@scriptc/runtime 0.0.0 → 0.0.2

package/src/scr_watch.c

@@ -0,0 +1,568 @@
+/* fs.watch — the OPTIONAL file-watching half of the event loop: FSWatcher
+ * handles over the unit's own kqueue (EVFILT_VNODE on an O_EVTONLY fd).
+ * This translation unit links ONLY into binaries whose IR uses fs.watch
+ * (moduleUsesFsWatch — the scr_events/scr_net gating precedent), and the
+ * emitted main calls scr_watch_install() before %main, which points the
+ * loop's nullable watch hooks (scr_async.c) here. Watch-free programs pay
+ * zero bytes and keep their exact link line.
+ *
+ * Semantics, matched against Node on macOS and pinned by corpus:
+ * - fs.watch(path, cb) THROWS Node's fs error synchronously when the path
+ *   cannot be opened ("ENOENT: no such file or directory, watch 'x'" —
+ *   the portless polling-fallback catch shape).
+ * - The callback fires with "change" (NOTE_WRITE/EXTEND/ATTRIB) or
+ *   "rename" (NOTE_RENAME/DELETE/REVOKE — rename wins when a drain
+ *   coalesces both, libuv's precedence).
+ * - An open watcher KEEPS THE LOOP ALIVE until close() (Node's persistent
+ *   default; unref has no lowering — nothing exercises it).
+ * - close() is idempotent: the fd closes (dropping its kqueue filter),
+ *   listeners release, nothing fires again.
+ *
+ * Platform honesty: kqueue watches the INODE behind the opened fd, not
+ * the name — the same stance as libuv's kqueue backend. A
+ * rewrite-in-place (writeFileSync truncate, append) delivers "change"
+ * reliably. A replace-by-rename delivers ONE "rename" for the old inode,
+ * after which this unit REOPENS the path best-effort: when a new file
+ * took the name (the atomic-rewrite idiom), watching continues on the
+ * new inode; when the name is gone, the watcher stays registered but
+ * silent (Node's kqueue backend goes silent the same way — its FSEvents
+ * backend on macOS would keep reporting, a documented divergence).
+ *
+ * The Linux arm is inotify, matched against Linux Node (the differential
+ * container's oracle — probed, not assumed): one inotify fd for the
+ * unit, one wd per watcher, libuv's exact event mapping — a mask with
+ * IN_ATTRIB or IN_MODIFY fires "change", ANY other delivered mask fires
+ * "rename", including IN_DELETE_SELF and the trailing IN_IGNORED (Linux
+ * Node really does fire two 'rename's for an unlink, and a leading
+ * 'change' for the link-count IN_ATTRIB). After IN_IGNORED the wd is
+ * dead and the watcher goes silent but stays open (holds the loop) until
+ * close() — Node does NOT reopen on Linux, so neither does this arm; the
+ * kqueue reopen dance is macOS-only, exactly like libuv. The wd → watcher
+ * route trusts the kernel's cyclic wd allocation (idr_alloc_cyclic): a
+ * close()-queued IN_IGNORED can never collide with a freshly issued wd.
+ *
+ * The Windows arm is ReadDirectoryChangesW (see the SCR_HAVE_RDCW block
+ * below): directory-handle watching by NAME, so replace-by-rename keeps
+ * delivering without any reopen dance — Node-on-Windows's own stance.
