diff --git a/library/modules/std/sched.sx b/library/modules/std/sched.sx
index 97970b7a..60f337fd 100644
--- a/library/modules/std/sched.sx
+++ b/library/modules/std/sched.sx
@@ -613,7 +613,19 @@ impl Io for Scheduler {
     // `*void`; `arg` is its single payload pointer. Returns the `*Fiber` as the
     // opaque handle (the `Future` stashes it in `task`). The fiber inherits this
     // call's context (Phase 0), so the worker sees this scheduler as `context.io`.
+    // CONTRACT: `arg` must outlive the fiber's FIRST RUN (i.e. survive until
+    // `run`/`poll` schedules it), not merely until `spawn_raw` returns — the fiber
+    // dereferences it later, on its own stack. A stack local that dies before the
+    // driver loop runs is a use-after-free (the `async` layer keeps it alive by
+    // owning it on the heap `Future`).
     spawn_raw :: (self: *Scheduler, entry: *void, arg: *void, opts: SpawnOpts) -> *void {
+        // A null entry would jump to 0x0 once the fiber runs — bail loudly at the
+        // spawn site (where the caller is) instead of crashing deep in the loop.
+        // (A non-fn garbage pointer is undetectable — the caller owns that.)
+        if entry == null {
+            print("sched: spawn_raw called with a null entry function\n");
+            abort();
+        }
         f := self.spawn(() => {
             entry_fn : (*void) -> void = xx entry;
             entry_fn(arg);
@@ -640,10 +652,20 @@ impl Io for Scheduler {
 
     // Advance the world one step (the unit `run` loops over). Drain every
     // currently-ready fiber, then fire the earliest pending timer (advancing the
-    // virtual clock). Returns the new clock after a timer fire, or -1 when nothing
-    // ran and no timer is pending. NOTE: fd-readiness (kqueue/epoll) blocking is
-    // NOT driven here yet — `run` remains the canonical driver for fd-bound work;
-    // `poll` covers the compute + virtual-timer path the async layer needs.
+    // virtual clock). Returns the new virtual clock after a timer fire, or -1 when
+    // the ready queue AND the timer set are both empty (the program is quiescent).
+    //
+    // `deadline_ms` is RESERVED — the virtual-time single-step jumps the clock to
+    // the earliest timer regardless of any real-time budget, so there is nothing
+    // for a deadline to bound here (it matches `CBlockingIo.poll`'s stub shape; a
+    // real-clock `Io` would honor it). Named for protocol conformance, intentionally
+    // unread — NOT a silently-dropped argument.
+    //
+    // fd-readiness (kqueue/epoll) is NOT driven by `poll` — `run` remains the
+    // canonical driver for fd-bound work. To avoid the silent-wrong-answer where a
+    // `-1` ("quiescent") is indistinguishable from "fibers are fd-blocked", bail
+    // loudly if any io-waiter is pending: a `poll`-only driver can't make progress
+    // on fds and must use `run` instead.
     poll :: (self: *Scheduler, deadline_ms: i64) -> i64 {
         while self.ready_head != null {
             f := dequeue(self);
@@ -666,6 +688,10 @@ impl Io for Scheduler {
             self.wake(t.fiber);
             return self.clock_ms;
         }
+        if self.io_waiters.len > 0 {
+            print("sched: poll() cannot drive fd-readiness — use run() for fd-bound fibers\n");
+            abort();
+        }
         return -1;
     }
 
@@ -678,6 +704,13 @@ impl Io for Scheduler {
     // null (a timer-cancel handle is Phase 3). The fired timer wakes the fiber
     // through the same `wake` path the run loop uses.
     arm_timer :: (self: *Scheduler, deadline_ms: i64, park: ParkToken) -> *void {
+        // MUST be called from inside a fiber — the timer wakes `self.current`, so a
+        // null current would arm a `fiber = null` timer that segfaults `wake` when
+        // it fires. Bail loudly, exactly like `sleep`/`suspend_self`/`block_on_fd`.
+        if self.current == null {
+            print("sched: arm_timer() called outside a fiber (no running fiber to wake)\n");
+            abort();
+        }
         t : Timer = .{ deadline_ms = deadline_ms, fiber = self.current };
         self.timers.append(t, self.own_allocator);
         return null;
@@ -753,11 +786,18 @@ T
 // regardless of linkage), never by an external name — so it needs the convention,
 // not a public symbol. `abi(.c)` states that intent and keeps the symbol private.
 //
-// One consequence of the C-ABI boundary: an `abi(.c)` fn has no implicit
-// `context` param, so `self.body()` runs under the static `__sx_default_context`
-// — NOT whatever `push Context { allocator = ... }` was in force at the
-// `run()` call site. Fiber bodies do not inherit a caller-scoped allocator; a
-// body that needs one must capture it explicitly (the long-lived-container rule).
+// The C-ABI boundary means this frame has no implicit `context` param, so the
+// body would otherwise run under the static `__sx_default_context`. To make a
+// fiber inherit its SPAWNER's context (so `context.io` reaches a fiber-backed
+// scheduler, and `context.allocator`/`context.data` carry through), `spawn`
+// snapshots the live `context` into `Fiber.dctx` and the body below re-pushes it.
+// LIFETIME CONTRACT (consequence of the snapshot): every capability in the
+// spawn-time context — `allocator`, `io`, `data` — must OUTLIVE the fiber, since
+// the body dereferences them later through `dctx`. Protocol values (`io`,
+// `allocator`) borrow a receiver (often a stack local via `xx`); that receiver,
+// and any `data` pointee, must stay alive until the fiber is reaped. A fiber that
+// grows a long-lived container through the inherited `context.allocator` is
+// subject to the usual long-lived-container rule (CLAUDE.md).
 fib_dispatch :: (self: *Fiber) abi(.c) {
     // Run the body under the context captured at spawn (`Fiber.dctx`), so the
     // fiber inherits its spawner's `context` — notably `context.io` resolves to a