diff --git a/build.zig b/build.zig
index b6e6dea..ae18855 100644
--- a/build.zig
+++ b/build.zig
@@ -926,6 +926,22 @@ pub fn build(b: *std.Build) void {
     });
     test_step.dependOn(&b.addRunArtifact(time_tests).step);
 
+    // runtime.Thread's lock/condvar state machines (Mutex/Condition/RwLock/WaitGroup). Its
+    // Futex seam falls back to std.Thread.Futex off the danos target, so the tests exercise
+    // them with real host threads (docs/threading-plan.md M11). Like time.zig it pulls in
+    // system.zig (syscall wrappers), which needs the `abi` module.
+    const thread_tests = b.addTest(.{
+        .root_module = b.createModule(.{
+            .root_source_file = b.path("library/runtime/thread.zig"),
+            .target = target,
+            .optimize = optimize,
+            .imports = &.{
+                .{ .name = "abi", .module = abi_module },
+            },
+        }),
+    });
+    test_step.dependOn(&b.addRunArtifact(thread_tests).step);
+
     // Convenience: `zig build gen-xkeyboard-config` regenerates the layout tables from the
     // vendored data (offline). `fetch` (the network step) stays a manual script run.
     const gen_xkb = b.addSystemCommand(&.{ "python3", "tools/make-xkeyboard-config.py", "generate" });
diff --git a/docs/README.md b/docs/README.md
index 5778206..96663c8 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -107,7 +107,7 @@ Start with the north star:
 - **[threading.md](threading.md) — threads, the std-shaped way.** **Built** (M1–M6):
   `runtime.Thread` mirrors `std.Thread`'s API (spawn/join/detach, Mutex/Condition/
   Semaphore) over a **private** thread ABI — several tasks sharing one address space via
-  a `thread_spawn` syscall, futex-backed blocking, aspace refcounting. Why it's the
+  a `thread_spawn` syscall, futex-backed blocking, address-space refcounting. Why it's the
   native type and not literal `std.Thread` (the [private ABI](syscall.md)), and why
   threads stay a narrow opt-in against the [resilience](resilience.md) default. Build
   plan + gates: [threading-plan.md](threading-plan.md).
diff --git a/docs/display-v2-plan.md b/docs/display-v2-plan.md
index f08d75d..ca97ce4 100644
--- a/docs/display-v2-plan.md
+++ b/docs/display-v2-plan.md
@@ -62,7 +62,7 @@ is the only backend), and `zig build test` stays green.
 - [x] [abi.zig](../system/abi.zig): `shm_create` (34) / `shm_map` (35) syscalls + a
       `shm_test` service id. Handlers in process.zig: `shm_create(len)` allocates contiguous,
       zeroed, **cacheable** frames, wraps them in a refcounted object, installs a capability
-      handle, maps them into the caller's shm arena → returns vaddr + handle; `shm_map(cap)`
+      handle, maps them into the caller's shm arena → returns virtual_address + handle; `shm_map(cap)`
       maps the same physical pages into the receiver. Reclaimed on death (see below).
 - [x] The capability core (ipc-synchronous.zig) is now **kind-tagged**: `scheduler.Task`'s
       handle table holds `HandleObject{kind, ptr}`; `closeHandles` and `shareCapability`
diff --git a/docs/display-v2.md b/docs/display-v2.md
index 42dedad..ada8c3e 100644
--- a/docs/display-v2.md
+++ b/docs/display-v2.md
@@ -77,9 +77,9 @@ deferred (docs/display.md, "What v1 does not do"). v2 builds it: the natural gen
 of M13 capability-passing from *endpoints* to *memory objects* —
 
 ```
-shm_create(len) -> {handle, vaddr}     // a shareable, page-aligned RAM region
+shm_create(len) -> {handle, virtual_address}     // a shareable, page-aligned RAM region
 … pass `handle` as the send_cap on an ipc_call …
-shm_map(cap)    -> vaddr                // the receiver maps the same physical pages
+shm_map(cap)    -> virtual_address                // the receiver maps the same physical pages
 ```
 
 The payoff is leverage: the **same** primitive unlocks **both** native GPU drivers *and*
diff --git a/docs/display.md b/docs/display.md
index 3f4fd1b..b4c6dad 100644
--- a/docs/display.md
+++ b/docs/display.md
@@ -220,8 +220,8 @@ both are clean additions behind the interfaces v1 establishes.
   to render into its *own* buffer and hand the compositor a *reference*, not a stream of
   commands. That needs the missing cross-process shared-memory primitive — best built as
   the natural generalization of the existing M13 [capability passing](driver-model.md)
-  from *endpoints* to *memory objects* (`shm_create(len) → {cap, vaddr}`, pass `cap` on
-  an `ipc_call`, receiver `shm_map(cap) → vaddr`). v1 avoids it because server-owned
+  from *endpoints* to *memory objects* (`shm_create(len) → {cap, virtual_address}`, pass `cap` on
+  an `ipc_call`, receiver `shm_map(cap) → virtual_address`). v1 avoids it because server-owned
   surfaces already prove the whole pipeline.
 
 - **Runtime mode-setting (a native backend).** Detecting the EDID mode list and changing
diff --git a/docs/driver-model.md b/docs/driver-model.md
index 9016e1b..23e1946 100644
--- a/docs/driver-model.md
+++ b/docs/driver-model.md
@@ -240,8 +240,8 @@ once per page, maps writeback-cached, and never reveals a physical address.
 **The fix.**
 
 ```
-dma_alloc(len, flags) -> vaddr (rax), paddr (rdx)
-dma_free(vaddr, len)  -> 0
+dma_alloc(len, flags) -> virtual_address (rax), physical_address (rdx)
+dma_free(virtual_address, len)  -> 0
 
 flags:  dma_coherent  (1)  uncacheable; the default and the only one that's portable
         dma_wc        (2)  write-combining — needs PAT programmed; for framebuffers
diff --git a/docs/drivers.md b/docs/drivers.md
index 74ea9bd..44749d2 100644
--- a/docs/drivers.md
+++ b/docs/drivers.md
@@ -74,7 +74,7 @@ The driver syscall numbers (`system/abi.zig`) with the device types they carry
 |---|------|---------|
 | 11 | `device_enumerate(buf, max) -> total` | Snapshot the device table |
 | 12 | `device_claim(id) -> ok` | Take **exclusive** ownership |
-| 13 | `mmio_map(id, res_idx) -> vaddr` | Map a claimed device's register window |
+| 13 | `mmio_map(id, res_idx) -> virtual_address` | Map a claimed device's register window |
 | 14 | `irq_bind(id, res_idx, endpoint)` | Deliver that device's IRQ as a notification |
 | 15 | `irq_ack(id, res_idx)` | Re-arm the IRQ after servicing the device |
 | 16 | `device_register(parent_id, desc) -> id` | Publish a child of a device you claimed |
diff --git a/docs/threading-plan.md b/docs/threading-plan.md
index 7b06344..3f245c8 100644
--- a/docs/threading-plan.md
+++ b/docs/threading-plan.md
@@ -105,28 +105,28 @@ first unchecked box.
