cgraphx 2.0.2 → 2.0.3

package/dist/core/code/engine/storage/index-lock.js

@@ -0,0 +1,664 @@
+"use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.acquireIndexLock = exports.sweepStagingArtifacts = exports.isLockUnwritableCode = exports.LOCK_UNWRITABLE_CODES = exports.IndexLockTimeoutError = void 0;
+/**
+ * Cross-process single-writer lock for a GitNexus index directory (#2658).
+ *
+ * `analyze` is the only writer of a `.cgraphx/code/` (or `branches/<slug>/`) slot,
+ * but nothing stopped two `analyze` runs — e.g. two editor/agent SessionStart
+ * hooks firing on the same repo at once — from wiping and rebuilding the same
+ * store concurrently. They raced on `lbug` and its sidecars, wasted N× CPU
+ * producing one index, and left orphaned WAL fragments (#2637). This module
+ * gives the write path an exclusive, index-directory-scoped lock so a second
+ * writer waits for the first instead of colliding; after acquiring, the caller
+ * re-runs its normal freshness check, so a run whose work the holder already
+ * did exits up-to-date rather than rebuilding (single-flight coalescing).
+ *
+ * Ownership lives with the process that runs the pipeline (the heap-respawn
+ * child when a respawn happens, the original otherwise) — NOT a supervising
+ * parent — so the entity the OS tracks for liveness is always the real writer.
+ * See run-analyze.ts for the acquire site.
+ *
+ * TWO BACKENDS behind the {@link acquireIndexLock} seam:
+ *
+ *  - **socket** (Windows named pipe / Linux abstract socket, via `net`) — the
+ *    preferred, KERNEL-OWNED lock. Holding it = holding a listening endpoint the
+ *    kernel binds to this process; `EADDRINUSE` therefore means a *live* holder,
+ *    and the kernel drops the binding the instant the holder exits for ANY reason
+ *    (clean exit, crash, OOM, SIGKILL). That makes it provably race-free: no
+ *    stale detection, no pid-reuse guess, no takeover, and — since the endpoint
+ *    lives outside the index dir — no filesystem write, so it works unchanged on
+ *    a read-only index mount. This is the same class of kernel object as the
+ *    Windows named mutex the issue's reporter used as an external workaround, but
+ *    built from Node's stdlib `net`, so it adds NO native dependency and cannot
+ *    break `npx cgraph` install anywhere.
+ *
+ *  - **file** (`O_EXCL` pidfile) — the portable fallback for macOS/BSD (no
+ *    abstract sockets; filesystem sockets don't release cleanly on death) and
+ *    for any environment where the socket backend can't bind. It uses pid-
+ *    liveness staleness, an atomic rename-steal reclaim, bounded malformed-file
+ *    handling, read-only tolerance, and a finite wait timeout (a reused pid can
+ *    masquerade as live where process start-time isn't verifiable, so waiting is
+ *    bounded rather than a hang). Its stale-takeover has an irreducible narrow
+ *    race — inherent to file-based advisory locks — which is precisely why the
+ *    socket backend is preferred; only a kernel primitive closes it.
+ *
+ * Scope: cross-process, same logical index dir. The file backend never steals a
+ * foreign-host lock (pid liveness is meaningless across hosts); the socket
+ * backend is single-host by nature. The motivating case (local hook-driven
+ * re-index) is single-host. See AcquireOptions.timeoutMs for the wait ceiling.
+ */
+const node_fs_1 = require("node:fs");
+const node_net_1 = __importDefault(require("node:net"));
+const node_path_1 = __importDefault(require("node:path"));
+const node_os_1 = __importDefault(require("node:os"));
+const node_crypto_1 = require("node:crypto");
+const LOCK_FILENAME = 'analyze.lock';
+const LOCK_RECORD_VERSION = 1;
+/** Base poll interval while waiting for a live holder; jittered per attempt. */
+const DEFAULT_POLL_MS = 250;
+/** How often to re-emit the "still waiting for pid N" diagnostic. */
+const DIAGNOSTIC_INTERVAL_MS = 15_000;
+/**
+ * Default wait ceiling (10 min). Generous enough to sit behind a normal
+ * analyze, finite so a pid-reuse ghost on a platform without start-time
+ * verification can't wedge acquisition forever (see AcquireOptions.timeoutMs).
