diff --git a/cranelift/runtime/src/platform.rs b/cranelift/runtime/src/platform.rs
index a79b0897..bf45ebd7 100644
--- a/cranelift/runtime/src/platform.rs
+++ b/cranelift/runtime/src/platform.rs
@@ -68,6 +68,9 @@ pub unsafe extern "C" fn pith_input() -> *mut i8 {
 
 /// Execute a command and return exit code
 ///
+/// the child runs to completion via the process pool (see `process`), so a
+/// green worker is not held for however long the command takes.
+///
 /// # Safety
 /// command must be a valid null-terminated C string
 #[no_mangle]
@@ -87,24 +90,34 @@ pub unsafe extern "C" fn pith_exec(command: *const i8) -> i64 {
         cmd.args(&parts[1..]);
     }
 
-    if let Ok(status) = cmd.status() {
-        if let Some(code) = status.code() {
-            return code as i64;
-        }
-        return 0;
+    match crate::process::command_status(cmd) {
+        Some(status) => status.code().unwrap_or(0) as i64,
+        None => -1,
     }
-    -1
+}
+
+/// the multiplier/increment of the LCG behind `pith_random_float` (knuth's
+/// MMIX constants).
+const LCG_MUL: u64 = 6364136223846793005;
+const LCG_INC: u64 = 1;
+
+fn lcg_step(s: u64) -> u64 {
+    s.wrapping_mul(LCG_MUL).wrapping_add(LCG_INC)
 }
 
 /// Random float between 0.0 and 1.0
 #[no_mangle]
 pub extern "C" fn pith_random_float() -> f64 {
-    use std::num::Wrapping;
-
-    let s = RANDOM_SEED.load(Ordering::Relaxed);
-    let new_s = Wrapping(s) * Wrapping(6364136223846793005) + Wrapping(1);
-    RANDOM_SEED.store(new_s.0, Ordering::Relaxed);
-    (new_s.0 >> 11) as f64 / (1u64 << 53) as f64
+    // step the seed with a compare-exchange rather than a load/store pair: two
+    // tasks on different workers drawing at once must each claim a distinct
+    // step of the sequence, not both advance from the same seed and return the
+    // identical "random" value. relaxed ordering is enough — the value itself
+    // is the only payload, nothing else is published through it.
+    let prev = RANDOM_SEED
+        .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |s| Some(lcg_step(s)))
+        .unwrap_or(0); // unreachable: the closure never returns None
+    let new_s = lcg_step(prev);
+    (new_s >> 11) as f64 / (1u64 << 53) as f64
 }
 
 /// Seed the random number generator
@@ -152,6 +165,50 @@ pub unsafe extern "C" fn pith_random_string(len: i64) -> *mut i8 {
 mod tests {
     use super::*;
 