 ## M1 — Address-space refcount (kernel foundation, no API, no behaviour change) ✅
 
 The one invariant change threads require, landed and proven **before** anything shares
-an address space. Today aspace is 1:1 with a task and teardown destroys it on any user
+an address space. Today address space is 1:1 with a task and teardown destroys it on any user
 task's exit; make destruction happen on the **last** exit.
 
 - [x] A refcount keyed by the address-space root, held in `scheduler.zig`
-      (`aspace_refs`): `retainAspace` takes a reference in `spawnUserLocked` (on the
+      (`address_space_refs`): `retainAddressSpace` takes a reference in `spawnUserLocked` (on the
       success path, after the slot + stack are secured), all under the big kernel lock.
 - [x] Both task-teardown paths ([scheduler.zig](../system/kernel/scheduler.zig):
       `exitUserLocked` and `destroyTaskLocked`) call `releaseAspace`, which decrements
       and only `destroyAddressSpace`s at **zero**; an unretained space (hand-built test
       spaces) is destroyed directly, preserving prior behaviour.
-- [x] `-Dtest-case=aspace-refcount`: spawn and reap several ring-3 processes in sequence
-      and assert (via test-observable `liveAspaceCount`/`aspaceDestroyCount`) that the
+- [x] `-Dtest-case=address-space-refcount`: spawn and reap several ring-3 processes in sequence
+      and assert (via test-observable `liveAddressSpaceCount`/`addressSpaceDestroyCount`) that the
       live-space count returns to **baseline** and destructions advance by exactly that
       many — each space destroyed exactly once, no leak, no double-free. (Refcount
       observables, not raw frame counts, since kernel stacks are still leaked on exit.)
 
-**Gate (met):** `python3 test/qemu_test.py aspace-refcount` passes
-(`aspace-refcount: spaces released to baseline ok` → `DANOS-TEST-RESULT: PASS`), and the
+**Gate (met):** `python3 test/qemu_test.py address-space-refcount` passes
+(`address-space-refcount: spaces released to baseline ok` → `DANOS-TEST-RESULT: PASS`), and the
 full guardrail set passes unchanged — 13/13 (`smoke`, `sched`, `priority`, `smp`,
 `affinity`, `process`, `process-kill`, `supervision`, `fault-recovery`,
 `vfs-client-death`, `ipc`, `ipc-cap`, `display-service`); default `zig build` clean,
-`zig build test` green. The reframing is invisible until an aspace is actually shared.
+`zig build test` green. The reframing is invisible until an address space is actually shared.
 
 ## M2 — `thread_spawn` + `thread_exit`: a thread runs in the shared address space ✅
 
@@ -135,7 +135,7 @@ space and exits cleanly.
 
 - [x] [abi.zig](../system/abi.zig): `thread_spawn = 37`, `thread_exit = 38`. Handlers in
       process.zig; `thread_spawn` calls `scheduler.spawnThread` (shares the caller's
-      aspace, `retainAspace`); `thread_exit` ends the task like a process `exit(0)`
+      address space, `retainAddressSpace`); `thread_exit` ends the task like a process `exit(0)`
       (`terminateCurrent` → `releaseAspace`). The closure pointer is delivered in the new
       thread's **rdi** via a new `jump_to_user_arg` asm path (`t.user_arg`, 0 for a
       process) — no naked runtime asm.
@@ -152,14 +152,14 @@ space and exits cleanly.
       address space.
 
 **Gate (met):** `python3 test/qemu_test.py thread-spawn` passes
-(`thread-test: child ran in shared aspace ok` → `DANOS-TEST-RESULT: PASS`); guardrail set
+(`thread-test: child ran in shared address space ok` → `DANOS-TEST-RESULT: PASS`); guardrail set
 16/16 green (incl. `args`/`init`/`process`, which exercise the new `jump_to_user_arg`
-process path with arg 0) plus `aspace-refcount`; `zig build` clean, `zig build test`
+process path with arg 0) plus `address-space-refcount`; `zig build` clean, `zig build test`
 green.
 
 > **Note (deferred to M3+):** the mmap arena is per-*task* (`heap_next`), so two threads
-> in one aspace that both `mmap` would collide. Fine for M2 (only the parent maps, for the
-> child's stack); make the arena per-aspace and the runtime heap thread-safe alongside the
+> in one address space that both `mmap` would collide. Fine for M2 (only the parent maps, for the
+> child's stack); make the arena per-address-space and the runtime heap thread-safe alongside the
 > `Mutex` work (M5).
 
 ## M3 — `join` + `detach` + real parallelism ✅
@@ -184,7 +184,7 @@ green.
 **Gate (met):** `python3 test/qemu_test.py thread-join` passes (`thread-test: join ok` →
 `DANOS-TEST-RESULT: PASS`), robust across 4 runs; guardrail 17/17 green (incl. `smp`,
 `affinity`, `process-kill`, and `args`/`init`/`process` on the exit-endpoint spawn path)
-plus `aspace-refcount`/`thread-spawn`; `zig build` clean, `zig build test` green.
+plus `address-space-refcount`/`thread-spawn`; `zig build` clean, `zig build test` green.
 
 > **Note (deferred):** a detached thread's stack is freed only at process exit (not by the
 > reaper on thread exit) — kernel user-stack tracking + reclaim is a later refinement. And
@@ -212,7 +212,7 @@ plus `aspace-refcount`/`thread-spawn`; `zig build` clean, `zig build test` green
 **Gate (met):** `python3 test/qemu_test.py thread-futex` passes, robust across 3 runs —
 the case's **ordered** regex asserts `waiting → waking → woke → PASS` on the serial
 stream (the handoff proof), and `thread-futex: timeout ok` confirms the timeout.