+ * A repo whose analyze legitimately runs longer can raise
+ * CGRAPH_INDEX_LOCK_TIMEOUT_MS (or set it ≤ 0 for unbounded).
+ */
+const DEFAULT_TIMEOUT_MS = 600_000;
+/**
+ * How long a lock file must stay unreadable (empty/partial JSON) before we
+ * treat it as a crash orphan and reclaim it. Tolerates the microsecond
+ * create→write→close window of a *live* owner (see acquireIndexLock), so we
+ * never steal a lock that is a poll-interval away from being written. Scaled
+ * off the poll interval, floored at 1s.
+ */
+const malformedGraceMs = (pollMs) => Math.max(1000, pollMs * 2);
+class IndexLockTimeoutError extends Error {
+    holder;
+    /**
+     * Whether `holder` carries a real, identifiable owner. False on the socket
+     * backend (and the file backend's malformed/vanished-lock timeouts), where the
+     * holder is a placeholder (`pid -1`) — the OS socket lock exposes no owner
+     * metadata (#2658 review M3). Consumers must not present `holder.pid` as a real
+     * pid when this is false.
+     */
+    holderKnown;
+    constructor(holder, waitedMs, holderKnown = true) {
+        super(holderKnown
+            ? `Timed out after ${waitedMs}ms waiting for another cgraph analyze ` +
+                `(pid ${holder.pid} on ${holder.hostname}, invocation ${holder.invocationId}) ` +
+                `to release the index lock.`
+            : `Timed out after ${waitedMs}ms waiting for another cgraph analyze ` +
+                `(holder identity unknown) to release the index lock.`);
+        this.name = 'IndexLockTimeoutError';
+        this.holder = holder;
+        this.holderKnown = holderKnown;
+    }
+}
+exports.IndexLockTimeoutError = IndexLockTimeoutError;
+const HOSTNAME = node_os_1.default.hostname();
+/** Linux: field 22 of /proc/<pid>/stat (starttime). null elsewhere / on error. */
+const readProcStartTime = (pid) => {
+    if (process.platform !== 'linux')
+        return null;
+    try {
+        const stat = (0, node_fs_1.readFileSync)(`/proc/${pid}/stat`, 'utf8');
+        // comm (field 2) is parenthesized and may contain spaces/')' — split after
+        // the last ')' so the remaining fields align to their documented numbers.
+        const afterComm = stat
+            .slice(stat.lastIndexOf(') ') + 2)
+            .trim()
+            .split(' ');
+        // afterComm[0] is field 3 (state); starttime is field 22 → index 19.
+        return afterComm[19] ?? null;
+    }
+    catch {
+        return null;
+    }
+};
+/** true if the pid exists (signal 0). EPERM means it exists but isn't ours. */
+const pidAlive = (pid) => {
+    try {
+        process.kill(pid, 0);
+        return true;
+    }
+    catch (err) {
+        return err.code === 'EPERM';
+    }
+};
+const buildRecord = () => ({
+    v: LOCK_RECORD_VERSION,
+    pid: process.pid,
+    hostname: HOSTNAME,
+    startTime: readProcStartTime(process.pid),
+    token: (0, node_crypto_1.randomBytes)(16).toString('hex'),
+    invocationId: (0, node_crypto_1.randomUUID)(),
+    acquiredAt: new Date().toISOString(),
+});
+const readRecord = (lockPath) => {
+    try {
+        const raw = (0, node_fs_1.readFileSync)(lockPath, 'utf8');
+        const parsed = JSON.parse(raw);
+        // `typeof NaN === 'number'`, so a bare number check lets NaN/0/-1/Infinity/
+        // fractional pids reach process.kill (#2658 review L4): a garbled or crafted
+        // lock file with `{"pid":0}` reads as a live holder and wedges a real analyze
+        // for the full wait timeout. A real pid is a positive integer.