+ *
+ * Event COALESCING is delivery-granular everywhere: one drain pass may
+ * merge what Node reports as two events — corpus programs observe "fired
+ * at least once", never exact counts. */
+#define _XOPEN_SOURCE 700
+#define _DARWIN_C_SOURCE 1
+#include "scr_runtime.h"
+
+#include <errno.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+#if defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || \
+    defined(__OpenBSD__) || defined(__DragonFly__)
+#define SCR_HAVE_KQUEUE 1
+#include <sys/event.h>
+#elif defined(__linux__)
+#define SCR_HAVE_INOTIFY 1
+#include <sys/inotify.h>
+#elif defined(_WIN32)
+/* The Windows arm is ReadDirectoryChangesW, libuv's exact shape: a FILE
+ * watch opens its PARENT directory and filters deliveries by basename, a
+ * DIRECTORY watch opens the path itself (non-recursive, fs.watch's
+ * default); one overlapped read per watcher is outstanding at a time,
+ * its manual-reset event polled at dispatch (the loop's win32 idle sleep
+ * caps while watchers pend — scr_async.c). Action mapping matches libuv:
+ * FILE_ACTION_MODIFIED fires "change", every other action (added/
+ * removed/renamed either direction) fires "rename". Directory-handle
+ * watching survives replace-by-rename without the kqueue reopen dance —
+ * the NAME is watched, not the inode, which is Node-on-Windows's own
+ * stance. */
+#define SCR_HAVE_RDCW 1
+#include <stdalign.h>
+#include <windows.h>
+#endif
+
+#ifndef O_EVTONLY
+#define O_EVTONLY O_RDONLY /* non-macOS: a plain read fd watches fine */
+#endif
+
+static void scr_watch_oom(void) {
+  fputs("scriptc: out of memory\n", stderr);
+  abort();
+}
+
+typedef struct {
+  ScrClosure *cb;  /* owned */
+  ScrWatchFn fn;   /* adapter: event name borrowed */
+} ScrWatchL;
+
+struct ScrWatcher {
+  size_t rc;
+  int fd;       /* -1 once closed, or after a failed reopen (silent) */
+#ifdef SCR_HAVE_INOTIFY
+  int wd;       /* the inotify watch descriptor; -1 once dead (IN_IGNORED
+                 * arrived or close() removed it) */
+#endif
+#ifdef SCR_HAVE_RDCW
+  HANDLE dh;       /* the watched directory handle; INVALID once closed or
+                    * silent (a failed reissue) */
+  OVERLAPPED ovl;  /* the one outstanding ReadDirectoryChangesW; hEvent is
+                    * a manual-reset event polled at dispatch */
+  wchar_t *filter; /* file watches: the watched basename (UTF-16, ordinal
+                    * case-insensitive compare); NULL for directory watches */
+  bool inflight;   /* a read is outstanding (ovl/buf are the kernel's) */
+  alignas(DWORD) char buf[4096]; /* FILE_NOTIFY_INFORMATION entries */
+#endif
+  bool closed;  /* close() called: off the liveness count, listeners gone */
+  ScrStr *path; /* the watched path — the reopen-after-rename target */
+  ScrWatchL *ls;
+  size_t n, cap;
+  struct ScrWatcher *next; /* the open-watcher registry (+1 each) */
+};
+
+static ScrWatcher *scr_watchers = NULL;
+static size_t scr_watchers_open = 0; /* liveness: open (unclosed) watchers */
+
+#ifdef SCR_HAVE_KQUEUE
+static int scr_watch_kq = -1; /* created with the first watch; lives forever */
+
+/* Arms the vnode filter for a watcher's current fd. EV_CLEAR: state resets