-Guardrail 18/18 green (incl. `sleep`/`event`/`ipc` blocking paths) + `aspace-refcount`,
+Guardrail 18/18 green (incl. `sleep`/`event`/`ipc` blocking paths) + `address-space-refcount`,
 `thread-spawn`, `thread-join`; `zig build` clean, `zig build test` green.
 
 > **Note:** the kernel test checks only the freshest verdict marker via `bufferHas` (the
@@ -252,11 +252,11 @@ green.
 
 - [x] `getCurrentId` via a small `thread_self = 42` syscall (`runtime.Thread.getCurrentId`
       returns the kernel task id). **Per-thread `threadlocal` TLS is deferred** — no
-      consumer needs it, and it would require context-switching `fs.base` per task (real
-      kernel + per-switch cost) for an unused feature; threaded binaries have run fine
+      consumer needs it, and it would require context-switching the thread pointer per task
+      (real kernel + per-switch cost) for an unused feature; threaded binaries have run fine
       without it through M2–M5. threading.md's TLS reasoning already scoped it as
       deferred-unless-needed. When a consumer appears, the shape is: `thread_spawn`
-      allocates a per-thread TLS block, sets `fs.base`, and the context switch saves/
+      allocates a per-thread TLS block, sets the thread pointer, and the context switch saves/
       restores it.
 - [x] `RwLock` / `WaitGroup` deferred (no consumer yet); they slot onto the same
       `Futex`/`Mutex`/`Condition` when wanted.
@@ -279,8 +279,11 @@ green.
 cross-core parallelism, futex, and `Mutex`/`Condition`/`Semaphore`, all over a private
 thread ABI behind the runtime.
 
-**Phase 2 (M7–M11): planned below** — hardening the deferred parts so threads are safe
-for real workloads and reclaimed like everything else danos owns.
+**Phase 2 (M7–M11): built.** Thread-safe allocation (M7), a task reaper that reclaims dead
+tasks' kernel stacks (M8), endpoint-free `thread_join` (M9), the per-thread thread pointer (M10),
+and `RwLock`/`WaitGroup` + host-testable sync (M11). Two things stay deferred by design
+(no consumer): the Zig `threadlocal` *compiler* layer (M10) and detached-thread user-stack
+reclaim (M9) — both noted in place.
 
 ---
 
@@ -290,112 +293,170 @@ The organising principle, so Phase 2 reinforces danos's goals rather than erodin
 
 - **Everything a thread owns is reclaimed on process death.** Thread stacks, TLS blocks,
   and futex words live in the process's **address space**, and the kernel's per-process
-  state is keyed by the aspace root — so the M1 refcount + `destroyAddressSpace` already
+  state is keyed by the address-space root — so the M1 refcount + `destroyAddressSpace` already
   free all of it when the last thread exits. A crashed or killed threaded process leaves
-  **nothing** behind. Phase 2 closes the one thing that is *not* aspace-owned — the
+  **nothing** behind. Phase 2 closes the one thing that is *not* address-space-owned — the
   per-task **kernel** stack (kernel heap) — with a reaper (M8). This is the
   [resilience](resilience.md) restart guarantee, extended to threads.
 - **Kernel owns mechanism; the runtime owns policy.** The kernel maps pages, saves/
-  restores `fs.base`, and reaps dead tasks; the runtime decides allocation, TLS layout,
+  restores the thread pointer, and reaps dead tasks; the runtime decides allocation, TLS layout,
   and lock algorithms. Every new kernel entry stays a private syscall behind the runtime
   ([syscall.md](syscall.md)) — the ABI stays renumberable.
 - **The process is still the isolation and restart boundary.** Threads share fate within
   one process; Phase 2 never adds a way for one process to reach into another (the
   cross-process futex stays explicitly out of scope, below).
 
-### M7 — Thread-safe allocation (the correctness gap)
+### M7 — Thread-safe allocation (the correctness gap) ✅
 
 Today the mmap arena cursor is per-*task* and the runtime heap is unlocked, so two
 threads in one process that both allocate corrupt each other. The thread *machinery*
 avoids this (closure on the stack, stacks mmap'd only by the spawner), but real
-multi-threaded code would hit it. Close it:
+multi-threaded code would hit it. Closed it:
 
-- [ ] **Kernel — per-address-space mmap arena.** Grow M1's `aspace_refs` entry into a
-      small address-space object holding the `mmap`/`mmio` arena cursors (moved off
-      `Task`); `systemMmap`/`mmio_map` bump the *aspace's* cursor under the big lock, so
-      sibling threads get disjoint, serialized grants. Freed at refcount zero, so the
-      cursors vanish with the process.
-- [ ] **Runtime — thread-safe heap.** Guard the allocator with a `Thread.Mutex`, gated on
+- [x] **Kernel — per-address-space mmap arena.** Grew M1's `address_space_refs` entry into the
+      per-address-space object holding the `mmap`/`mmio` arena cursors (moved off `Task`);
+      `scheduler.addressSpaceMmapNextPtr`/`addressSpaceDeviceMapNextPtr` expose them. `systemMmap`
+      reserves a disjoint range under a *brief* lock, then maps **per page** under a
+      short-held lock — not the whole grant — because the big lock is held with interrupts
+      disabled, so pinning it across a multi-MiB memset+map froze other cores (it timed
+      the `affinity` scenario out mid-bring-up). Freed at refcount zero, so the cursors
+      vanish with the process.
+- [x] **Runtime — thread-safe heap.** The allocator's two free-list mutators
+      (`rawAlloc`/`rawFree`) take a `Thread.Mutex`, gated on
       `!@import("builtin").single_threaded` so single-threaded binaries compile it out and
-      pay nothing. (The heap grows via mmap, now safe per above.)
-- [ ] `-Dtest-case=thread-alloc` (`smp: 4`): N threads each do many `alloc`/`free` of
-      varied sizes, write a per-thread pattern, verify it, and free; assert every block
-      round-trips intact and all memory returns — no corruption under concurrent
-      allocation. A direct check confirms two threads' concurrent `mmap`s are disjoint.
+      pay nothing. Uncontended acquisition is a single CAS (no syscall).