+    // both random tests reseed the one process-global sequence, so they must
+    // not interleave with each other (cargo runs tests concurrently).
+    static RANDOM_TEST_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
+
+    // seeding pins the whole sequence: same seed, same draws. this is the
+    // contract the CAS step must preserve for single-threaded callers.
+    #[test]
+    fn seeded_draws_are_reproducible() {
+        let _guard = RANDOM_TEST_LOCK.lock().unwrap_or_else(|p| p.into_inner());
+        pith_random_seed(42);
+        let first: Vec<u64> = (0..8).map(|_| pith_random_float().to_bits()).collect();
+        pith_random_seed(42);
+        let second: Vec<u64> = (0..8).map(|_| pith_random_float().to_bits()).collect();
+        assert_eq!(first, second);
+    }
+
+    // the race this guards against: with a load/store seed update, two threads
+    // can both step from the same seed and return the identical draw. the
+    // compare-exchange gives every draw its own step of the sequence, so
+    // concurrent draws are distinct (up to the astronomically unlikely 53-bit
+    // output collision between different seeds).
+    #[test]
+    fn concurrent_draws_do_not_repeat() {
+        let _guard = RANDOM_TEST_LOCK.lock().unwrap_or_else(|p| p.into_inner());
+        pith_random_seed(7);
+        let threads: Vec<_> = (0..4)
+            .map(|_| {
+                std::thread::spawn(|| {
+                    (0..2000)
+                        .map(|_| pith_random_float().to_bits())
+                        .collect::<Vec<u64>>()
+                })
+            })
+            .collect();
+        let mut all: Vec<u64> = threads
+            .into_iter()
+            .flat_map(|t| t.join().expect("drawing thread panicked"))
+            .collect();
+        let total = all.len();
+        all.sort_unstable();
+        all.dedup();
+        assert_eq!(all.len(), total, "two concurrent draws returned the same value");
+    }
+
     #[test]
     fn exec_rejects_null_and_invalid_utf8() {
         let invalid = [0xffu8, 0x00];
diff --git a/cranelift/runtime/src/process.rs b/cranelift/runtime/src/process.rs
index 06ea73bf..e174c6ce 100644
--- a/cranelift/runtime/src/process.rs
+++ b/cranelift/runtime/src/process.rs
@@ -10,14 +10,17 @@
 //! worth waiting for is pollable. the pipes are pipes, and linux will hand out
 //! a `pidfd` that becomes readable exactly when the child exits. so both waits
 //! go to the epoll reactor that already serves sockets (`netpoll`), and no
-//! thread is parked on a child anywhere in the runtime.
+//! green worker is parked on a spawned child anywhere in the runtime.
 //!
-//! that also settles the question of what may run off the task's own thread:
-//! nothing does. the reactor never runs pith code — it only marks a task
-//! runnable, and the task resumes on the worker it was pinned to — so every
-//! handle, `Bytes`, and C string here is built where it always was.
-
-use crate::blocking;
+//! the run-to-completion calls (`process.output`, `exec`) cannot use the
+//! reactor: `Command::output` and `Command::status` spawn, drain, and wait in
+//! one opaque std call. those go to a small thread pool instead, the same
+//! shape dns and file i/o use — see `command_output`/`command_status` and the
+//! rule in `blocking` about what may cross to a pool thread. the `Command`
+//! going in and the `Output` coming back are plain owned data; every handle,
+//! `Bytes`, and C string is still built on the calling task's thread.
+
+use crate::blocking::{self, Pool};
 use crate::collections::list::PithList;
 use crate::concurrency::scheduler::{backend, Backend};
 use crate::fdio;
@@ -31,6 +34,25 @@ use std::process::{Child, ChildStderr, ChildStdin, ChildStdout, Command, Stdio};
 use std::sync::atomic::{AtomicI64, Ordering};
 use std::sync::OnceLock;
 
+/// the threads that run whole-child commands (`Command::output`/`status`) for
+/// green tasks. a child decides for itself when it exits, so one of these can
+/// be held for an unbounded time — its own small pool keeps that from queueing
+/// behind (or in front of) file i/o or dns.
+static POOL: Pool = Pool::new("pith-proc", 4);
+
+/// run `command.output()` — spawn, drain both pipes, wait — without holding a
+/// green worker for the child's lifetime. off a green task it runs inline,
+/// exactly as the callers always did. `None` is any spawn/read failure.
+pub(crate) fn command_output(mut command: Command) -> Option<std::process::Output> {
+    blocking::run(&POOL, move || command.output().ok())
+}
+
+/// run `command.status()` — spawn and wait, stdio inherited — with the same
+/// offload-or-inline split as `command_output`.
+pub(crate) fn command_status(mut command: Command) -> Option<std::process::ExitStatus> {
+    blocking::run(&POOL, move || command.status().ok())
+}
+
 /// which of a child's three pipes a caller wants.
 #[derive(Clone, Copy, PartialEq, Eq)]
 pub(crate) enum Pipe {
@@ -232,18 +254,20 @@ pub unsafe extern "C" fn pith_process_output_argv(
     env_keys: PithList,
     env_values: PithList,
 ) -> i64 {
-    let Some(mut command) = pith_build_command(program, argv, cwd, env_keys, env_values) else {
+    let Some(command) = pith_build_command(program, argv, cwd, env_keys, env_values) else {
         return 0;
     };
 
