook contract.
+ * Fail matrix:
+ * - Linux proc scan: owner found -> fail-closed (skip augment); budget exhausted
+ *   (GITNEXUS_HOOK_LINUX_PROC_BUDGET_MS) -> fail-CLOSED (#2180). This is a
+ *   deliberate change from the old "timeout -> fail-open then try lsof" path.
+ *   End-to-end the busy-host outcome is unchanged: the old code's lsof fallback
+ *   ETIMEDOUT'd on the very hosts where the scan ran out of budget and ALSO
+ *   failed closed there — the lsof leg only ever added 1-2s of dead work plus
+ *   the orphan-storm risk it caused (#2163). What changes is that an overloaded
+ *   host now self-throttles immediately (the throttle the incident needed)
+ *   instead of paying for a doomed lsof. Mid-load hosts that used to fall
+ *   through to a successful lsof now answer from the scan directly (faster) or,
+ *   if even the scan can't finish in budget, fail closed (self-throttle) — a
+ *   bounded, documented tradeoff, never an orphan.
+ * - macOS / other Unix: fail-open on most errors; fail-closed only on lsof
+ *   ETIMEDOUT, matching the hook contract.
+ * - Windows: fail-closed only on PowerShell ETIMEDOUT.
+ *
+ * Unix subprocess containment contract (#2163):
+ * - lsof/ps are wrapped in coreutils `timeout`/`gtimeout` when a working
+ *   wrapper is found (`timeout -k 1 <budget> lsof ...`). If this hook process
+ *   is itself SIGKILLed (e.g. by the runner's 10s hook timeout) the wrapper
+ *   survives, SIGTERMs its child at the budget (2s lsof / 1s ps) and SIGKILLs
+ *   it 1s later — orphan lifetime is bounded at ~3s instead of unbounded.
+ * - GITNEXUS_HOOK_TIMEOUT_PATH: the sentinel value `disabled` switches the
+ *   wrapper off deterministically; any other value is adopted only when it
+ *   exists AND passes a one-shot `-k` exit-propagation self-test — otherwise
+ *   resolution FALLS THROUGH to the built-in candidate list (first self-test
+ *   pass wins), so no malformed value of any shape can silently disable
+ *   orphan containment.
+ * - The gitnexus server is lazy-open + sticky-hold: an idle MCP server holds
+ *   ZERO lbug fds until the repo's first MCP query, then keeps the fd open.
+ *   A probe before that first query is therefore always false — a known,
+ *   pre-existing race, not a bug in this probe.
+ * - resolveUnixGuardTimeout is exported so the hook adapters can wrap the
+ *   `gitnexus augment` CLI child — the longest-lived hook subprocess (7s
+ *   local / 12s npx inner budgets) — in the same guard; see runGitNexusCli
+ *   in the adapters (#2163 follow-up).
  */
 
 const fs = require('fs');
@@ -25,6 +68,16 @@ function isGitNexusServerCommand(command) {
   return hasServerMode && hasGitNexus;
 }
 
+// GITNEXUS_DEBUG-gated stderr diagnostics. Reuses the exact gating predicate the
+// Windows ps1-load warning already uses (===' 1' / ==='true') so there is one
+// debug convention in this file, and writes via process.stderr.write (NOT a
+// spawn) so it never perturbs the windowsHide spawn-count invariant.
+function debugLog(msg) {
+  if (process.env.GITNEXUS_DEBUG === '1' || process.env.GITNEXUS_DEBUG === 'true') {
+    process.stderr.write(`[GitNexus hook] ${msg}\n`);
+  }
+}
+
 function resolveHookBinary(tool) {
   const envKey = tool === 'lsof' ? 'GITNEXUS_HOOK_LSOF_PATH' : 'GITNEXUS_HOOK_PS_PATH';
   const fromEnv = process.env[envKey];
@@ -46,6 +99,106 @@ function resolveHookBinary(tool) {
   return tool;
 }
 
+function hasMissingHookBinaryOverride(tool) {
+  const envKey = tool === 'lsof' ? 'GITNEXUS_HOOK_LSOF_PATH' : 'GITNEXUS_HOOK_PS_PATH';
+  const fromEnv = process.env[envKey];
+  if (!fromEnv || !String(fromEnv).trim()) return false;
+  try {
+    return !fs.existsSync(String(fromEnv).trim());
+  } catch {
+    return true;
+  }
+}
+
+// Sentinel:
+//   undefined = not resolved yet (resolve lazily, on first lsof/ps fallback)
+//   string    = self-tested coreutils timeout/gtimeout path (use as wrapper)
+//   null      = no usable wrapper (disabled, none found, or self-test failed)
+let unixGuardTimeoutCache;
+
+/**
+ * Resolve a coreutils `timeout`/`gtimeout` binary to wrap lsof/ps with
+ * (#2163). Unix-only by contract: the probe's win32 dispatch returns before
+ * reaching it, and the exported callers (the adapters' runGitNexusCli,
+ * #2163 follow-up) must check the platform first — the self-test below
+ * spawns /bin/sh. The memoized result is module-wide, so probe and adapter
+ * share one lazy self-test per hook process.