+- [x] `-Dtest-case=thread-alloc` (`smp: 4`): 4 threads each do 500 `alloc`/fill/verify/
+      `free` cycles of varied sizes; each block is filled with a per-thread pattern and
+      verified before free, so any overlap between concurrent allocations is caught.
 
-**Gate:** `thread-alloc` passes; guardrail + all `thread-*` cases green.
+**Gate (met):** `thread-alloc` passes (3× non-flaky); full guardrail 23/23 green,
+`zig build`/`zig build test` clean.
 
-### M8 — The task reaper (cleanup + resilience)
+> **Also fixed here:** the `affinity` guardrail's fixed-count busy-loop (`while (spins <
+> 3e9)`) had codegen-dependent wall-time — adding a function to `tests.zig` flipped how
+> the optimiser compiled it, swinging affinity from ~4 s to ~63 s and timing it out.
+> Reworked it (and the settle loop) to wait on the wall clock instead, so its duration is
+> independent of unrelated code changes.
 
-A dead task's **kernel** stack is currently leaked ("no reaper yet") — every process
-*and* thread death loses one, so a crash loop bleeds kernel memory. A reaper fixes it and
-serves the [resilience](resilience.md) restart goal directly:
+### M8 — The task reaper (cleanup + resilience) ✅
 
-- [ ] A dying task cannot free the kernel stack it runs on, so it hands itself to a
-      **reap list** and switches away; the kernel stack (and, for a detached thread, its
-      user stack) is reclaimed from another context — a low-priority reaper step drained
-      on the scheduler tick and when a core goes idle. Extends the existing
-      `reap_task_hook`/`destroyTaskLocked` path rather than inventing a parallel one.
-- [ ] `-Dtest-case=task-reap`: spawn and exit many threads and processes; assert the
-      kernel-heap free bytes (a new test observable) return to **baseline** — kernel
-      stacks reclaimed, no leak — and that the `fault-recovery`/kill paths reclaim too.
+A dead task's **kernel** stack was leaked ("no reaper yet") — every process *and* thread
+death lost one, so a crash loop bled kernel memory. The reaper fixes it and serves the
+[resilience](resilience.md) restart goal directly:
 
-**Gate:** `task-reap` passes; `fault-recovery`, `supervision`, `process-kill`,
-`aspace-refcount` still green.
+- [x] A dying task cannot free the kernel stack it runs on, so `exit()`/`exitUserLocked`
+      record it in a **per-core `reap_after_switch` slot** and switch away; the task that
+      resumes on that core frees the stack in `switchTo`'s tail (it's on its own stack, the
+      big lock is still held so the slot can't have been reused). A **tick-time drain**
+      (`reapKillPendingLocked`) is the safety net for the case where the next task is
+      *fresh* (enters via the trampoline, bypassing `switchTo`'s tail). A task killed while
+      *not* running is freed immediately in `destroyTaskLocked`. A `live_stack_bytes`
+      counter is the observable. *(Detached-thread user-stack reclaim moves to M9, which
+      adds the joinable/detached flag.)*
+- [x] `-Dtest-case=task-reap` (`smp: 4`): spawn and kill 12 processes; poll the
+      test-observable `scheduler.liveStackBytes()` until it returns to **baseline** (a
+      correct reaper gets there in a few ms; a genuine leak times out) — every kernel
+      stack reclaimed, no leak. Threads exit through the same `exitUserLocked`, so covered.
+
+**Gate (met):** `task-reap` passes (5× isolated + 2× in the full batch); `fault-recovery`,
+`supervision`, `process-kill`, `address-space-refcount`, `smp`, `affinity` all still green (24/24
+full guardrail); `zig build`/`zig build test` clean.
+
+> **Bug found + fixed here (touches every context switch):** the post-`switchContext` reap
+> first read the `pc` **parameter**, but a task that migrated cores carries a *stale* `pc`
+> in its saved `switchTo` frame — so it read the wrong core's slot and freed a live stack
+> (a #GP under SMP). Fixed to re-fetch `thisCpu()` after the switch (the switch only swaps
+> stacks on the current core).
 
 ### M9 — Futex-completion join (retire the per-thread endpoint)
 
 With the reaper (M8) able to act *after* a thread is fully off its stack, migrate `join`
 to the std shape and drop M3's per-thread exit endpoint:
 
-- [ ] `thread_spawn` takes a user **completion word** (in the `Thread` handle's memory)
-      and a joinable/detached flag. On reap the kernel writes 0 to that word and
-      `futex_wake`s it (a `CLONE_CHILD_CLEARTID` equivalent — safe now the thread is off
-      its stack). `join` = `futex_wait` on the word, then `munmap` the stack; a
-      **detached** thread's stack is `munmap`ped by the reaper instead. No IPC endpoint
-      per thread.
-- [ ] `thread-join` passes on the new path; a check confirms joining N threads creates no
-      per-thread endpoints (handle count stable).
+- [x] A **`thread_join(tid)` syscall** (not a user futex word): it blocks the caller until
+      the task with id `tid` exits, and the exit paths call `wakeJoinersLocked`. `join`
+      only reclaims the joined thread's **user** stack, which the thread vacates the moment
+      it enters the kernel to exit — so waking at *exit* time (not reap time) is safe, and
+      no reaper/address-space juggling or user-memory write is needed. This is equally
+      std-shaped (like `pthread_join`) and much simpler/safer than the planned
+      reaper-written completion word. `thread_spawn` no longer takes an exit endpoint (the
+      runtime passes `no_cap`); the per-thread IPC endpoint is gone.
+- [x] `thread-join` passes on the new path, and its join mode now runs **40 spawn+join
+      cycles** — under the old per-thread-endpoint scheme those leaked handles would
+      exhaust the 16-slot handle table; here they all succeed, proving join is endpoint-free.
 
-**Gate:** `thread-join`/`thread-mutex` green on futex-completion join; guardrail green.
+**Gate (met):** `thread-join` passes (3× isolated) on the `thread_join` path; full
+guardrail 26/26 (incl. `process-kill`, `supervision`, `fault-recovery`, `task-reap`);
+`zig build`/`zig build test` clean.