-    match command.output() {
-        Ok(output) => {
+    // the child runs to completion on a pool thread (or inline, off green); the
+    // pith-visible handle is built here either way.
+    match command_output(command) {
+        Some(output) => {
             let status = output.status.code().unwrap_or(-1) as i64;
             let stdout = String::from_utf8_lossy(&output.stdout).to_string();
             let stderr = String::from_utf8_lossy(&output.stderr).to_string();
             pith_store_process_output(status, stdout, stderr)
         }
-        Err(_) => 0,
+        None => 0,
     }
 }
 
@@ -439,6 +463,26 @@ pub extern "C" fn pith_process_close(handle: i64) {
 mod tests {
     use super::*;
 
+    // unit tests never run on a green worker, so these exercise the inline arm
+    // of the offload split; the pool arm is the same closure on another thread.
+    #[test]
+    fn command_output_captures_both_pipes_and_the_exit_code() {
+        let mut cmd = Command::new("/bin/sh");
+        cmd.args(["-c", "echo out; echo err >&2; exit 3"]);
+        let output = command_output(cmd).expect("sh should run");
+        assert_eq!(output.status.code(), Some(3));
+        assert_eq!(output.stdout, b"out\n");
+        assert_eq!(output.stderr, b"err\n");
+    }
+
+    #[test]
+    fn command_status_reports_the_exit_code() {
+        let mut cmd = Command::new("/bin/sh");
+        cmd.args(["-c", "exit 5"]);
+        assert_eq!(command_status(cmd).expect("sh should run").code(), Some(5));
+        assert!(command_status(Command::new("/nonexistent-program")).is_none());
+    }
+
     #[test]
     fn invalid_process_handles_return_safe_defaults() {
         assert_eq!(pith_process_wait(12345), -1);
diff --git a/cranelift/runtime/src/runtime_core.rs b/cranelift/runtime/src/runtime_core.rs
index 415c66c9..e0b37fad 100644
--- a/cranelift/runtime/src/runtime_core.rs
+++ b/cranelift/runtime/src/runtime_core.rs
@@ -621,6 +621,26 @@ mod tests {
         }
     }
 
+    // structs travel between tasks, and tasks run on different worker threads,
+    // so the last strong owner can die on a thread the weak reader never runs
+    // on. the dead flag is an atomic for exactly this shape; the join gives the
+    // load a happens-after edge, so it must observe the death.
+    #[test]
+    fn weak_load_sees_a_release_from_another_thread() {
+        unsafe {
+            let s = pith_struct_alloc(1);
+            pith_struct_weak_retain(s);
+            assert_eq!(pith_struct_weak_load(s), s);
+
+            std::thread::spawn(move || pith_struct_release(s))
+                .join()
+                .expect("releasing thread panicked");
+
+            assert_eq!(pith_struct_weak_load(s), 0);
+            pith_struct_weak_release(s);
+        }
+    }
+
     #[test]
     fn weak_load_on_a_dead_or_invalid_pointer_is_none() {
         unsafe {
@@ -1689,6 +1709,15 @@ unsafe fn struct_weak(base: *mut u8) -> &'static std::sync::atomic::AtomicU32 {
     &*(base.add(STRUCT_OFF_WEAK) as *const std::sync::atomic::AtomicU32)
 }
 