+ *
+ * GITNEXUS_HOOK_TIMEOUT_PATH semantics: the sentinel `disabled` turns the
+ * wrapper off; any other value is only a CANDIDATE — an existing file path
+ * is tried first, but it must pass the `-k` exit-propagation self-test to
+ * be adopted. On any failure (non-existent path, directory, non-executable
+ * file, wrapper without `-k` support, always-exit-0 stub, …) resolution
+ * falls through to the built-in candidates below, tried in order, first
+ * self-test pass wins. This is strictly stronger than the sibling
+ * GITNEXUS_HOOK_LSOF_PATH / GITNEXUS_HOOK_PS_PATH overrides (which only
+ * check existence): no bad env value of ANY shape can silently disable
+ * orphan containment.
+ *
+ * Lazy self-test: candidates are probed only when the lsof/ps fallback is
+ * first reached, and the result is memoized. A candidate is adopted only
+ * when `timeout -k 1 1 /bin/sh -c 'exit 42'` exits 42 — i.e. it must RUN
+ * the wrapped command AND PROPAGATE its exit status. This rejects two
+ * failure shapes: wrappers without the coreutils `-k` flag — busybox <1.34,
+ * toybox, broken symlinks — which would exit with a usage error without
+ * ever running lsof, silently converting the lsof-ETIMEDOUT fail-closed
+ * contract into fail-open (#1492 regression); and always-exit-0 stubs
+ * (/bin/true shapes), which would otherwise be adopted and "succeed" every
+ * wrapped spawn instantly without running it — a constant no-owner probe
+ * answer and, worse, a silently dead augment (status 0, empty stderr passes
+ * the adapters' success check with no context; #2163 follow-up review).
+ * Only when EVERY candidate fails does the probe fall back to the unwrapped
+ * status quo (memoized null). busybox ≥1.34 passes the test and is fully
+ * usable for everything THIS file spawns (lsof/ps are the guard's direct
+ * children) and for the adapters' direct-exec arm. The adapters' npx arm
+ * additionally relies on coreutils' process-GROUP signalling for its
+ * `-s KILL` grandchild reaping; busybox signals only its direct child, and
+ * this self-test deliberately does not probe that capability — see the
+ * adapter docblocks for the residual-gap statement.
+ */
+function passesGuardSelfTest(guard) {
+  try {
+    const selfTest = spawnSync(guard, ['-k', '1', '1', '/bin/sh', '-c', 'exit 42'], {
+      encoding: 'utf-8',
+      timeout: 3000,
+      stdio: ['ignore', 'ignore', 'ignore'],
+      windowsHide: true,
+    });
+    return !selfTest.error && selfTest.status === 42;
+  } catch {
+    return false;
+  }
+}
+
+function resolveUnixGuardTimeout() {
+  if (unixGuardTimeoutCache !== undefined) return unixGuardTimeoutCache;
+  unixGuardTimeoutCache = null;
+  const fromEnv = process.env.GITNEXUS_HOOK_TIMEOUT_PATH;
+  const trimmed = fromEnv ? String(fromEnv).trim() : '';
+  if (trimmed === 'disabled') return unixGuardTimeoutCache;
+  const candidates = [];
+  if (trimmed && fs.existsSync(trimmed)) candidates.push(trimmed);
+  for (const builtin of [
+    '/usr/bin/timeout',
+    '/bin/timeout',
+    '/opt/homebrew/bin/gtimeout',
+    '/usr/local/bin/gtimeout',
+  ]) {
+    try {
+      if (fs.existsSync(builtin)) candidates.push(builtin);
+    } catch {
+      /* ignore */
+    }
+  }
+  for (const candidate of candidates) {
+    if (passesGuardSelfTest(candidate)) {
+      unixGuardTimeoutCache = candidate;
+      break;
+    }
+  }
+  return unixGuardTimeoutCache;
+}
+
 function resolveWindowsPowerShellPath() {
   const fromEnv = process.env.GITNEXUS_HOOK_POWERSHELL_PATH;
   if (fromEnv && String(fromEnv).trim() && fs.existsSync(String(fromEnv).trim())) {
@@ -132,76 +285,374 @@ function hasGitNexusServerOwnerWindows(dbPathAbs, myPid) {
   return false;
 }
 
-function readLinuxCmdline(pidStr) {
+// The procfs root every Linux scan path reads from. Production is always /proc;
+// GITNEXUS_HOOK_PROC_ROOT only exists so unit tests can inject a fixture tree
+// (comm + cmdline + fd symlinks) and assert the three-phase logic without
+// scanning the real, ~hundreds-of-process /proc of the test host.