 
-### M10 — Per-thread TLS (`threadlocal`)
+> **Reaper hardened here (fixes an M8 flake).** M8's single per-core reap slot could be
+> *overwritten* by a second death on that core before the first drained (a fresh-task/SMP
+> timing window) — an intermittent one-stack leak (`task-reap` flaked ~20%). Replaced it
+> with a per-core reap **list** plus a `.reaping` task state so a pending slot can't be
+> reused before its stack is freed. `task-reap` now 11/11 isolated + 2× in the batch.
 
-Give each thread its own `threadlocal` storage — the piece self-hosting Zig
-([zig-self-hosting.md](zig-self-hosting.md)) will force:
+> **Deferred:** detached-thread **user-stack** reclaim (still freed at process exit, as in
+> M3). Doing it in the reaper needs the saved address space + stack range and a
+> translate/unmap in a not-currently-loaded address space — real complexity for a bounded leak.
+> A follow-up when a consumer needs it.
 
-- [ ] **Runtime** allocates a per-thread TLS block from the binary's `PT_TLS` template
-      (linker symbols: copy `.tdata`, zero `.tbss`, variant-II TCB self-pointer) and hands
-      its thread pointer to `thread_spawn`; the main thread sets its own via a new
-      `set_thread_pointer` syscall in `_start`. The block is aspace memory → reclaimed on
-      teardown.
-- [ ] **Kernel** stores `fs_base` on `Task`, loads it at first entry and restores it on
-      context switch only when it changes (the same conditional-load pattern as CR3).
-      `getCurrentId` can then read a TLS self-slot instead of a syscall.
-- [ ] `-Dtest-case=thread-tls`: two threads each write and read their own `threadlocal`
-      slot with no cross-talk, and observe distinct `getCurrentId`.
+### M10 — Per-thread TLS: the thread-pointer mechanism ✅
+
+Give each thread its own thread pointer and private TLS storage — the foundation
+self-hosting Zig ([zig-self-hosting.md](zig-self-hosting.md)) will build `threadlocal` on.
+
+- [x] **Kernel** stores `thread_pointer` on `Task` and restores it on every context switch
+      **only when it changes** (the same conditional-load discipline as CR3;
+      `architecture.setThreadPointer` → `wrmsr IA32_FS_BASE` on x86_64). A
+      `set_thread_pointer(addr)` = 44 syscall sets the caller's `thread_pointer` and loads it
+      now. The kernel never touches FS, so there is no swapgs complication.
+- [x] **Runtime** lays a small per-thread TLS block at the top of each thread's stack
+      (self-pointer at `%fs:0` + scratch slots) and the thread trampoline calls
+      `set_thread_pointer` before any user code — so every spawned thread has a private,
+      switch-stable thread pointer. Reclaimed with the stack.
+- [x] `-Dtest-case=thread-tls` (`smp: 4`): two threads each write a unique marker to their
+      own `%fs:8` slot and — after both have written — read it back; a shared (non-per-thread)
+      FS base would clobber one and cause cross-talk. Both read their own marker → pass.
+
+**Gate (met):** `thread-tls` passes (3×); full guardrail 25/25 (the switch-time thread-pointer
+restore touches every context switch); `zig build`/`zig build test` clean.
+
+> **Deferred: the Zig `threadlocal` *compiler* layer.** Real `threadlocal` variables need
+> the ELF **variant-II TLS** surface — `.tdata`/`.tbss` sections + a `PT_TLS` program header
+> in `user.ld`, a runtime that copies the template with exact negative-offset layout, and
+> the `.large`-code-model TLS section names — a high-uncertainty lift for a feature with
+> **no consumer today** (threading.md scopes it "only if a consumer needs it"). What lands
+> here is the load-bearing piece — the per-thread thread pointer, context-switched — so adding the
+> compiler layer later is purely runtime+linker work on top, no kernel change. `getCurrentId`
+> stays the `thread_self` syscall (M6) rather than an fs self-slot (which would need the
+> main thread's TLS set up in `_start` too).
 
 **Gate:** `thread-tls` passes; full `thread-*` suite + guardrail green.
 
-### M11 — `RwLock`, `WaitGroup`, and host-testable sync
+### M11 — `RwLock`, `WaitGroup`, and host-testable sync ✅
 
-- [ ] `runtime.Thread.RwLock` and `WaitGroup` on the existing `Futex`/`Mutex`/
-      `Condition`.
-- [ ] A compile-time `Futex` seam: syscalls on the danos target, a host-backed impl under
-      `zig build test`, so the `Mutex`/`Condition`/`RwLock` state machines run as host
-      unit tests (fast iteration, no QEMU).
-- [ ] `-Dtest-case=thread-rwlock` (`smp: 4`): many readers + writers over an `RwLock` keep
-      an invariant (a reader never observes a half-written value); host tests cover the
-      lock transitions.
+- [x] `runtime.Thread.RwLock` (reader-preferring: `>0` readers / `-1` writer / `0` free,
+      with `lock`/`tryLock`/`unlock` + `lockShared`/`tryLockShared`/`unlockShared`) and
+      `WaitGroup` (`start`/`finish`/`wait`), both on the existing `Mutex`/`Condition`.
+- [x] A compile-time `Futex` seam gated on `builtin.os.tag == .freestanding`: the futex
+      syscalls on danos, a spin+yield mock off-target (Zig 0.16 has no `std.Thread.Futex`;
+      `wake` is a no-op since the state machines re-check). `thread.zig` is wired into
+      `zig build test`, so `Mutex`/`RwLock`/`WaitGroup` run as **host unit tests** with real
+      `std.Thread` threads (`test` blocks only compile under test).
+- [x] `-Dtest-case=thread-rwlock` (`smp: 4`): 2 writers set both halves of a value under
+      the exclusive lock while 3 readers check the halves match under the shared lock —
+      zero half-write observations across ~150k reads. Host tests cover the Mutex,
+      RwLock, and WaitGroup state machines.
 
-**Gate:** host `zig build test` covers the sync primitives; `thread-rwlock` passes;
-guardrail green.
+**Gate (met):** `zig build test` covers the sync primitives (host threads); `thread-rwlock`
+passes (3×); full Done gate **26/26** (whole `thread-*` suite + guardrail); `zig build`
+clean.
 
 ---
 
 ## Deferred (explicitly not in this plan)
 
-- **Cross-process shared-memory futex** — the `(aspace, vaddr)` key can become a
+- **Cross-process shared-memory futex** — the `(address_space, virtual_address)` key can become a
   physical-address key so two processes share a futex through an [shm](display-v2.md)
   region. Not needed for intra-process threads.