+        if (!Number.isInteger(parsed.pid) || parsed.pid <= 0)
+            return null;
+        if (typeof parsed.token !== 'string')
+            return null;
+        return parsed;
+    }
+    catch {
+        // Missing (won the race, file gone) or malformed/half-written → treat as
+        // "no readable holder"; the caller retries the O_EXCL create.
+        return null;
+    }
+};
+/**
+ * A same-host holder is stale iff its process is gone, or (Linux) its pid is
+ * alive but was reused — a different start time. A live holder is never stolen
+ * on age alone (a large repo legitimately analyzes for many minutes), and a
+ * foreign-host holder is never stale (its liveness is unknowable here). Where
+ * start-time verification is unavailable (non-Linux), a reused pid cannot be
+ * distinguished from a genuine live holder, so it is NOT stolen — the finite
+ * acquire timeout is what bounds that case instead (see AcquireOptions).
+ */
+const isStale = (holder) => {
+    if (holder.hostname !== HOSTNAME)
+        return false;
+    if (!pidAlive(holder.pid))
+        return true;
+    const now = readProcStartTime(holder.pid);
+    if (holder.startTime && now && holder.startTime !== now)
+        return true; // pid reused
+    return false;
+};
+/**
+ * Reclaim a lock file we judged reclaimable — a dead holder (`expected` = its
+ * record) or a malformed/unreadable crash-orphan (`expected` = null) — moving
+ * the exact inode aside in ONE `rename` syscall to a token-unique name so two
+ * waiters reclaiming the same orphan can't both win (the loser's rename ENOENTs).
+ *
+ * CRITICAL (#2658 review): the reclaim must not act on a STALE judgment. The
+ * staleness decision (`isStale` / malformed-grace) happened a few syscalls ago;
+ * a live writer may have O_EXCL-created its own lock at `lockPath` since. Blindly
+ * renaming that live lock aside would delete it and admit a SECOND writer — the
+ * exact double-writer this lock exists to prevent (reproduced: ~18%/round under
+ * 4-way reclaim contention on the file backend). So:
+ *   1. re-read `lockPath` immediately BEFORE the rename and confirm it still holds
+ *      exactly what we judged (same token, or still-unreadable) — shrinking the
+ *      window to the single gap between this read and the rename;
+ *   2. after the rename, confirm what we ACTUALLY moved matches the judgment; if a
+ *      live lock slipped into that residual gap, RESTORE it (rename back) so its
+ *      holder is never displaced, and lose the reclaim.
+ * A concurrent creator whose fresh lock the restore overwrites is caught by the
+ * acquire loop's post-write read-back verify (see acquireViaFile), so it backs
+ * off rather than proceeding as a second writer.
+ *
+ * Returns true if we won the reclaim (caller retries the create), false if we
+ * lost the race or the judgment went stale (caller re-loops and re-reads).
+ */
+const matchesJudgment = (record, expected) => expected === null ? record === null : record?.token === expected.token;
+const stealLock = (lockPath, me, expected) => {
+    // (1) Re-verify the judgment still holds right before we move anything.
+    if (!matchesJudgment(readRecord(lockPath), expected))
+        return false;
+    if (expected === null && !(0, node_fs_1.existsSync)(lockPath))
+        return false; // malformed → but now vanished
+    const aside = `${lockPath}.dead.${me.token}`;
+    try {
+        (0, node_fs_1.renameSync)(lockPath, aside);
+    }
+    catch (err) {
+        if (err.code === 'ENOENT')
+            return false; // another stealer won
+        throw err;
+    }
+    // (2) Confirm what we moved is what we judged; if a live lock slipped into the
+    // read→rename gap, put it back — a live holder must never be displaced.