+ * at delivery, so an unserviced event never busy-spins the idle poll (the
+ * dispatch drain is the single consumer). */
+static bool scr_watch_arm(ScrWatcher *w) {
+  if (scr_watch_kq < 0) {
+    scr_watch_kq = kqueue();
+    if (scr_watch_kq < 0) return false;
+  }
+  struct kevent ev;
+  EV_SET(&ev, (uintptr_t)w->fd, EVFILT_VNODE, EV_ADD | EV_CLEAR,
+         NOTE_WRITE | NOTE_EXTEND | NOTE_ATTRIB | NOTE_RENAME | NOTE_DELETE | NOTE_REVOKE,
+         0, w);
+  struct timespec zero = {0, 0};
+  return kevent(scr_watch_kq, &ev, 1, NULL, 0, &zero) == 0;
+}
+#endif
+
+#ifdef SCR_HAVE_INOTIFY
+static int scr_watch_ino = -1; /* created with the first watch; lives forever */
+
+/* libuv's subscription mask: name-level events for directory watches,
+ * self events for the file case, IN_IGNORED arriving for free. */
+#define SCR_WATCH_IN_MASK                                                        \
+  (IN_ATTRIB | IN_CREATE | IN_MODIFY | IN_DELETE | IN_DELETE_SELF |              \
+   IN_MOVE_SELF | IN_MOVED_FROM | IN_MOVED_TO)
+
+static ScrWatcher *scr_watch_by_wd(int wd) {
+  if (wd < 0) return NULL;
+  for (ScrWatcher *w = scr_watchers; w != NULL; w = w->next) {
+    if (!w->closed && w->wd == wd) return w;
+  }
+  return NULL;
+}
+#endif
+
+#ifdef SCR_HAVE_RDCW
+static wchar_t *scr_watch_widen(const char *s) {
+  int n = MultiByteToWideChar(CP_UTF8, 0, s, -1, NULL, 0);
+  if (n <= 0) return NULL;
+  wchar_t *w = malloc((size_t)n * sizeof *w);
+  if (!w) scr_watch_oom();
+  MultiByteToWideChar(CP_UTF8, 0, s, -1, w, n);
+  return w;
+}
+
+/* Issues the next overlapped read. False leaves the watcher SILENT (the
+ * kqueue arm's exhaustion stance): open, holding the loop, never firing. */
+static bool scr_watch_issue(ScrWatcher *w) {
+  ResetEvent(w->ovl.hEvent);
+  if (!ReadDirectoryChangesW(w->dh, w->buf, (DWORD)sizeof w->buf, FALSE,
+                             FILE_NOTIFY_CHANGE_FILE_NAME | FILE_NOTIFY_CHANGE_DIR_NAME |
+                                 FILE_NOTIFY_CHANGE_ATTRIBUTES | FILE_NOTIFY_CHANGE_SIZE |
+                                 FILE_NOTIFY_CHANGE_LAST_WRITE | FILE_NOTIFY_CHANGE_CREATION |
+                                 FILE_NOTIFY_CHANGE_SECURITY,
+                             NULL, &w->ovl, NULL)) {
+    return false;
+  }
+  w->inflight = true;
+  return true;
+}
+
+/* Tears down the win32 backend state — close(), the leak path in
+ * release(), and the atexit cleanup all funnel here; idempotent. */
+static void scr_watch_backend_drop(ScrWatcher *w) {
+  if (w->dh != INVALID_HANDLE_VALUE) {
+    CancelIo(w->dh); /* the queued completion dies with the handle */
+    CloseHandle(w->dh);
+    w->dh = INVALID_HANDLE_VALUE;
+  }
+  if (w->ovl.hEvent != NULL) {
+    CloseHandle(w->ovl.hEvent);
+    w->ovl.hEvent = NULL;
+  }
+  w->inflight = false;
+  free(w->filter);
+  w->filter = NULL;
+}
+#endif
+
+ScrWatcher *scr_watcher_retain(ScrWatcher *w) {
+  if (w->rc != SIZE_MAX) w->rc++;
+  return w;
+}
+
+static void scr_watcher_drop_listeners(ScrWatcher *w) {
+  for (size_t i = 0; i < w->n; i++) scr_closure_release(w->ls[i].cb);
+  free(w->ls);
+  w->ls = NULL;
+  w->n = w->cap = 0;
+}
+
+void scr_watcher_release(ScrWatcher *w) {
+  if (!w || w->rc == SIZE_MAX) return;