 - **Per-thread priorities / affinity distinct from the process** — threads inherit the
diff --git a/docs/threading.md b/docs/threading.md
index abcd350..35ba304 100644
--- a/docs/threading.md
+++ b/docs/threading.md
@@ -2,12 +2,13 @@
 
 A note on danos **threads** — several tasks sharing one address space — provided by a
 `runtime.Thread` type that mirrors the shape of Zig's `std.Thread` while keeping every
-kernel entry behind the [runtime](../library/runtime). **Built** (M1–M6, see
-[threading-plan.md](threading-plan.md)): `spawn`/`join`/`detach`, cross-core
-parallelism, a futex (`futex_wait`/`futex_wake`), and a futex-backed
-`Mutex`/`Condition`/`Semaphore`, plus `getCurrentId`/`currentCore`. Deferred by design
-(no consumer yet): per-thread `threadlocal` TLS, `RwLock`/`WaitGroup`, and migrating
-`join` to a futex completion word — see the plan's M5/M6 notes. The analysis is against
+kernel entry behind the [runtime](../library/runtime). **Built** (M1–M11, see
+[threading-plan.md](threading-plan.md)): `spawn`/`join`/`detach`, cross-core parallelism,
+a futex, `Mutex`/`Condition`/`Semaphore`/`RwLock`/`WaitGroup`, `getCurrentId`/`currentCore`,
+per-thread thread-pointer TLS, thread-safe allocation, and a task reaper that reclaims dead
+tasks' kernel stacks. Deferred by design (no consumer yet): the Zig `threadlocal`
+*compiler* layer (the per-thread thread pointer is in place, so it's runtime+linker work on top) and
+detached-thread user-stack reclaim — see the plan's M9/M10 notes. The analysis is against
 **Zig 0.16** (the pinned toolchain); `std.Thread`'s internals move between releases, so
 treat upstream shapes as "0.16.x."
 
@@ -147,10 +148,10 @@ Plus one invariant change with no new syscall: **address-space reference countin
 
 ### Address-space reference counting
 
-Today an address space is 1:1 with a task: `spawnUserLocked` records `aspace` on the
-Task, and teardown does `destroyAddressSpace(t.aspace)` when **any** user task exits
+Today an address space is 1:1 with a task: `spawnUserLocked` records `address_space` on the
+Task, and teardown does `destroyAddressSpace(t.address_space)` when **any** user task exits
 ([scheduler.zig](../system/kernel/scheduler.zig)). With threads, several tasks share
-one `aspace`, so the first to exit would rip the address space out from under its
+one `address_space`, so the first to exit would rip the address space out from under its
 siblings.
 
 Fix: a small refcount keyed by the address-space root (`createAddressSpace` in
@@ -161,7 +162,7 @@ that must land and be proven before anything shares an address space.
 
 ### `thread_spawn` and the trampoline
 
-The scheduler already accepts an arbitrary `aspace` and does **not** smuggle values
+The scheduler already accepts an arbitrary `address_space` and does **not** smuggle values
 through registers — `startUserTask` reads the entry/stack from the Task and
 `jumpToUser`s ([scheduler.zig](../system/kernel/scheduler.zig)). That makes the thread
 path clean:
@@ -170,7 +171,7 @@ path clean:
    `{ fn_ptr, args_tuple, completion }`, the std "Instance" pattern — and writes the
    closure pointer to the **top word of the new stack**.
 2. It calls `thread_spawn(entry = &threadTrampoline, stack_top, arg = closure_ptr)`.
-   The kernel calls the same `spawnUserLocked` path with the **caller's aspace**
+   The kernel calls the same `spawnUserLocked` path with the **caller's address space**
    (refcount++), `entry`, and `user_sp = stack_top`.
 3. `threadTrampoline` (a small runtime shim) reads the closure off its stack, calls
    the user function, then calls `thread_exit`. No new register ABI — the closure
@@ -184,7 +185,7 @@ Unlike a process start, there is **no** System V argc/argv/auxv block
 
 - **`thread_exit`** marks the task dead and hands the kernel the thread's user-stack
   range. The kernel reaps the task on the scheduler (already running on a *kernel*
-  stack, so it can safely unmap the user stack), decrements the aspace refcount, and
+  stack, so it can safely unmap the user stack), decrements the address-space refcount, and
   frees the task slot.
 - **`join` — Stage 1** reuses the existing exit-notification machinery
   ([process-lifecycle.md](process-lifecycle.md)): `spawn` passes a per-thread
@@ -205,12 +206,12 @@ Unlike a process start, there is **no** System V argc/argv/auxv block
 call the futex wrappers on the slow path — the same construction `std.Thread` uses,
 so the algorithms port directly.
 
-Keying: threads share an address space, so a **virtual address within that aspace**
-identifies a futex uniquely; the kernel keys its wait queue by `(aspace_root, vaddr)`.
-Keying by the **physical** address instead (translate `vaddr -> paddr` on entry) is a
+Keying: threads share an address space, so a **virtual address within that address space**
+identifies a futex uniquely; the kernel keys its wait queue by `(address_space_root, virtual_address)`.
+Keying by the **physical** address instead (translate `virtual_address -> physical_address` on entry) is a
 deliberate forward door: it lets two *processes* share a futex through an
 [shm](display-v2.md) region later, without changing the API. We start with the
-private-per-aspace key and note the physical-key upgrade.
+private-per-address-space key and note the physical-key upgrade.
 
 No spinning: a contended lock parks the task in the kernel and the core is free to run
 other work or `hlt` ([halting.md](halting.md)). This is why futex is a locked
@@ -218,12 +219,12 @@ decision, not a "maybe later."
 
 ### TLS and `getCurrentId`
 
-danos sets up no `fs.base` TLS today (fine under `single_threaded`). Two scoped needs:
+danos sets up no thread-pointer TLS today (fine under `single_threaded`). Two scoped needs:
 
 - **`getCurrentId`** returns the kernel task id — either a trivial syscall or, better,
   a value the runtime stashes in a per-thread control block.