+//
+// Test-only gate (F4): the override is honored ONLY under a test runner —
+// vitest injects VITEST="true" and NODE_ENV="test" into every worker (verified;
+// a production hook is `node <file>.cjs` with neither set). Without the gate, a
+// production env that accidentally leaked GITNEXUS_HOOK_PROC_ROOT (pointing at an
+// empty/bad tree) would make readdirSync find no pids -> 'not-owned' -> Linux
+// owner detection silently OFF (fail-OPEN: augment races the real server for the
+// lbug, the #1492 class). Gating to the test signal makes that leak inert in
+// production (always /proc) while the fake-procfs unit tests, which run under
+// vitest, still inject freely. Unset env (or non-test context) => /proc, so the
+// production path is byte-for-byte the historical behavior.
+function isTestContext() {
+  return (
+    process.env.VITEST === 'true' || process.env.VITEST === '1' || process.env.NODE_ENV === 'test'
+  );
+}
+function getProcRoot() {
+  if (!isTestContext()) return '/proc';
+  const raw = process.env.GITNEXUS_HOOK_PROC_ROOT;
+  return raw && String(raw).trim() ? String(raw) : '/proc';
+}
+
+// Max bytes read from /proc/<pid>/cmdline in Phase 1. Bounded by default so a
+// D-state holder wedged on mmap_lock, or a process with a pathological multi-MB
+// argv, can't stall the hook. 16 KiB comfortably clears a realistic
+// `node <abs path to .../node_modules/gitnexus/dist/cli/index.js> mcp` line
+// (the `mcp`/`serve` mode token lives at the very tail, so the cap must be large
+// enough to reach it — see PROC_CMDLINE_FLOOR escalation below). Overridable for
+// tests; never goes below PROC_CMDLINE_FLOOR.
+const PROC_CMDLINE_FLOOR = 4096;
+function getCmdlineMaxBytes() {
+  const raw = process.env.GITNEXUS_HOOK_PROC_CMDLINE_MAX;
+  // Number() (not parseInt) so "8e3" reads as 8000, not 8 (parseInt stops at
+  // 'e'). The `raw && String(raw).trim()` guard keeps empty/whitespace on the
+  // default; trailing garbage ("8abc") now -> NaN -> default (stricter).
+  const n = raw && String(raw).trim() ? Number(String(raw).trim()) : NaN;
+  if (Number.isFinite(n) && n >= PROC_CMDLINE_FLOOR) return n;
+  return 16384;
+}
+
+// Phase 0 comm prefilter. /proc/<pid>/comm is the kernel task->comm string,
+// capped at 16 bytes INCLUDING the trailing NUL — i.e. at most 15 visible
+// chars, truncated by the kernel with no marker. So a process whose real name
+// is longer than 15 chars shows a 15-char prefix here. The match below is
+// therefore truncation-safe in BOTH directions (a whitelist name that is a
+// prefix of comm, or comm that is a prefix of a whitelist name, both count) to
+// guarantee we never drop a real owner at this cheap stage — Phase 2's dev+ino
+// fd check is the real authority; Phase 0/1 only exist to skip the overwhelming
+// majority (kernel threads, shells, editors) cheaply.