+// the dead flag must be an atomic like its neighbors: structs cross task (and
+// therefore worker-thread) boundaries, so the last strong owner can set it on
+// one thread while a weak reference reads it on another. the release store
+// pairs with the acquire load in `pith_struct_weak_load`, so a reader that
+// observes the value as dead also observes everything the destructor did.
+unsafe fn struct_dead(base: *mut u8) -> &'static std::sync::atomic::AtomicU32 {
+    &*(base.add(STRUCT_OFF_DEAD) as *const std::sync::atomic::AtomicU32)
+}
+
 // drop one unit of weak; free the allocation when it reaches zero. this is the
 // single place a struct header is deallocated, so the strong path and every
 // weak reference funnel their final drop through here.
@@ -1787,9 +1816,11 @@ pub unsafe extern "C" fn pith_struct_release(ptr: i64) {
         f(ptr);
     }
     crate::perf_count(&crate::PERF_STRUCT_FREES, 1);
-    // the value is now dead; a weak read after this returns none. drop the
-    // implicit weak unit, which frees the header if no weak refs remain.
-    (base.add(STRUCT_OFF_DEAD) as *mut u32).write(1);
+    // the value is now dead; a weak read after this returns none. the release
+    // store publishes the destructor's writes to any thread whose weak load
+    // sees the flag. then drop the implicit weak unit, which frees the header
+    // if no weak refs remain.
+    struct_dead(base).store(1, std::sync::atomic::Ordering::Release);
     struct_weak_drop(base);
 }
 
@@ -1828,7 +1859,9 @@ pub unsafe extern "C" fn pith_struct_weak_load(ptr: i64) -> i64 {
     let Some(base) = struct_base(ptr) else {
         return 0;
     };
-    if (base.add(STRUCT_OFF_DEAD) as *const u32).read() != 0 {
+    // acquire pairs with the release store in `pith_struct_release`: a load
+    // that sees the target dead also sees its destructor's effects.
+    if struct_dead(base).load(std::sync::atomic::Ordering::Acquire) != 0 {
         return 0;
     }
     ptr
diff --git a/cranelift/runtime/src/utility.rs b/cranelift/runtime/src/utility.rs
index ecd57365..42405893 100644
--- a/cranelift/runtime/src/utility.rs
+++ b/cranelift/runtime/src/utility.rs
@@ -45,6 +45,9 @@ pub unsafe extern "C" fn pith_log_error(msg: *const i8) {
 }
 
 /// Execute command and capture output — returns stdout as C string
+///
+/// the child runs to completion via the process pool (see `process`), so a
+/// green worker is not held for however long the command takes.
 #[no_mangle]
 pub unsafe extern "C" fn pith_exec_output(cmd: *const i8) -> *mut i8 {
     let Some(cmd_str) = cstr_str(cmd) else {
@@ -52,15 +55,13 @@ pub unsafe extern "C" fn pith_exec_output(cmd: *const i8) -> *mut i8 {
     };
     let parts: Vec<&str> = cmd_str.split_whitespace().collect();
     if parts.is_empty() {
-        std::ptr::null_mut()
-    } else {
-        match std::process::Command::new(parts[0])
-            .args(&parts[1..])
-            .output()
-        {
-            Ok(output) => crate::pith_copy_bytes_to_cstring(&output.stdout),
-            Err(_) => std::ptr::null_mut(),
-        }
+        return std::ptr::null_mut();
+    }
+    let mut command = std::process::Command::new(parts[0]);
+    command.args(&parts[1..]);
+    match crate::process::command_output(command) {
+        Some(output) => crate::pith_copy_bytes_to_cstring(&output.stdout),
+        None => std::ptr::null_mut(),
     }
 }
 
diff --git a/docs/concurrency.md b/docs/concurrency.md
index 9dd50e26..58fc6a46 100644
--- a/docs/concurrency.md
+++ b/docs/concurrency.md
@@ -216,13 +216,15 @@ matters more here than anywhere else: a wait has no bound at all, and four
 `sleep 3600`s would have emptied a four-thread pool. a child that has already
 finished by the time you wait for it is collected on the spot, with no park.
 
-`process.output` is the exception, along with everything routed through it
-(`run`, `text`, `output_checked`) and the two shell helpers `run_shell` and
-`output_shell`. each of those runs a child to completion inside one call while
-draining both of its pipes, and draining two pipes at once without blocking
-needs a reactor wait covering more than one fd, which is not built yet. until
-it is, `start` plus your own reads and `wait` yields properly; `output_ctx` is
-that pattern already written out.
+`process.output` — along with everything routed through it (`run`, `text`,
+`output_checked`) and the two shell helpers `run_shell` and `output_shell` —
+runs a child to completion inside one call while draining both of its pipes,
+which the reactor cannot cover (it waits on one fd at a time). those calls go
+to a small process pool of their own instead, the dns/file shape: the command
+crosses to a blocking thread, the task parks, and the worker stays free for
+however long the child runs. `exec` and `exec_output` take the same route.
+`start` plus your own reads and `wait` still parks on the reactor directly,
+with no pool thread involved.
 