-- **`threadlocal` variables** need a real per-thread TLS block and `fs.base` set per
-  thread. `thread_spawn` sets `fs.base` to a runtime-allocated per-thread block; full
+- **`threadlocal` variables** need a real per-thread TLS block and the thread pointer set per
+  thread. `thread_spawn` sets the thread pointer to a runtime-allocated per-thread block; full
   `threadlocal` support is Stage 3, only if a consumer needs it. Nothing in the core
   spawn/join/mutex path requires `threadlocal`.
 
@@ -237,14 +238,14 @@ it may call `runtime.Thread.spawn`. Everyone else stays single-threaded and lean
 ## Interaction with the rest of the kernel
 
 - **Scheduler / SMP** ([scheduling.md](scheduling.md), [smp.md](smp.md)): a thread is
-  just another `Task` with an `aspace` shared with its siblings; the existing
+  just another `Task` with an `address_space` shared with its siblings; the existing
   per-core ready queues, priorities, and affinity apply unchanged. Threads of one
   process can run on different cores simultaneously — that is the point.
 - **Halting** ([halting.md](halting.md)): futex-parked waiters keep the "idle core
   halts" property intact under lock contention — no busy-wait.
 - **Lifecycle** ([process-lifecycle.md](process-lifecycle.md)): killing a process
-  must kill *all* its threads and only then drop the last aspace ref. The kill path
-  already targets a process; it fans out to every task on that aspace.
+  must kill *all* its threads and only then drop the last address-space ref. The kill path
+  already targets a process; it fans out to every task on that address space.
 - **Resilience** ([resilience.md](resilience.md)): a faulting thread kills its whole
   process (shared fate). The supervisor restarts the **process**, which respawns its
   threads from a known-good state — restart granularity stays the process.
@@ -257,8 +258,8 @@ The ordered, `/loop`-runnable milestones live in
 a verifiable gate (`python3 test/qemu_test.py <case>`, asserting serial markers;
 `zig build test` for host unit tests). The stages below are the shape it expands.
 
-- **Stage 0 — address-space refcount.** Refcount on the aspace root; teardown destroys
-  at zero. No API yet; nothing shares an aspace, so refcount is 1 everywhere.
+- **Stage 0 — address-space refcount.** Refcount on the address-space root; teardown destroys
+  at zero. No API yet; nothing shares an address space, so refcount is 1 everywhere.
   *Gate:* the full QEMU suite stays green (no regression) — proves the reframing is
   invisible until used.
 - **Stage 1 — spawn / join / detach.** `thread_spawn` + `thread_exit`, the trampoline,
@@ -272,7 +273,7 @@ a verifiable gate (`python3 test/qemu_test.py <case>`, asserting serial markers;
   word. *Gate:* `-Dtest-case=thread-mutex` — a bounded producer/consumer over a
   `Mutex` + `Condition` moves K items with no lost wakeups and no busy-wait (assert
   the consumer blocked, e.g. via a low idle tick count).
-- **Stage 3 — polish.** Per-thread TLS / `fs.base` and `threadlocal` (only if a
+- **Stage 3 — polish.** Per-thread TLS / thread pointer and `threadlocal` (only if a
   consumer needs it), `RwLock`/`WaitGroup` as demanded, and this doc's cases wired
   into [test/qemu_test.py](../test/qemu_test.py).
 
@@ -292,7 +293,7 @@ are — user code never names a syscall.
 - **No thread priorities distinct from the process.** Threads inherit the process
   priority; per-thread priority is a later question if it ever earns its keep.
 - **No cross-process shared-memory futex yet** — the physical-address key leaves the
-  door open, but the first cut is private-per-aspace.
+  door open, but the first cut is private-per-address-space.
 - **No `pthread`/POSIX surface.** The API is `std.Thread`-shaped Zig, nothing more.
 
 ## The self-hosting endgame
diff --git a/docs/vdso.md b/docs/vdso.md
index 34203cd..3c2d89c 100644
--- a/docs/vdso.md
+++ b/docs/vdso.md
@@ -119,7 +119,7 @@ One table entry per kernel call, C ABI (System V AMD64), names prefixed
 returns are `u64`, errors return as negative values exactly as today.
 
 The calls that return two values in `rax:rdx` today — `dma_alloc`
-(vaddr + paddr), `msi_bind` (address + data), `shm_create` (vaddr + handle) —
+(virtual_address + physical_address), `msi_bind` (address + data), `shm_create` (virtual_address + handle) —
 become functions returning a two-`u64` struct. The System V ABI returns a
 16-byte struct in `rax:rdx`, so the stub is a plain `syscall; ret` — the
 C-ABI spelling of the existing convention, at zero cost.
diff --git a/library/runtime/heap.zig b/library/runtime/heap.zig
index 0ff4fd7..0e525b7 100644
--- a/library/runtime/heap.zig
+++ b/library/runtime/heap.zig
@@ -9,15 +9,32 @@
 //! — `grow` asks the kernel for pages via `mmap` instead of mapping frames
 //! itself, and the kernel picks the base address.
 //!
-//! Single-threaded and 16-byte maximum alignment, exactly like the kernel heap; a
-//! lock and larger alignments come when user programs gain threads.
+//! 16-byte maximum alignment, exactly like the kernel heap. The free list is guarded by
+//! a `Thread.Mutex` **only in multi-threaded binaries** (`addThreadedUserBinary`): the
+//! guard is gated on `builtin.single_threaded`, so an ordinary single-threaded binary
+//! compiles it out and pays nothing, while a threaded one can allocate safely from
+//! several threads at once (docs/threading-plan.md M7). The lock lives at the two
+//! free-list mutators — `rawAlloc`/`rawFree` — which every entry point funnels through.
 
 const std = @import("std");
+const builtin = @import("builtin");
 const abi = @import("abi");
 const system_calls = @import("system.zig");
+const Mutex = @import("thread.zig").Thread.Mutex;
 
 const page_size = abi.page_size;
 
+/// Guards `free_list`. A no-op in single-threaded builds (compiled out); a real futex
+/// mutex in threaded ones. Uncontended acquisition is a single CAS — no syscall.
+var heap_mutex: Mutex = .{};
+
+inline fn lockHeap() void {
+    if (comptime !builtin.single_threaded) heap_mutex.lock();
+}
+inline fn unlockHeap() void {
+    if (comptime !builtin.single_threaded) heap_mutex.unlock();
+}
+
 /// A block header, at the start of every block; while free it also links the
 /// free list via `next`.
 const Block = extern struct {
@@ -84,8 +101,11 @@ fn insertFree(block: *Block) void {
     }
 }
 
-/// Allocate `len` bytes (16-byte aligned), or null if out of memory.