+//
+// The whitelist is calibrated against what a real `gitnexus mcp`/`serve` server
+// actually reports for comm. Observed on production hosts: the server renames
+// its main thread, so comm reads `MainThread` (via @ladybugdb/core's
+// worker_threads setup), NOT `node` — omitting it would blind the probe to
+// every real server (#1492-class owner miss). We also keep the plausible
+// launcher/runtime basenames in case a future build does not rename the thread.
+// Conservative by design: over-collecting a few extra candidates only costs a
+// bounded number of Phase 1 cmdline reads.
+const COMM_CANDIDATES = ['node', 'gitnexus', 'bun', 'deno', 'npm', 'npx', 'MainThread'];
+function commLooksLikeServer(comm) {
+  const c = comm.trim();
+  if (!c) return false;
+  for (const name of COMM_CANDIDATES) {
+    if (name === c || name.startsWith(c) || c.startsWith(name)) return true;
+  }
+  return false;
+}
+
+function readProcComm(procRoot, pidStr) {
   try {
-    return fs.readFileSync(`/proc/${pidStr}/cmdline`, 'utf8').replace(/\0+/g, ' ').trim();
+    return fs
+      .readFileSync(path.join(procRoot, pidStr, 'comm'), 'utf8')
+      .replace(/\0+/g, '')
+      .trim();
   } catch {
     return '';
   }
 }
 
-function linuxProcScanFindGitNexusServer(dbPathAbs, myPid) {
+// Timeout sentinel for readLinuxCmdline (F3). MUST be distinct from the
+// "unreadable/empty" return value (''): '' flows through isGitNexusServerCommand
+// as a NON-candidate (both regexes are false on ''), so the Phase 1 caller
+// `continue`s past it — correct for a raced/openSync-failed pid, but a FAIL-OPEN
+// bug if it ever meant "I ran out of budget mid-read" (a real owner whose
+// escalation timed out would be silently dropped, racing the lbug -> #1492). A
+// unique Symbol can never collide with any cmdline string, so the caller can
+// branch on it explicitly and map a mid-read timeout to the tri-state 'timeout'
+// (fail-CLOSED) instead of swallowing it as a non-candidate.
+const CMDLINE_TIMEOUT = Symbol('gitnexus.cmdline.timeout');
+
+// Bounded /proc/<pid>/cmdline read for Phase 1. openSync+readSync (not
+// readFileSync) so a D-state holder cannot stall the hook on a huge or
+// never-EOF argv: we read at most `cap` bytes and stop. cmdline separates argv
+// with NULs; convert to spaces for isGitNexusServerCommand.
+//
+// Owner-miss guard for the 4 KB cap: the `gitnexus` token usually sits in the
+// first path component while the `mcp`/`serve` mode token is the LAST argv, so
+// a naive 4 KB read could clip the mode token off a server launched with a very
+// long interpreter path and silently miss a real owner. We mitigate two ways:
+// (a) the default cap (16 KiB) already clears realistic lines; (b) if the first
+// read fills the cap AND already contains the `gitnexus` token but no mode
+// token yet, we keep reading in bounded chunks (up to a hard ceiling) until the
+// mode token appears or the file ends — so a genuine server is never missed for
+// want of a few more bytes, while non-candidates still pay only the initial
+// bounded read.
+//
+// Budget (F3): the escalation loop above is the one place a SINGLE pathological
+// candidate could read up to HARD_CEIL (256 KiB) before the next scan-level
+// budget check, weakening the timeout contract. `outOfBudget` (the scan's shared
+// deadline callback) is checked once per escalation iteration; on expiry we
+// return CMDLINE_TIMEOUT (NOT '') so the caller can fail-closed honestly rather
+// than mistake the partial read for a non-candidate. Reads that simply can't
+// open / error out still return '' (genuinely "not a readable candidate").
+function readLinuxCmdline(procRoot, pidStr, cap, outOfBudget) {
+  const file = path.join(procRoot, pidStr, 'cmdline');
+  let fd;
+  try {
+    fd = fs.openSync(file, 'r');
+  } catch {
+    return '';
+  }
+  try {
+    const HARD_CEIL = 262144; // 256 KiB absolute ceiling for the escalation path
+    let collected = Buffer.alloc(0);
+    let offset = 0;
+    let chunkCap = cap;
+    for (;;) {
+      // allocUnsafe is safe here: readSync fills exactly [0, bytes), only
+      // buf.subarray(0, bytes) is consumed, and Buffer.concat deep-copies that
+      // slice into `collected`, so the uninitialized tail never reaches decode.