+    if (!matchesJudgment(readRecord(aside), expected)) {
+        try {
+            (0, node_fs_1.renameSync)(aside, lockPath); // restore; an overwritten concurrent creator's read-back backs it off
+        }
+        catch {
+            /* slot re-taken between our move and restore — leave it; we lost the reclaim */
+        }
+        return false;
+    }
+    try {
+        (0, node_fs_1.unlinkSync)(aside); // uniquely ours by token → safe; best-effort
+    }
+    catch {
+        /* leftover .dead.<token> is inert (not analyze.lock, not swept) — harmless */
+    }
+    return true;
+};
+/**
+ * Placeholder holder for an {@link IndexLockTimeoutError} thrown while the lock
+ * file exists but no valid record can be read (malformed/partial), or it keeps
+ * vanishing — there is no real holder to name, but the error still needs one so
+ * the CLI's `err.holder.pid` stays defined. This path is a rare backstop:
+ * malformed files are reclaimed within {@link MALFORMED_GRACE_MS}.
+ */
+const unknownHolder = () => ({
+    v: LOCK_RECORD_VERSION,
+    pid: -1,
+    hostname: HOSTNAME,
+    startTime: null,
+    token: '',
+    invocationId: '<unreadable>',
+    acquiredAt: '',
+});
+/**
+ * Filesystem-create error codes we tolerate by proceeding lock-free: a
+ * read-only mount (EROFS) or a denied create (EACCES/EPERM). Such a filesystem
+ * rejects every index WRITE in the same directory too, so no concurrent writer
+ * can exist and the lock is moot — an already-indexed repo on a `:ro` mount
+ * must still reach its `alreadyUpToDate` fast path (#2658). A genuinely-needed
+ * write fails later exactly as it would have without the lock.
+ */
+exports.LOCK_UNWRITABLE_CODES = new Set(['EROFS', 'EACCES', 'EPERM']);
+const isLockUnwritableCode = (code) => code !== undefined && exports.LOCK_UNWRITABLE_CODES.has(code);
+exports.isLockUnwritableCode = isLockUnwritableCode;
+/** A lock handle that owns nothing — returned when the filesystem refuses to
+ *  create the lock file (see {@link LOCK_UNWRITABLE_CODES}). Release is a no-op. */
+const noopHandle = (record) => ({ record, release: () => { } });
+/**
+ * Delete orphaned build/staging artifacts left in the lock directory by a
+ * crashed prior writer. Safe precisely because we hold the exclusive lock: no
+ * other writer can be creating these here right now, so anything present is a
+ * crash orphan. Matches this slot's staging files ONLY — never `lbug` itself,
+ * never `lbug.wal`/`lbug.shadow` (the LIVE index's own sidecars), and never a
+ * `branches/<slug>/` sub-slot (which owns its own lock + sweep). Non-recursive.
+ */
+const sweepStagingArtifacts = (lockDir, log) => {
+    // Matches `lbug.new`, `lbug.new.wal`, `lbug.staging.<id>`, `lbug.staging.<id>.wal`, …
+    // Does NOT match `lbug`, `lbug.wal`, `lbug.shadow`.
+    const stagingRe = /^lbug\.(staging\..+|new(\..+)?)$/;
+    let removed = 0;
+    let entries;
+    try {
+        entries = (0, node_fs_1.readdirSync)(lockDir);
+    }
+    catch {
+        return;
+    }
+    for (const name of entries) {
+        if (!stagingRe.test(name))
+            continue;
+        try {
+            (0, node_fs_1.unlinkSync)(node_path_1.default.join(lockDir, name));
+            removed++;
+        }
+        catch {
+            /* best-effort */
+        }
+    }
+    if (removed > 0) {
+        log?.(`Cleared ${removed} orphaned index-staging file(s) from a prior interrupted analyze.`);
+    }
+};
+exports.sweepStagingArtifacts = sweepStagingArtifacts;
+const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
+/** Poll delay with jitter (avoids two waiters lock-stepping), clamped so it
+ *  never overshoots the remaining timeout budget. Callers guarantee
+ *  `waited < timeoutMs`, so the result is ≥ 1. */
+const jitteredDelay = (pollMs, timeoutMs, waited) => {
+    const jitter = Math.floor(Math.random() * pollMs);
+    const remaining = timeoutMs - waited;
+    return Math.max(1, Math.min(pollMs + jitter, remaining));
+};
+/**
+ * Resolve the wait ceiling. Explicit `opt` wins; else
+ * CGRAPH_INDEX_LOCK_TIMEOUT_MS; else {@link DEFAULT_TIMEOUT_MS}. A value ≤ 0
+ * (from either source) means unbounded.
