diff --git a/docs/ISSUE_reload_persistent_coroutine_segv.md b/docs/ISSUE_reload_persistent_coroutine_segv.md
new file mode 100644
index 0000000..5c7f9d0
--- /dev/null
+++ b/docs/ISSUE_reload_persistent_coroutine_segv.md
@@ -0,0 +1,154 @@
+# Known issue — SIGSEGV during pool reload when the bootloader spawns a long‑lived coroutine (#93)
+
+Status: reload **deadlock** is fixed (see "Fix that shipped"). The **intermittent
+crash** it exposes is now **root-caused and filed as an ext/async bug**:
+[true-async/php-async#176](https://github.com/true-async/php-async/issues/176).
+
+**Root cause (confirmed — code + ASAN + gdb):** `async_thread_create_closure`
+(`thread.c`) does `memcpy(&func, copy->func, sizeof(zend_op_array))` — a *shallow*
+copy of only the **top-level** op_array. Its `dynamic_func_defs` (nested closures
+declared inside the transferred closure, e.g. the `spawn(fn: while(true) delay())`
+inside the bootloader) are copied as a pointer → **all workers share the same
+nested op_array** from the persistent snapshot (proven: same `func` address
+`0x531…` on every worker; top-level closures are per-worker `0x7fff…`). Then
+`zend_create_closure` writes a per-thread arena `run_time_cache` **direct pointer**
+into that shared op_array (`zend_closures.c:817`), which is read from another
+thread / after that worker's arena is freed → dangling read → SEGV. Fix belongs
+in ext/async: deep-copy `dynamic_func_defs` per worker (like the top-level).
+Everything below is the raw investigation trail.
+
+---
+
+## 1. Background: the original deadlock (fixed)
+
+`HttpServer::reload()` under the built‑in worker pool (`mode=pool`, sync mode)
+rotates the cohort: it bumps the epoch, each worker retires (drain #74 → stop →
+`start()` returns), the ThreadPool `reload()` waits for every old worker to post
+its exit token, then submits the fresh cohort.
+
+A worker only posts its exit token after its scheduler has fully drained
+(`ZEND_ASYNC_RUN_SCHEDULER_AFTER_MAIN` in `thread_pool_worker_handler`). That
+drain loops while `has_handles` is true. If the **bootloader spawned a
+never‑ending coroutine into the worker's main scope** — e.g. laravel‑spawn's
+async DB pool arms a recurring `PDO::ATTR_POOL_HEALTHCHECK_INTERVAL` timer, or
+any `Async\spawn(fn: while(true) …)` — that coroutine's live reactor handle keeps
+`has_handles` true forever. Nothing cancels it on retire (drain #74 only touches
+`server_scope`, not the main scope), so the drain never completes, the exit token
+is never sent, and `ThreadPool::reload()` blocks forever. The user sees:
+
+```
+reload.start workers=8
+[true-async-server] worker thread exited cleanly   ×8   (old cohort — normal)
+<no reload.done, server stops serving>
+```
+
+Reproduced on Linux with the real laravel‑spawn app (the 8× "exited cleanly" is
+the normal old cohort retiring; the bug is the missing `reload.done`).
+
+## 2. Fix that shipped
+
+- **ext/async `thread_pool.c`** — at the worker's `done:` exit, for sync‑mode
+  pools only (`!pool->coroutine_mode`), call `start_graceful_shutdown()` before
+  `RUN_SCHEDULER_AFTER_MAIN`. It cancels the main‑scope stragglers and arms the
+  drain's escape valve, so the drain completes and the exit token is posted.
+  Gated to sync mode so coroutine‑mode pools keep draining their in‑flight task
+  coroutines (test `thread_pool/077`).
+- **server `http_server_class.c`** — lifecycle logs so the journal reads
+  `worker shutting down (reason=reload) → worker exited` instead of the
+  confusing bare `worker thread exited cleanly ×N`.
+
+Result: deadlock gone. **Real laravel‑spawn server: 10/10 reloads clean.**
+Server hot‑reload phpt 5/5, ThreadPool suite 78/79 (077 green via the gate),
+zero compiler warnings.
+
+## 3. The remaining problem: intermittent SIGSEGV under stress
+
+Under an aggressive synthetic harness (below), **~50 % of reloads SIGSEGV.** The
+real server did not hit it in 10 reloads, but it is a real latent crash.