+      const buf = Buffer.allocUnsafe(chunkCap);
+      const bytes = fs.readSync(fd, buf, 0, chunkCap, offset);
+      if (bytes <= 0) break;
+      collected = Buffer.concat([collected, buf.subarray(0, bytes)]);
+      offset += bytes;
+      const text = collected.toString('utf8').replace(/\0+/g, ' ');
+      // Stop early when we can already decide "owner": has both the gitnexus
+      // token and a mode token. Keep going only when gitnexus is present but
+      // the mode token might be just past the boundary.
+      const hasGitNexus =
+        /(?:^|[/\\\s])gitnexus(?:\.cmd)?(?:\s|$)/.test(text) ||
+        /node_modules[/\\]gitnexus[/\\]/.test(text);
+      const hasMode = /(?:^|\s)(mcp|serve)(?:\s|$)/.test(text);
+      if (hasMode) break; // decided (positive); isGitNexusServerCommand re-checks below
+      if (bytes < chunkCap) break; // EOF: full cmdline read, definitive
+      if (!hasGitNexus) break; // not a candidate; do not escalate the read
+      if (offset >= HARD_CEIL) break; // bounded escalation only
+      // Budget gate the escalation: a single huge-argv candidate must not burn
+      // the whole scan deadline before we re-check. Return the timeout sentinel
+      // (never '') so the caller fails closed instead of treating us as a
+      // non-candidate. The sole caller (linuxProcScanFindGitNexusServer) always
+      // passes outOfBudget, so no presence guard is needed.
+      if (outOfBudget()) return CMDLINE_TIMEOUT;
+      chunkCap = cap; // keep reading more in cap-sized chunks
+    }
+    return collected.toString('utf8').replace(/\0+/g, ' ').trim();
+  } catch {
+    return '';
+  } finally {
+    try {
+      fs.closeSync(fd);
+    } catch {
+      /* ignore */
+    }
+  }
+}
+
+function resolveLinuxProcBudgetMs() {
   const raw = process.env.GITNEXUS_HOOK_LINUX_PROC_BUDGET_MS;
-  const budget = Number(raw && String(raw).trim()) ? Number.parseInt(String(raw), 10) : 1200;
+  // Gate on the STRING's emptiness, NOT the parsed number's truthiness — the
+  // old `Number(raw && trim()) ? ... : 1200` form treated "0" as falsy and
+  // silently fell back to 1200 (#2180). Use Number() (not parseInt) so "16e3"
+  // reads as 16000, not 16 (parseInt stops at 'e'). The `&& String(raw).trim()`
+  // guard is load-bearing: without it a set-but-empty/whitespace value would be
+  // `Number("")===0` => budget 0 => immediate fail-CLOSED timeout (augment
+  // permanently skipped). With it, ''/whitespace => NaN => 1200 default, while a
+  // finite "0" still parses to an explicit, deterministic "no budget" =>
+  // immediate timeout. Non-numeric / unset => default 1200.
+  const n = raw != null && String(raw).trim() ? Number(String(raw).trim()) : NaN;
+  if (!Number.isFinite(n)) return 1200;
+  return n; // may be <= 0, meaning "out of budget on the first check"
+}
+
+// Returns one of: 'owned' (a non-self process with a GitNexus-server cmdline
+// holds the target lbug fd), 'not-owned' (scan completed, no such owner), or
+// 'timeout' (the per-scan budget was exhausted before a verdict). The name is
+// pinned by a source-contract test; only the return TYPE changed (#2180:
+// boolean -> tri-state, so the dispatcher can fail-closed on 'timeout').
+function linuxProcScanFindGitNexusServer(dbPathAbs, myPid) {
+  const budget = resolveLinuxProcBudgetMs();
+  // A non-positive budget is an explicit, deterministic "no time to scan" =>
+  // immediate timeout (the #2180 test vector, and the only correct reading of
+  // the fixed parse: "0" must NOT mean 1200). Returning before any procfs read
+  // keeps it instantaneous regardless of host load.