 a task waits a few microseconds on-CPU before it actually parks. a read the page
 cache answers comes back faster than the two thread wakeups a park costs, so
diff --git a/docs/limitations.md b/docs/limitations.md
index 6fa42b72..8d53e5ab 100644
--- a/docs/limitations.md
+++ b/docs/limitations.md
@@ -107,6 +107,24 @@ correctness story:
   cross-module map reads, set codegen, negative float literals like `-1.0`) were
   re-checked and all pass; they are now pinned by regression tests
   (`tests/cases/test_xmod_float.pith` and friends).
+- `os.set_env` after tasks have started is a libc-level race. glibc's `setenv`
+  and `getenv` are not synchronized against each other, and the runtime's own
+  pool threads read the environment behind your back — `getaddrinfo` on the
+  dns pool consults `RES_OPTIONS` and friends on every lookup. rust's internal
+  env lock only covers rust-side accesses, so a `set_env` concurrent with a
+  dial can crash in libc. set environment variables before spawning tasks. the
+  candidate fixes (snapshot the environment at startup, or make late `set_env`
+  write to an overlay that child processes inherit) both change observable
+  semantics, so this is documented rather than decided for now.
+- closing an fd-backed handle races a concurrent call on the same handle. a
+  task parked on a socket or pipe that another task closes is woken with an
+  error (the reactor's close teardown), which is the common case and safe.
+  what remains is the standard raw-fd hazard: between one task's handle
+  lookup and its syscall, a close plus a new open can recycle the fd number,
+  and the syscall lands on the wrong fd. present on both backends and
+  inherited from fd semantics; closing it needs handles that carry liveness
+  (a generation, like task handles have) rather than raw fd numbers. do not
+  share a connection between tasks without coordinating its close.
 
 ## the green backend, now the default on linux
 
@@ -134,14 +152,22 @@ thread would. on a slow network mount that reasoning does not hold and a task
 doing one of them holds its worker until it returns. making that decision
 adaptive rather than fixed is the open work.
 
-the other is `process.output` and the calls built on it (`run`, `text`,
-`output_checked`, `run_shell`, `output_shell`). each runs a child to completion
-inside one runtime call and holds the worker for as long as the child lives.
-draining a child's stdout and stderr at the same time without blocking needs a
-reactor wait that covers more than one fd, and the reactor waits on one; adding
-that is the fix. `start` plus explicit reads and `wait` yields properly today,
-and `output_ctx` is that pattern already written out. `PITH_GREEN=0` is the
-blunt workaround for either.
+the other is `sleep`, and everything built on it. `time.delay` maps straight
+to the blocking sleep, so a green task that sleeps holds its worker for the
+duration — and the waiting loops built on delay hold one too: a `select` with
+no ready arm sleeps a millisecond per probe until an arm fires, and the
+`concurrent.after`/`ticker` workers and a context's deadline watcher do the
+same. each slice is short and correctness is unaffected, but on a two-worker
+host two idle selects can occupy both workers while they poll. the fix is a
+real timer in the reactor — the deadline heap `netpoll` already keeps could
+serve a pure-timeout wait with no fd — so a sleeping task parks the way a
+socket read does. `PITH_GREEN=0` is the blunt workaround.
+
+(`process.output` and the calls built on it — `run`, `text`, `output_checked`,
+`run_shell`, `output_shell`, plus `exec` and `exec_output` — used to be on
+this list for holding a worker while a child ran to completion; they now hand
+the whole spawn-drain-wait to a process pool of their own and the caller
+parks, the same shape dns and file i/o use.)
 
 the calling task pays for that. a file call made from inside a task now costs a
 thread handoff it did not before, so a task reading a small cached file in a