+  if (--w->rc == 0) {
+    scr_watcher_drop_listeners(w); /* only reachable pre-close via leaks */
+    if (w->fd >= 0) close(w->fd);
+#ifdef SCR_HAVE_RDCW
+    scr_watch_backend_drop(w);
+#endif
+    scr_str_release(w->path);
+    free(w);
+  }
+}
+
+void *scr_watcher_retain_v(void *p) { return scr_watcher_retain((ScrWatcher *)p); }
+void scr_watcher_release_v(void *p) { scr_watcher_release((ScrWatcher *)p); }
+
+/* fs.watch(path, cb?) — opens the path NOW (throwing Node's fs error on
+ * failure), registers the watcher with the loop, and arms the vnode
+ * filter. cb (nullable) moves in as the first listener. Result +1. */
+ScrWatcher *scr_fs_watch(ScrStr *path, ScrClosure *cb /*moves, nullable*/, ScrWatchFn fn) {
+#ifdef SCR_HAVE_INOTIFY
+  /* The wd IS the validation: Node's Linux throw comes straight from
+   * inotify_add_watch's errno, the same message shape as the open below. */
+  if (scr_watch_ino < 0) scr_watch_ino = inotify_init1(IN_NONBLOCK | IN_CLOEXEC);
+  int wd = scr_watch_ino >= 0 ? inotify_add_watch(scr_watch_ino, path->data, SCR_WATCH_IN_MASK)
+                              : -1;
+  if (wd < 0) {
+    if (cb != NULL) scr_closure_release(cb);
+    scr_fs_throw(errno, "watch", path);
+    return NULL; /* exception pending */
+  }
+#elif defined(SCR_HAVE_RDCW)
+  /* Existence is the validation (Node's win32 throw): a missing path is
+   * the synchronous ENOENT. A file watch then opens the PARENT directory
+   * and remembers the basename to filter deliveries by. */
+  wchar_t *wpath = scr_watch_widen(path->data);
+  DWORD attrs = wpath != NULL ? GetFileAttributesW(wpath) : INVALID_FILE_ATTRIBUTES;
+  if (attrs == INVALID_FILE_ATTRIBUTES) {
+    free(wpath);
+    if (cb != NULL) scr_closure_release(cb);
+    scr_fs_throw(ENOENT, "watch", path);
+    return NULL; /* exception pending */
+  }
+  wchar_t *wfilter = NULL;
+  if ((attrs & FILE_ATTRIBUTE_DIRECTORY) == 0) {
+    wchar_t *base = NULL;
+    for (wchar_t *p = wpath; *p != L'\0'; p++)
+      if (*p == L'\\' || *p == L'/') base = p;
+    if (base != NULL) {
+      wfilter = _wcsdup(base + 1);
+      if (wfilter == NULL) scr_watch_oom();
+      *base = L'\0'; /* wpath is now the parent directory */
+    } else {
+      wfilter = wpath; /* bare name: watch the cwd, filter by it */
+      wpath = _wcsdup(L".");
+      if (wpath == NULL) scr_watch_oom();
+    }
+  }
+  HANDLE dh = CreateFileW(wpath, FILE_LIST_DIRECTORY,
+                          FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, NULL,
+                          OPEN_EXISTING, FILE_FLAG_BACKUP_SEMANTICS | FILE_FLAG_OVERLAPPED,
+                          NULL);
+  free(wpath);
+  if (dh == INVALID_HANDLE_VALUE) {
+    free(wfilter);
+    if (cb != NULL) scr_closure_release(cb);
+    scr_fs_throw(ENOENT, "watch", path);
+    return NULL; /* exception pending */
+  }
+#else
+  int fd = open(path->data, O_EVTONLY | O_NONBLOCK | O_CLOEXEC);
+  if (fd < 0) {
+    if (cb != NULL) scr_closure_release(cb);
+    scr_fs_throw(errno, "watch", path);
+    return NULL; /* exception pending */
+  }
+#endif
+  ScrWatcher *w = calloc(1, sizeof *w);
+  if (!w) scr_watch_oom();
+  w->rc = 1;
+  w->path = scr_str_retain(path);