+  if (budget <= 0) return 'timeout';
+  const procRoot = getProcRoot();
+  const cmdlineCap = getCmdlineMaxBytes();
   const start = Date.now();
+  const outOfBudget = () => Date.now() - start > budget;
+
   let targetStat;
   try {
     targetStat = fs.statSync(dbPathAbs);
   } catch {
-    return false;
+    // Caller already existsSync'd the path; a stat failure here is a transient
+    // race, treat as no owner (historical semantics).
+    return 'not-owned';
   }
+
   let procEntries;
   try {
-    procEntries = fs.readdirSync('/proc', { withFileTypes: true });
+    procEntries = fs.readdirSync(procRoot, { withFileTypes: true });
   } catch {
-    return false;
+    return 'not-owned';
   }
+
+  // Phase 0 + Phase 1: collect the few PIDs whose comm AND cmdline look like a
+  // GitNexus server, without touching any fd yet.
+  const candidates = [];
   for (const ent of procEntries) {
-    if (Date.now() - start > budget) return false;
+    if (outOfBudget()) return 'timeout';
     if (!ent.isDirectory() || !/^\d+$/.test(ent.name)) continue;
     const pid = Number.parseInt(ent.name, 10);
     if (!Number.isFinite(pid) || pid === myPid) continue;
-    const fdDir = path.join('/proc', ent.name, 'fd');
+
+    // Phase 0: cheap comm prefilter.
+    const comm = readProcComm(procRoot, ent.name);
+    if (!comm) continue; // unreadable comm (kernel thread, raced exit) -> skip
+    if (!commLooksLikeServer(comm)) continue;
+
+    // Phase 1: bounded cmdline read + isGitNexusServerCommand prefilter.
+    if (outOfBudget()) return 'timeout';
+    const cmdline = readLinuxCmdline(procRoot, ent.name, cmdlineCap, outOfBudget);
+    // F3: a mid-read budget timeout returns the CMDLINE_TIMEOUT sentinel (a
+    // Symbol, never a string). Fail CLOSED on it rather than letting it fall
+    // through isGitNexusServerCommand as a non-candidate — a real owner whose
+    // escalation timed out must not be silently dropped (would fail-OPEN).
+    if (cmdline === CMDLINE_TIMEOUT) return 'timeout';
+    if (!isGitNexusServerCommand(cmdline)) continue;
+    candidates.push(ent.name);
+  }
+
+  // Phase 2: only now stat the fds of the (typically 0-2) survivors.
+  for (const pidStr of candidates) {
+    if (outOfBudget()) return 'timeout';
+    const fdDir = path.join(procRoot, pidStr, 'fd');
     let fds;
     try {
       fds = fs.readdirSync(fdDir);
-    } catch {
+    } catch (err) {
+      // F1: the old code returned 'owned' for EVERY non-ENOENT error. That was
+      // a correctness bug: /proc/<pid>/fd is owner-only (mode 0500), so a
+      // cross-user/root `gitnexus mcp` serving a DIFFERENT repo passes Phase 0+1
+      // (its cmdline matches) and then EACCES'es here — yet its dev+ino was
+      // NEVER compared against THIS lbug. Claiming 'owned' lets it permanently,
+      // silently suppress augment for a repo it does not actually lock. We now
+      // distinguish the failure shapes (all still fail-closed where we can't
+      // prove non-ownership, but 'timeout' is the HONEST verdict for
+      // "inconclusive", not the false-positive 'owned'):
+      const code = err && err.code;
+      if (code === 'ENOENT') {
+        // Process raced away between the candidate scan and now -> genuinely no
+        // longer an owner. Move on.
+        continue;
+      }
+      if (code === 'EACCES' || code === 'EPERM') {
+        // Permission-denied fd dir: cannot read fds, so ownership is
+        // UNVERIFIABLE. Fail closed honestly via 'timeout' (the dispatcher maps
+        // timeout -> true, same protective skip as before) WITHOUT lying that we
+        // confirmed ownership. Do NOT degrade to not-owned/fail-open: if this
+        // really is the owner, fail-open re-opens the #1492 lbug race; augment
+        // is optional context, so a conservative skip costs little.
+        debugLog(
+          `fd dir unreadable for candidate pid ${pidStr} (${code}); ownership ` +
+            `unverifiable, probe inconclusive -> fail-closed (timeout)`,
+        );
+        return 'timeout';
+      }
+      if (code === 'EIO' || code === 'ESTALE') {
+        // Genuine transient I/O against this candidate's fd dir — not evidence
+        // it does NOT hold the lbug. Treat as inconclusive and fail closed
+        // (timeout) rather than continue, so a real owner mid-I/O-blip is not
+        // dropped (would fail-open).