+
+### Proven facts (gdb, release `-O2 -g` build)
+
+Backtrace is stable:
+
+```
+#0 ZEND_INIT_FCALL_SPEC_CONST_HANDLER  Zend/zend_vm_execute.h:4186
+       fbc = CACHED_PTR(opline->result.num);        // result.num = 0
+#1 execute_ex
+#2 zend_call_function                   Zend/zend_execute_API.c:1155
+#3 async_coroutine_execute              ext/async/coroutine.c:532   // first start (SET_STARTED just above)
+#4 fiber_entry                          ext/async/scheduler.c:1822
+```
+
+Inspecting the crashing coroutine:
+
+- It is the **bootloader‑spawned persistent coroutine itself** — its closure is
+  `{closure:{closure:…repro:25}:30}` (the `while(true){ delay() }`).
+- The closure and its op_array are **VALID**: `fci_cache.function_handler` is a
+  live `ZEND_USER_FUNCTION` (type 2), the closure object `refcount = 2`.
+  **It is NOT freed** (an earlier "op_array freed" guess was wrong).
+- Coroutine flags: `STARTED = 1` (set at coroutine.c:524 immediately before the
+  crash — i.e. this is its **first** execution), `CANCELLED = 0`, `MAIN = 0`.
+- The fault: `CACHED_PTR(0)` dereferences a **NULL run‑time cache**, even though
+  `op_array.cache_size = 8`. So the op_array expects a run‑time cache but the
+  `ZEND_MAP_PTR` resolves to NULL in this execution.
+
+### What was ruled out
+
+- **"Coroutine never ticked before reload"** — ruled out. A first‑run probe
+  shows both workers' persistent coroutines run *during serving* (before the
+  reload trigger). The crash happens later, in the rotation phase (a 3rd
+  first‑run — the replacement worker re‑booting the bootloader — is observed
+  after `reload.start`, then the crash).
+- **Cancelled‑coroutine‑executed** — ruled out; the crashing coroutine has
+  `CANCELLED = 0`.
+
+### Failed fix attempt (do not repeat blindly)
+
+Cancelling the main scope earlier, server‑side, in `start()`'s post‑wakeup path
+(right after the #74 `server_scope` drain, on the still‑valid `start()`
+coroutine, with the root coroutine `SET_PROTECTED`) made the crash **worse**
+(~75 %). So the crash is **not** simply "the context is torn down by the time we
+reach the pool `done:`": executing/cancelling this coroutine is unsafe at both
+points. Reverted.
+
+### Open questions for next session
+
+1. Why does a **cloned/transferred closure** (created from the bootloader
+   snapshot in the worker) have a **NULL run‑time cache** on the execution that
+   crashes, while the *same* coroutine executed fine during serving?
+2. Who **fresh‑starts** this coroutine during the rotation drain, and why is it a
+   fresh start (not a resume) when it already ran during serving? Is there a
+   second instance (replacement worker) vs the old instance being re‑entered?
+3. Is the fix better placed in ext/async (make the drain skip/deallocate
+   never‑resumable coroutines, or ensure the run‑time cache is bound before
+   execution) rather than in the server?
+
+## 4. Reproduction
+
+Build the release stack (php‑src `true-async-stable` + ext/async `main` →
+`/home/edmond/php-release`, then the server extension against it). Run the
+synthetic harness — a trivial bootloader that leaves one persistent coroutine,
+which is all it takes:
+
+```php
+// repro_persistent.php  (2 workers, watcher hot‑reload)
+$config = (new HttpServerConfig())
+    ->addListener('127.0.0.1', (int)$argv[1])
+    ->setWorkers(2)
+    ->setLogSeverity(TrueAsync\LogSeverity::INFO)->setLogStream(STDERR)
+    ->setBootloader(static function () use ($code) {
+        define('APP_V', (string) include $code);
+        Async\spawn(static function () {          // <-- the trigger: a never‑ending
+            while (true) { Async\delay(1000); }   //     main‑scope coroutine
+        });
+    })
+    ->enableHotReload([$dir], ['php'], 150, 1000);
+// a spawned driver curls until v1, rewrites the watched file to v2, curls for v2.
+```
+
+Loop it on fresh ports; ~half the runs SIGSEGV (rc 139), the rest complete with
+`reload.done` + `after=v2`. The crash needs the parent reactor busy (the driver's
+blocking `shell_exec(curl)` on the parent supplies the timing); the real server,
+whose requests come from external processes, has not reproduced it.