+  if (cb != NULL) {
+    w->ls = malloc(sizeof *w->ls);
+    if (!w->ls) scr_watch_oom();
+    w->ls[0].cb = cb;
+    w->ls[0].fn = fn;
+    w->n = w->cap = 1;
+  }
+#ifdef SCR_HAVE_INOTIFY
+  w->fd = -1; /* inotify holds the watch; there is no per-watcher fd */
+  w->wd = wd;
+#elif defined(SCR_HAVE_KQUEUE)
+  w->fd = fd;
+  if (!scr_watch_arm(w)) {
+    /* Filter refused (kqueue exhaustion): the watcher exists but never
+     * fires — Node's kqueue backend surfaces this as a throw; the honest
+     * cheap stance here is silence plus liveness, and nothing real hits
+     * it. */
+    close(w->fd);
+    w->fd = -1;
+  }
+#elif defined(SCR_HAVE_RDCW)
+  w->fd = -1;
+  w->dh = dh;
+  w->filter = wfilter;
+  w->ovl.hEvent = CreateEventW(NULL, TRUE, FALSE, NULL);
+  if (w->ovl.hEvent == NULL || !scr_watch_issue(w)) {
+    /* the kqueue exhaustion stance: open, holding the loop, silent */
+    scr_watch_backend_drop(w);
+  }
+#else
+  close(fd); /* no backend: open validated the path; no events exist */
+  w->fd = -1;
+#endif
+  w->next = scr_watchers;
+  scr_watchers = scr_watcher_retain(w);
+  scr_watchers_open++;
+  return w;
+}
+
+/* watcher.close() — idempotent: drops the filter (closing the fd deletes
+ * its knote AND its queued events), the listeners, and the liveness
+ * count. The registry's reference is NOT released here: a listener may
+ * close its own watcher while the dispatch drain still holds the pointer
+ * in its current event batch — the sweep between batches (and the atexit
+ * cleanup) unlinks closed watchers at safe points instead. */
+void scr_watcher_close(ScrWatcher *w) {
+  if (w->closed) return;
+  w->closed = true;
+#ifdef SCR_HAVE_INOTIFY
+  if (w->wd >= 0) {
+    /* The queued IN_IGNORED this rm produces finds no open watcher and is
+     * skipped; wd numbers are issued cyclically, so it cannot collide
+     * with a watch created afterwards. */
+    if (scr_watch_ino >= 0) (void)inotify_rm_watch(scr_watch_ino, w->wd);
+    w->wd = -1;
+  }
+#endif
+#ifdef SCR_HAVE_RDCW
+  scr_watch_backend_drop(w);
+#endif
+  if (w->fd >= 0) {
+    close(w->fd);
+    w->fd = -1;
+  }
+  scr_watcher_drop_listeners(w);
+  scr_watchers_open--;
+}
+
+/* Unlinks and releases closed watchers — called only where no fetched
+ * event batch can still point at them. */
+static void scr_watch_sweep(void) {
+  ScrWatcher **link = &scr_watchers;
+  while (*link) {
+    ScrWatcher *w = *link;
+    if (w->closed) {
+      *link = w->next;
+      w->next = NULL;
+      scr_watcher_release(w); /* the registry's reference */
+    } else {
+      link = &w->next;
+    }
+  }
+}
+
+/* The runtime-provided listener adapters. */
+void scr_watch_thunk0(ScrClosure *cb, const char *event) {
+  (void)event;
+  ((void (*)(ScrClosure *))cb->fn)(cb);
+}
+void scr_watch_thunk_event(ScrClosure *cb, const char *event) {
+  /* The listener owns its +1 param per the universal convention. */
+  ScrStr *e = scr_str_new(event, strlen(event));
+  ((void (*)(ScrClosure *, ScrStr *))cb->fn)(cb, e);
+}
+
+/* Fires one watcher's listeners with the event name (snapshot — a
+ * listener closing the watcher mid-emit still finishes this delivery,
+ * and close() drops only the LIVE list). */