+ */
+const resolveTimeoutMs = (opt) => {
+    const raw = typeof opt === 'number'
+        ? opt
+        : (() => {
+            const env = process.env.CGRAPH_INDEX_LOCK_TIMEOUT_MS;
+            if (env === undefined || env === '')
+                return DEFAULT_TIMEOUT_MS;
+            const n = Number(env);
+            return Number.isFinite(n) ? n : DEFAULT_TIMEOUT_MS;
+        })();
+    return raw <= 0 ? Number.POSITIVE_INFINITY : raw;
+};
+/**
+ * Acquire the exclusive write lock for `lockDir` (the resolved index slot
+ * directory, e.g. `<repo>/.cgraphx/code` or `<repo>/.cgraphx/code/branches/<slug>`).
+ *
+ * Blocks until the lock is held (waiting only on live holders, stealing dead
+ * ones immediately), then sweeps orphaned staging files under the lock and
+ * returns a handle. Rejects with `IndexLockTimeoutError` if `timeoutMs` is
+ * exceeded while a live holder still holds the lock.
+ */
+/**
+ * File-based (O_EXCL pidfile) backend. The portable fallback used on platforms
+ * without the socket backend (macOS/BSD) or when the OS socket lock is
+ * unavailable. Carries the pid-liveness staleness, atomic rename-steal reclaim,
+ * bounded malformed-file handling, and read-only tolerance. Its stale-takeover
+ * has an irreducible (narrow) race — see the module header — which is why the
+ * socket backend is preferred where available.
+ */
+const acquireViaFile = async (lockDir, me, opts) => {
+    try {
+        (0, node_fs_1.mkdirSync)(lockDir, { recursive: true });
+    }
+    catch (err) {
+        // Read-only / denied filesystem → proceed lock-free (see LOCK_UNWRITABLE_CODES).
+        if ((0, exports.isLockUnwritableCode)(err.code))
+            return noopHandle(me);
+        throw err;
+    }
+    const lockPath = node_path_1.default.join(lockDir, LOCK_FILENAME);
+    const pollMs = opts.pollMs ?? DEFAULT_POLL_MS;
+    const timeoutMs = resolveTimeoutMs(opts.timeoutMs);
+    const startedAt = Date.now();
+    let announcedWait = false;
+    let lastDiagnosticAt = 0;
+    // When the lock file exists but has no readable record, the timestamp we
+    // first observed it unreadable — used to reclaim a crash-orphan after a grace.
+    let malformedSince = null;
+    for (;;) {
+        try {
+            // O_WRONLY | O_CREAT | O_EXCL — the atomic arbiter of ownership.
+            const fd = (0, node_fs_1.openSync)(lockPath, 'wx');
+            try {
+                (0, node_fs_1.writeSync)(fd, JSON.stringify(me));
+            }
+            finally {
+                (0, node_fs_1.closeSync)(fd);
+            }
+            // Read-back verify (#2658 review L5): if this process stalled (a >graceMs
+            // GC pause) between the O_EXCL create of the *empty* file and the write
+            // above, a waiter could have reclaimed the empty file (renamed it aside)
+            // and O_EXCL-created its own lock at `lockPath`. Our write then landed on
+            // the renamed-aside inode, not `lockPath`. Confirm `lockPath` still carries
+            // our token before claiming ownership; if it was stolen, contend normally.