+
+Full harness + gdb logs from the investigation live under the session scratchpad.
diff --git a/docs/repro_reload_persistent_coroutine_segv.php b/docs/repro_reload_persistent_coroutine_segv.php
new file mode 100644
index 0000000..e980b28
--- /dev/null
+++ b/docs/repro_reload_persistent_coroutine_segv.php
@@ -0,0 +1,58 @@
+<?php
+/* Proof: a bootloader that leaves a PERSISTENT async resource on the worker
+ * reactor (a never-ending background coroutine with a timer) — mimicking the
+ * laravel-spawn async DB pool's healthcheck timer + pooled connections. If
+ * reload() deadlocks here while the trivial define()-only bootloader works,
+ * the cause is confirmed: long-lived reactor handles block worker retirement. */
+
+use TrueAsync\HttpServer;
+use TrueAsync\HttpServerConfig;
+use TrueAsync\LogSeverity;
+use function Async\spawn;
+
+$port = (int) ($argv[1] ?? 18200);
+$mode = $argv[2] ?? 'persistent';   // 'persistent' or 'trivial'
+$dir  = sys_get_temp_dir() . '/tas_persist_' . getmypid();
+@mkdir($dir, 0777, true);
+$code = $dir . '/version.php';
+file_put_contents($code, "<?php return 'v1';");
+
+$config = (new HttpServerConfig())
+    ->addListener('127.0.0.1', $port)
+    ->setReadTimeout(5)->setWriteTimeout(5)
+    ->setWorkers(2)
+    ->setLogSeverity(LogSeverity::INFO)->setLogStream(STDERR)
+    ->setBootloader(static function () use ($code, $mode): void {
+        define('APP_V', (string) include $code);
+        if ($mode === 'persistent') {
+            // Persistent background coroutine — arms a recurring timer on the
+            // worker reactor that outlives any per-request scope.
+            spawn(static function () {
+                while (true) {
+                    \Async\delay(1000);
+                }
+            });
+        }
+    })
+    ->enableHotReload([$dir], ['php'], 150, 1000);
+
+$server = new HttpServer($config);
+$server->addHttpHandler(function ($req, $res) {
+    $res->setStatusCode(200)->setBody(APP_V);
+});
+
+spawn(function () use ($port, $code, $dir) {
+    $curl = static fn (): string => (string) shell_exec(sprintf(
+        'curl -s --max-time 2 http://127.0.0.1:%d/ 2>/dev/null', $port));
+    $v1 = '';
+    for ($i = 0; $i < 50 && $v1 !== 'v1'; $i++) { usleep(200000); $v1 = $curl(); }
+    fwrite(STDOUT, "PROOF before=$v1\n");
+    file_put_contents($code, "<?php return 'v2';");
+    $v2 = '';
+    for ($i = 0; $i < 60 && $v2 !== 'v2'; $i++) { usleep(200000); $v2 = $curl(); }
+    fwrite(STDOUT, "PROOF after=$v2 (iterations=$i)\n");
+    @unlink($code); @rmdir($dir);
+    posix_kill(getmypid(), SIGKILL);
+});
+
+$server->start();
diff --git a/php_true_async_server.h b/php_true_async_server.h
index d45cad6..7484d2d 100644
--- a/php_true_async_server.h
+++ b/php_true_async_server.h
@@ -24,8 +24,8 @@
  *
  * It is kept flat at the extension root (not under include/) on purpose:
  * php-src's static-build codegen (build/genif.sh + build/print_include.awk)
- * discovers each bundled extension's module_entry by globbing
- * ext/<name>/*.h non-recursively and grepping the matched files for a
+ * discovers each bundled extension's module_entry by globbing the
+ * ext/<name> directory for *.h files non-recursively and grepping them for a
  * "phpext_" line — it does not follow nested #include paths. Keeping the
  * module-entry declaration here, at ext/server/php_true_async_server.h,
  * lets static (non-shared) builds such as static-php-cli find
diff --git a/src/http_server_class.c b/src/http_server_class.c
index 276a82c..20ad9d6 100644
--- a/src/http_server_class.c
+++ b/src/http_server_class.c
@@ -609,6 +609,9 @@ static void http_server_reload_stop_entry(void)
     http_server_worker_inbox_retire(server);
 
     server->stopping = true;
+    fprintf(stderr, "[true-async-server] worker shutting down (reason=reload, grace=%us)\n",
+            server->shutdown_timeout_s);
+    fflush(stderr);
     http_server_do_stop(server, "reload");
 }
 
@@ -1927,11 +1930,12 @@ static void pool_worker_handler(zend_async_event_t *event, void *vctx)
             fflush(stderr);
             zend_clear_exception();
         } else {
-            /* Even no-exception exit while server should still be running
-             * is suspicious — e.g. a bailout that left EG(exception) NULL
-             * (OOM during cleanup). Log a lighter notice. */
+            /* Clean exit — the normal path when a worker retires on reload
+             * (paired with the "worker shutting down" line above) or on a
+             * graceful stop. A clean exit while the server should still be
+             * running would instead point at a swallowed bailout. */
             fprintf(stderr,
-                "[true-async-server] worker thread exited cleanly\n");
+                "[true-async-server] worker exited\n");
             fflush(stderr);
         }
     } else {