+static void scr_watch_fire(ScrWatcher *w, const char *event) {
+  size_t n = w->n;
+  if (n == 0) return;
+  ScrWatchL *snap = malloc(n * sizeof *snap);
+  if (!snap) scr_watch_oom();
+  for (size_t i = 0; i < n; i++) {
+    snap[i] = w->ls[i];
+    scr_closure_retain(snap[i].cb);
+  }
+  for (size_t i = 0; i < n; i++) {
+    if (!scr_exc_pending()) snap[i].fn(snap[i].cb, event);
+    scr_closure_release(snap[i].cb);
+  }
+  free(snap);
+}
+
+/* ── the loop hooks (scr_async.c) ────────────────────────────────────── */
+
+/* Liveness: every un-closed watcher holds the loop (Node's persistent
+ * default) — a silent (fd < 0) watcher included, exactly like Node's. */
+static bool scr_watch_pending(void) { return scr_watchers_open > 0; }
+
+static int scr_watch_pollfd(void) {
+#ifdef SCR_HAVE_KQUEUE
+  return scr_watch_kq;
+#elif defined(SCR_HAVE_INOTIFY)
+  return scr_watch_ino; /* readable while events pend, like a kqueue fd */
+#else
+  return -1; /* win32: nothing pollable — the loop's capped sleep serves */
+#endif
+}
+
+static void scr_watch_dispatch(void) {
+#ifdef SCR_HAVE_KQUEUE
+  if (scr_watch_kq < 0) {
+    scr_watch_sweep();
+    return;
+  }
+  for (;;) {
+    scr_watch_sweep();
+    struct kevent evs[64];
+    struct timespec zero = {0, 0};
+    int n = kevent(scr_watch_kq, NULL, 0, evs, 64, &zero);
+    if (n <= 0) return;
+    for (int i = 0; i < n; i++) {
+      ScrWatcher *w = (ScrWatcher *)evs[i].udata;
+      if (w == NULL || w->closed || w->fd < 0) continue;
+      uint32_t ff = evs[i].fflags;
+      bool rename = (ff & (NOTE_RENAME | NOTE_DELETE | NOTE_REVOKE)) != 0;
+      if (rename) {
+        /* The inode left the name: reopen best-effort so the atomic
+         * replace-by-rename idiom keeps delivering on the new file. The
+         * OLD fd closes first (its filter dies with it) — a coalesced
+         * change on the old inode is superseded by the rename. */
+        close(w->fd);
+        w->fd = open(w->path->data, O_EVTONLY | O_NONBLOCK | O_CLOEXEC);
+        if (w->fd >= 0 && !scr_watch_arm(w)) {
+          close(w->fd);
+          w->fd = -1;
+        }
+        scr_watch_fire(w, "rename");
+      } else if (ff & (NOTE_WRITE | NOTE_EXTEND | NOTE_ATTRIB)) {
+        scr_watch_fire(w, "change");
+      }
+      if (scr_exc_pending()) return; /* the loop surfaces it */
+    }
+    if (scr_loop_has_ready()) return; /* microtasks interleave first */
+  }
+#elif defined(SCR_HAVE_INOTIFY)
+  if (scr_watch_ino < 0) {
+    scr_watch_sweep();
+    return;
+  }
+  for (;;) {
+    scr_watch_sweep();
+    char buf[4096] __attribute__((aligned(__alignof__(struct inotify_event))));
+    ssize_t n = read(scr_watch_ino, buf, sizeof buf);
+    if (n <= 0) return; /* EAGAIN/EINTR: the next turn retries */
+    for (char *p = buf; p < buf + n;) {
+      const struct inotify_event *e = (const struct inotify_event *)p;
+      p += sizeof *e + e->len;
+      ScrWatcher *w = scr_watch_by_wd(e->wd);
+      if (w == NULL) continue; /* closed, or a stale post-close IN_IGNORED */
+      /* libuv's mapping, probed against Linux Node: IN_ATTRIB/IN_MODIFY
+       * fire "change"; every other delivered mask — IN_DELETE_SELF,
+       * IN_MOVE_SELF, the name-level events, and IN_IGNORED itself —