+            const confirmed = readRecord(lockPath);
+            if (!confirmed || confirmed.token !== me.token)
+                continue;
+            return {
+                record: me,
+                release: () => {
+                    const current = readRecord(lockPath);
+                    if (current && current.token !== me.token)
+                        return; // no longer ours
+                    try {
+                        (0, node_fs_1.unlinkSync)(lockPath);
+                    }
+                    catch {
+                        /* already gone */
+                    }
+                },
+            };
+        }
+        catch (err) {
+            const code = err.code;
+            if (code === 'EEXIST') {
+                // fall through to holder inspection / wait / reclaim below
+            }
+            else if ((0, exports.isLockUnwritableCode)(code)) {
+                return noopHandle(me); // read-only / denied → proceed lock-free
+            }
+            else {
+                throw err;
+            }
+        }
+        const holder = readRecord(lockPath);
+        const waited = Date.now() - startedAt;
+        if (holder) {
+            malformedSince = null;
+            if (isStale(holder)) {
+                opts.log?.(`Reclaiming stale index lock from dead analyze (pid ${holder.pid}, ` +
+                    `invocation ${holder.invocationId}).`);
+                stealLock(lockPath, me, holder); // reclaim ONLY this dead record; live locks are never stolen
+                continue;
+            }
+            // Live holder → wait.
+            if (!announcedWait) {
+                announcedWait = true;
+                opts.onWaitStart?.(holder);
+                opts.log?.(`Another cgraph analyze (pid ${holder.pid} on ${holder.hostname}) is ` +
+                    `refreshing this index — waiting for it to finish.`);
+            }
+            if (waited >= timeoutMs)
+                throw new IndexLockTimeoutError(holder, waited);
+            if (Date.now() - lastDiagnosticAt >= DIAGNOSTIC_INTERVAL_MS) {
+                lastDiagnosticAt = Date.now();
+                if (waited >= DIAGNOSTIC_INTERVAL_MS) {
+                    opts.log?.(`Still waiting for analyze pid ${holder.pid} (${Math.round(waited / 1000)}s elapsed).`);
+                }
+            }
+            await sleep(jitteredDelay(pollMs, timeoutMs, waited));
+            continue;
+        }
+        // holder === null: the lock file is either gone (vanished between the failed
+        // create and our read) or present-but-unreadable (a crash between the
+        // O_EXCL create and the record write, or a partial write). NEVER hot-loop
+        // here — both branches are bounded by sleep + timeout.
+        if (!(0, node_fs_1.existsSync)(lockPath)) {
+            malformedSince = null; // genuinely vanished → the next create likely wins
+            if (waited >= timeoutMs)
+                throw new IndexLockTimeoutError(unknownHolder(), waited, false);
+            await sleep(jitteredDelay(pollMs, timeoutMs, waited));
+            continue;
+        }
+        // Malformed orphan present. Reclaim only after a grace, so a live owner's
+        // microsecond create→write window is never mistaken for a crash.
+        if (malformedSince === null)
+            malformedSince = Date.now();
+        if (Date.now() - malformedSince >= malformedGraceMs(pollMs)) {
+            opts.log?.('Reclaiming a malformed/partial index lock file (no readable owner record).');
+            stealLock(lockPath, me, null); // reclaim ONLY while still unreadable; a live lock written since is left
+            malformedSince = null;
+            continue;
+        }
+        if (waited >= timeoutMs)
+            throw new IndexLockTimeoutError(unknownHolder(), waited, false);
+        await sleep(jitteredDelay(pollMs, timeoutMs, waited));
+    }
+};
+/** Signals that the OS socket backend can't be used here (e.g. abstract
+ *  namespace disabled, sandbox, or an unexpected bind error) so the caller
+ *  should fall back to the file backend. NOT thrown for EADDRINUSE (that is a
+ *  live holder → wait) or timeouts (those propagate as IndexLockTimeoutError). */
+class SocketLockUnavailable extends Error {
+    cause;
+    constructor(cause) {
+        super(`OS socket lock unavailable: ${cause.code ?? cause.message}`);
+        this.cause = cause;
+        this.name = 'SocketLockUnavailable';
+    }
+}
+/**
+ * Canonicalize a path to its real filesystem identity so lexical aliases of the
+ * same directory (a symlink, a bind-mount path, a Windows junction, a `\\?\`
+ * prefix) map to ONE name (#2658 review H1). `lockDir` (the index slot) often
+ * does not exist yet, so `realpathSync` the deepest existing ancestor and
+ * re-append the not-yet-created remainder. A path with no symlink components
+ * realpaths to itself, so the common (non-aliased) case is unchanged — a holder
+ * that used the old resolved name is never orphaned.