+        debugLog(
+          `fd dir transient I/O error for candidate pid ${pidStr} (${code}); ` +
+            `probe inconclusive -> fail-closed (timeout)`,
+        );
+        return 'timeout';
+      }
+      // Any other shape (ENOTDIR — fd path is not a directory at all, so this
+      // is not a plausible live-procfs owner — and the long tail) is treated as
+      // "this candidate is not an owner": move to the next candidate instead of
+      // the old blanket 'owned'. If no other candidate owns the lbug the scan
+      // ends not-owned (dispatcher fail-open) — acceptable because ENOTDIR means
+      // the fd entry is structurally not a real /proc/<pid>/fd.
+      debugLog(
+        `fd dir not a readable directory for candidate pid ${pidStr} ` +
+          `(${code || 'unknown'}); treating candidate as non-owner -> continue`,
+      );
       continue;
     }
-    let holds = false;
     for (const fd of fds) {
-      if (Date.now() - start > budget) return false;
+      if (outOfBudget()) return 'timeout';
       try {
         const st = fs.statSync(path.join(fdDir, fd));
         if (st.dev === targetStat.dev && st.ino === targetStat.ino) {
-          holds = true;
-          break;
+          return 'owned';
         }
       } catch {
-        /* ignore */
+        /* fd raced closed; ignore */
       }
     }
-    if (!holds) continue;
-    if (isGitNexusServerCommand(readLinuxCmdline(ent.name))) return true;
   }
-  return false;
+
+  return 'not-owned';
 }
 
 function unixLsofPsFindGitNexusServer(dbPathAbs, myPid) {
+  const guard = resolveUnixGuardTimeout();
+  // An explicit missing override models ENOENT and must fail open instead of
+  // falling through to a host binary with different process-table visibility.
+  if (hasMissingHookBinaryOverride('lsof')) return false;
   const lsofPath = resolveHookBinary('lsof');
-  const lsof = spawnSync(lsofPath, ['-nP', '-t', '--', dbPathAbs], {
+  // The spawnSync timeouts below (lsof 1000ms / ps 500ms) are deliberately
+  // SHORTER than the wrapper budgets (2s / 1s): on the supervised path Node's
+  // SIGTERM always fires first, so `error.code === 'ETIMEDOUT'` and the
+  // fail-closed contract are untouched. The wrapper only matters once this
+  // hook process has been SIGKILLed and can no longer deliver that SIGTERM.
+  const [lsofCmd, lsofArgs] = guard
+    ? [guard, ['-k', '1', '2', lsofPath, '-nP', '-t', '--', dbPathAbs]]
+    : [lsofPath, ['-nP', '-t', '--', dbPathAbs]];
+  const lsof = spawnSync(lsofCmd, lsofArgs, {
     encoding: 'utf-8',
     timeout: 1000,
     stdio: ['ignore', 'pipe', 'ignore'],
     windowsHide: true,
   });
   if (lsof.error) return lsof.error.code === 'ETIMEDOUT';
+  // Guard-mediated deaths map to "unresponsive holder" (fail-closed). Three
+  // result shapes, verified against coreutils 9.1:
+  //   - signal-death: when `-k` escalates to SIGKILL, coreutils timeout
+  //     SELF-RAISES the signal, so spawnSync reports {status: null, signal}
+  //     with no .error (spawnSync's own ETIMEDOUT was handled above). The
+  //     same shape appears when this hook is frozen >2s (SIGSTOP, laptop
+  //     suspend) and the guard expires while it sleeps. By construction, a
+  //     guard-wrapped probe that died by signal without spawnSync ETIMEDOUT
+  //     is a budget/kill outcome.
+  //   - 124: budget expired and the child exited after the plain SIGTERM.
+  //   - 137: NOT the coreutils -k path — only exit-code-propagating wrappers,
+  //     or a child SIGKILLed externally (e.g. the OOM killer).