+       * fires "rename". */
+      bool change = (e->mask & (IN_ATTRIB | IN_MODIFY)) != 0;
+      if (e->mask & IN_IGNORED) w->wd = -1; /* the kernel dropped the watch:
+                                             * silent (no reopen), like Node */
+      scr_watch_fire(w, change ? "change" : "rename");
+      if (scr_exc_pending()) return; /* the loop surfaces it */
+    }
+    if (scr_loop_has_ready()) return; /* microtasks interleave first */
+  }
+#elif defined(SCR_HAVE_RDCW)
+  scr_watch_sweep();
+  for (ScrWatcher *w = scr_watchers; w != NULL; w = w->next) {
+    if (w->closed || w->dh == INVALID_HANDLE_VALUE || !w->inflight) continue;
+    if (WaitForSingleObject(w->ovl.hEvent, 0) != WAIT_OBJECT_0) continue;
+    DWORD bytes = 0;
+    BOOL ok = GetOverlappedResult(w->dh, &w->ovl, &bytes, FALSE);
+    w->inflight = false;
+    /* Deliver from THIS buffer before reissuing (the reissue hands buf
+     * back to the kernel). libuv's action mapping: MODIFIED is "change",
+     * everything else — added, removed, renamed either direction — is
+     * "rename". File watches filter by basename (ordinal,
+     * case-insensitive — NTFS name matching); zero bytes is the
+     * overflow notification, delivered as a filterless "rename" burst
+     * signal the same way libuv degrades. */
+    if (ok && bytes == 0) {
+      scr_watch_fire(w, "rename");
+      if (scr_exc_pending()) return;
+    } else if (ok) {
+      char *p = w->buf;
+      for (;;) {
+        FILE_NOTIFY_INFORMATION *e = (FILE_NOTIFY_INFORMATION *)p;
+        bool match =
+            w->filter == NULL ||
+            CompareStringOrdinal(e->FileName, (int)(e->FileNameLength / sizeof(WCHAR)),
+                                 w->filter, -1, TRUE) == CSTR_EQUAL;
+        if (match) {
+          scr_watch_fire(w, e->Action == FILE_ACTION_MODIFIED ? "change" : "rename");
+          if (scr_exc_pending()) return;
+          if (w->closed) break; /* a listener closed its own watcher */
+        }
+        if (e->NextEntryOffset == 0) break;
+        p += e->NextEntryOffset;
+      }
+    }
+    if (!w->closed && w->dh != INVALID_HANDLE_VALUE && !scr_watch_issue(w)) {
+      scr_watch_backend_drop(w); /* reissue refused: silent, like kqueue exhaustion */
+    }
+  }
+  scr_watch_sweep();
+#endif
+}
+
+/* Exit-time registry cleanup (the events/net-unit precedent): watchers a
+ * program legitimately leaves open at exit release their listeners and
+ * registry references so the RC audit sees a clean heap. */
+static void scr_watch_cleanup_atexit(void) {
+  while (scr_watchers) {
+    ScrWatcher *w = scr_watchers;
+    scr_watchers = w->next;
+    w->next = NULL;
+    if (!w->closed) {
+      w->closed = true;
+      scr_watchers_open--;
+    }
+#ifdef SCR_HAVE_RDCW
+    scr_watch_backend_drop(w);
+#endif
+    if (w->fd >= 0) {
+      close(w->fd);
+      w->fd = -1;
+    }
+    scr_watcher_drop_listeners(w);
+    scr_watcher_release(w);
+  }
+}
+
+void scr_watch_install(void) {
+  static bool installed = false;
+  if (installed) return;
+  installed = true;
+  atexit(scr_watch_cleanup_atexit);
+  scr_loop_set_watch(&scr_watch_pending, &scr_watch_dispatch, &scr_watch_pollfd);
+}