+ */
+const canonicalizeDir = (p) => {
+    const resolved = node_path_1.default.resolve(p);
+    const tail = [];
+    let dir = resolved;
+    for (;;) {
+        try {
+            const real = (0, node_fs_1.realpathSync)(dir);
+            return tail.length ? node_path_1.default.join(real, ...tail.reverse()) : real;
+        }
+        catch {
+            const parent = node_path_1.default.dirname(dir);
+            if (parent === dir)
+                return resolved; // reached the root with nothing to resolve
+            tail.push(node_path_1.default.basename(dir));
+            dir = parent;
+        }
+    }
+};
+/**
+ * Stable OS-IPC endpoint name for an index directory. The name is derived from
+ * the REAL path (case-folded on Windows), so two processes targeting the same
+ * physical slot — even via different lexical aliases — collide, and separate
+ * worktrees/branches never do. The endpoint lives OUTSIDE the index directory
+ * (abstract namespace / pipe namespace), so the lock needs no filesystem write
+ * and is unaffected by a read-only index mount.
+ */
+const socketLockName = (lockDir) => {
+    const resolved = canonicalizeDir(lockDir);
+    const key = (0, node_crypto_1.createHash)('sha256')
+        .update(process.platform === 'win32' ? resolved.toLowerCase() : resolved)
+        .digest('hex')
+        .slice(0, 32);
+    return process.platform === 'win32'
+        ? `\\\\.\\pipe\\gitnexus-idx-${key}`
+        : `\0gitnexus-idx-${key}`; // Linux abstract socket (no filesystem entry)
+};
+/** Attempt to listen; resolve to null on success or the error on failure. */
+const tryListen = (server, name) => new Promise((resolve) => {
+    const onError = (err) => {
+        server.removeListener('listening', onListening);
+        resolve(err);
+    };
+    const onListening = () => {
+        server.removeListener('error', onError);
+        resolve(null);
+    };
+    server.once('error', onError);
+    server.once('listening', onListening);
+    server.listen(name);
+});
+/**
+ * OS-owned socket/pipe backend (Windows named pipe, Linux abstract socket).
+ * Holding the lock = holding a listening endpoint the kernel binds to this
+ * process; `EADDRINUSE` therefore means a *live* holder, and the kernel drops
+ * the binding the instant the holder exits (clean exit, crash, OOM, SIGKILL) —
+ * so there is no stale detection, no reclaim, and no takeover race. See the
+ * module header for why this is preferred over the file backend.
+ */
+const acquireViaSocket = async (lockDir, me, opts) => {
+    const name = socketLockName(lockDir);
+    const pollMs = opts.pollMs ?? DEFAULT_POLL_MS;
+    const timeoutMs = resolveTimeoutMs(opts.timeoutMs);
+    const startedAt = Date.now();
+    let announcedWait = false;
+    let lastDiagnosticAt = 0;
+    for (;;) {
+        const server = node_net_1.default.createServer();
+        // Never keep the process alive on the lock's account, and never hold an
+        // incoming connection (nothing should connect; drop any stray peer).
+        server.unref();
+        server.on('connection', (sock) => sock.destroy());
+        const listenErr = await tryListen(server, name);
+        if (!listenErr) {
+            return {
+                record: me,
+                release: () => {
+                    try {
+                        server.close();
+                    }
+                    catch {
+                        /* already closed / releasing on exit */
+                    }
+                },
+            };
+        }
+        // This server never bound (listen failed); release its handle before the
+        // next poll or the fallback, so a long contended wait doesn't churn one
+        // unclosed net.Server per iteration (#2658 review L3).
+        try {
+            server.close();
+        }
+        catch {
+            /* never listened */
+        }
+        // Only EADDRINUSE means "held by a live holder → wait". Anything else means
+        // this environment can't use the socket backend → fall back to the file one.