+  if (guard && lsof.status === null && lsof.signal) return true;
+  if (guard && (lsof.status === 124 || lsof.status === 137)) return true;
 
   const pids = (lsof.stdout || '').split(/\s+/).filter(Boolean);
+  const psMissing = hasMissingHookBinaryOverride('ps');
   const psPath = resolveHookBinary('ps');
   for (const pid of pids) {
     if (Number(pid) === myPid) continue;
-    const ps = spawnSync(psPath, ['-p', pid, '-o', 'command='], {
+    // Missing ps means we cannot verify that this pid is a GitNexus server.
+    if (psMissing) continue;
+    const [psCmd, psArgs] = guard
+      ? [guard, ['-k', '1', '1', psPath, '-p', pid, '-o', 'command=']]
+      : [psPath, ['-p', pid, '-o', 'command=']];
+    const ps = spawnSync(psCmd, psArgs, {
       encoding: 'utf-8',
       timeout: 500,
       stdio: ['ignore', 'pipe', 'ignore'],
@@ -211,6 +662,11 @@ function unixLsofPsFindGitNexusServer(dbPathAbs, myPid) {
       if (ps.error.code === 'ETIMEDOUT') return true;
       continue;
     }
+    // Same guard-mediated-death mapping as the lsof call above (signal-death
+    // from the -k escalation or a frozen hook; 124 budget expiry; 137 only
+    // for exit-code-propagating wrappers / external SIGKILL).
+    if (guard && ps.status === null && ps.signal) return true;
+    if (guard && (ps.status === 124 || ps.status === 137)) return true;
     if (isGitNexusServerCommand(ps.stdout || '')) return true;
   }
   return false;
@@ -229,8 +685,13 @@ function hasGitNexusDbLockedByGitNexusServer(dbPath, myPid) {
   }
 
   if (process.platform === 'linux') {
-    if (linuxProcScanFindGitNexusServer(dbPathAbs, myPid)) return true;
-    return unixLsofPsFindGitNexusServer(dbPathAbs, myPid);
+    // #2180: cmdline-first procfs scan, no lsof. 'timeout' fails CLOSED
+    // (overloaded host self-throttles — the throttle the orphan-storm incident
+    // needed; the old lsof fallback ETIMEDOUT'd and failed closed on these same
+    // hosts anyway, only slower and with the orphan risk). 'not-owned' is the
+    // only false. See the fail matrix in the file header.
+    const verdict = linuxProcScanFindGitNexusServer(dbPathAbs, myPid);
+    return verdict !== 'not-owned';
   }
 
   return unixLsofPsFindGitNexusServer(dbPathAbs, myPid);
@@ -238,4 +699,30 @@ function hasGitNexusDbLockedByGitNexusServer(dbPath, myPid) {
 
 module.exports = {
   hasGitNexusDbLockedByGitNexusServer,
+  // Exported for white-box unit tests that must assert the tri-state verdict
+  // ('owned' | 'not-owned' | 'timeout') directly — the dispatcher collapses
+  // timeout and owned to the same boolean true, so the boolean API alone cannot
+  // distinguish the F1 EACCES->timeout fix from the old EACCES->owned bug. The
+  // Probe interface already declares this optional. Linux-only by contract; the
+  // name is pinned by a source-contract test.
+  linuxProcScanFindGitNexusServer,
+  // #2163 follow-up: the hook adapters wrap the augment CLI in the same
+  // guard. Returns a self-tested wrapper path — the built-in candidates are
+  // always absolute; a GITNEXUS_HOOK_TIMEOUT_PATH override is adopted as the
+  // exact string that passed the self-test. Same string is also the same
+  // RESOLUTION for absolute paths and for slashless names (PATH lookup is
+  // cwd-independent); a slash-containing RELATIVE override, however, is
+  // existsSync-checked and self-tested against this process's cwd while the
+  // adapters spawn the CLI with a `cwd` option (chdir-before-exec), so such
+  // a value can pass here yet ENOENT at the augment call site — set the
+  // override to an absolute path. Returns null when the wrapper is
+  // disabled/unavailable. Never call on win32 (see its JSDoc).
+  resolveUnixGuardTimeout,
+  // Exported for white-box unit tests of the numeric-env parsing (#2183 review):
+  // Number()-not-parseInt so "16e3" reads as 16000, plus the empty/whitespace
+  // guard that keeps a set-but-empty budget on the 1200 default instead of an
+  // immediate fail-closed timeout. Tested directly because the values are
+  // otherwise only observable indirectly through scan timing/escalation.
+  getCmdlineMaxBytes,
+  resolveLinuxProcBudgetMs,
 };