+        if (listenErr.code !== 'EADDRINUSE')
+            throw new SocketLockUnavailable(listenErr);
+        if (!announcedWait) {
+            announcedWait = true;
+            opts.onWaitStart?.(me);
+            opts.log?.('Another cgraph analyze is refreshing this index — waiting for it to finish.');
+        }
+        const waited = Date.now() - startedAt;
+        // Socket backend exposes no owner metadata → holder identity is unknown (M3).
+        if (waited >= timeoutMs)
+            throw new IndexLockTimeoutError(unknownHolder(), waited, false);
+        if (Date.now() - lastDiagnosticAt >= DIAGNOSTIC_INTERVAL_MS) {
+            lastDiagnosticAt = Date.now();
+            if (waited >= DIAGNOSTIC_INTERVAL_MS) {
+                opts.log?.(`Still waiting for another analyze (${Math.round(waited / 1000)}s elapsed).`);
+            }
+        }
+        await sleep(jitteredDelay(pollMs, timeoutMs, waited));
+    }
+};
+/** Platforms whose OS IPC namespace gives a clean, auto-releasing lock via
+ *  `net`: Windows named pipes and Linux abstract sockets. Elsewhere (macOS/BSD)
+ *  the file backend is used (no abstract namespace; filesystem sockets don't
+ *  release cleanly on death). Override for tests via CGRAPH_INDEX_LOCK_BACKEND
+ *  = 'socket' | 'file'.
+ *
+ *  Scope caveat: the socket backend's mutual-exclusion domain is NOT uniform.
+ *  Windows `\\.\pipe\` names are machine-wide (all sessions); Linux abstract
+ *  sockets are network-namespace-scoped (network_namespaces(7)). So two writers
+ *  that share a bind-mounted index dir but sit in separate netns (e.g. two
+ *  containers, Docker's default) do NOT collide on Linux — "single-host" is
+ *  really "single-netns" here. That cross-netns-shared-mount case is the one
+ *  the file backend (shared-filesystem O_EXCL) would cover; set
+ *  CGRAPH_INDEX_LOCK_BACKEND=file there. The motivating case (local hook-
+ *  driven re-index) is single-netns, so the default socket backend covers it. */
+const selectBackend = () => {
+    const override = process.env.CGRAPH_INDEX_LOCK_BACKEND;
+    if (override === 'socket' || override === 'file')
+        return override;
+    return process.platform === 'win32' || process.platform === 'linux' ? 'socket' : 'file';
+};
+/**
+ * Acquire the exclusive write lock for `lockDir` (the resolved index slot
+ * directory). Uses the OS socket/pipe backend where available (Windows/Linux),
+ * falling back to the file backend otherwise or if the socket backend is
+ * unusable in this environment. After acquiring, sweeps orphaned staging files
+ * under the lock (best-effort; a no-op on a read-only mount). Rejects with
+ * `IndexLockTimeoutError` if `timeoutMs` elapses while another live holder holds
+ * the lock.
+ */
+const acquireIndexLock = async (lockDir, opts = {}) => {
+    const me = buildRecord();
+    let handle;
+    if (selectBackend() === 'socket') {
+        try {
+            handle = await acquireViaSocket(lockDir, me, opts);
+        }
+        catch (err) {
+            if (!(err instanceof SocketLockUnavailable))
+                throw err; // timeout etc. propagate
+            opts.log?.('Index lock: OS socket lock unavailable here — using the file lock.');
+            handle = await acquireViaFile(lockDir, me, opts);
+        }
+    }
+    else {
+        handle = await acquireViaFile(lockDir, me, opts);
+    }
+    // Reclaim crashed-build staging orphans while we hold the lock. Best-effort:
+    // a read-only mount (no orphans reachable) just no-ops.
+    try {
+        (0, exports.sweepStagingArtifacts)(lockDir, opts.log);
+    }
+    catch {
+        /* best-effort */
+    }
+    return handle;
+};
+exports.acquireIndexLock = acquireIndexLock;