+/// Allocate `len` bytes (16-byte aligned), or null if out of memory. Holds the heap lock
+/// across the free-list search and any `grow` (which also touches the free list).
 fn rawAlloc(len: usize) ?[*]u8 {
+    lockHeap();
+    defer unlockHeap();
     const need = alignUp(header_size + len, 16);
 
     var attempts: u32 = 0;
@@ -119,6 +139,8 @@ fn rawAlloc(len: usize) ?[*]u8 {
 }
 
 fn rawFree(ptr: [*]u8) void {
+    lockHeap();
+    defer unlockHeap();
     const block: *Block = @ptrFromInt(@intFromPtr(ptr) - header_size);
     insertFree(block);
 }
diff --git a/library/runtime/shm.zig b/library/runtime/shm.zig
index 1a5ec5b..a75056c 100644
--- a/library/runtime/shm.zig
+++ b/library/runtime/shm.zig
@@ -22,7 +22,7 @@ pub const Region = struct {
 };
 
 /// Grant `len` bytes (rounded up to whole pages) of shareable, zeroed, cacheable RAM.
-/// Returns the region or null on failure. Two return values — vaddr in rax, handle in rdx —
+/// Returns the region or null on failure. Two return values — virtual_address in rax, handle in rdx —
 /// so this is a hand-written stub like `dma.alloc`.
 pub fn create(len: usize) ?Region {
     var rax: usize = undefined;
diff --git a/library/runtime/thread.zig b/library/runtime/thread.zig
index 2f84ebb..0a59df0 100644
--- a/library/runtime/thread.zig
+++ b/library/runtime/thread.zig
@@ -11,19 +11,27 @@
 //! A binary must be built multi-threaded (`addThreadedUserBinary`) before it may spawn.
 
 const std = @import("std");
+const builtin = @import("builtin");
 const abi = @import("abi");
 const sc = @import("system-call.zig");
 const system = @import("system.zig");
-const ipc = @import("ipc.zig");
+
+/// True in a real danos binary; false when this module is compiled for host unit tests.
+/// The `Futex` seam and the test blocks below branch on it so the lock/condvar state
+/// machines can be exercised on the host against `std.Thread.Futex` (docs/threading-plan.md
+/// M11), while the danos build uses the futex syscalls.
+const on_danos = builtin.os.tag == .freestanding;
 
 /// A thread stack, if the caller does not override it. 64 KiB of mmap'd, zeroed pages.
 pub const default_stack_size: usize = 64 * 1024;
 
+/// Bytes reserved at the top of each thread's stack for its per-thread TLS block (the
+/// self-pointer plus scratch slots reachable via `%fs`). docs/threading-plan.md M10.
+const tls_block_size: usize = 64;
+
 pub const Thread = struct {
-    /// The kernel task id of the spawned thread.
+    /// The kernel task id of the spawned thread — what `join` waits on.
     tid: u32,
-    /// The endpoint the kernel notifies when this thread ends — what `join` blocks on.
-    exit_endpoint: ipc.Handle,
     /// The mmap'd stack, reclaimed by `join` (or at process exit after `detach`).
     stack_base: usize,
     stack_size: usize,
@@ -46,51 +54,52 @@ pub const Thread = struct {
     pub fn spawn(config: SpawnConfig, comptime function: anytype, args: anytype) SpawnError!Thread {
         const Args = @TypeOf(args);
         const Closure = struct {
+            tls_base: usize,
             args: Args,
-            /// Entered directly by the kernel with `self` in rdi (C ABI). Runs the user
-            /// function, then ends the thread — never returns.
+            /// Entered directly by the kernel with `self` in rdi (C ABI). Establishes this
+            /// thread's TLS pointer, runs the user function, then ends the thread.
             fn entry(self_addr: usize) callconv(.c) noreturn {
                 const self: *@This() = @ptrFromInt(self_addr);
+                setThreadPointer(self.tls_base); // per-thread thread pointer before any user code
                 @call(.auto, function, self.args);
                 exitThread();
             }
         };
 
-        // The endpoint the kernel posts this thread's exit notification to.
-        const endpoint = ipc.createIpcEndpoint() orelse return error.SystemResources;
-
         const base = system.mmap(config.stack_size, system.PROT_READ | system.PROT_WRITE);
         if (system.mmapFailed(base)) return error.SystemResources;
 
-        // Lay the closure at the very top of the thread's own stack, then start the
-        // thread's rsp just below it (16-aligned minus 8, the alignment a `call` leaves
-        // for a C-ABI entry) so the growing stack never overwrites the args.
+        // Top of the thread's own stack, downward: the closure, then a small per-thread TLS
+        // block (the thread pointer points here; slot 0 is the variant-II self-pointer, the rest is
+        // scratch for user TLS), then the stack proper (rsp starts below the TLS block, so
+        // the growing stack never overwrites either).
         var closure_addr = (base + config.stack_size) - @sizeOf(Closure);
         closure_addr &= ~@as(usize, @alignOf(Closure) - 1); // align the closure down
-        const closure: *Closure = @ptrFromInt(closure_addr);
-        closure.* = .{ .args = args };
 
-        var stack_top = closure_addr & ~@as(usize, 15); // 16-align below the closure
+        const tls_base = (closure_addr - tls_block_size) & ~@as(usize, 15);
+        const tls: [*]usize = @ptrFromInt(tls_base);
+        tls[0] = tls_base; // self-pointer (fs:0), as the x86_64 TLS ABI expects
+
+        const closure: *Closure = @ptrFromInt(closure_addr);
+        closure.* = .{ .tls_base = tls_base, .args = args };
+
+        var stack_top = tls_base & ~@as(usize, 15); // 16-align below the TLS block
         stack_top -= 8; // ...then rsp % 16 == 8 at the C entry
 
-        const tid = threadSpawn(@intFromPtr(&Closure.entry), stack_top, closure_addr, endpoint);
+        const tid = threadSpawn(@intFromPtr(&Closure.entry), stack_top, closure_addr);
         if (threadSpawnFailed(tid)) {
             _ = system.munmap(base, config.stack_size);
             return error.SystemResources;
         }
-        return .{ .tid = @intCast(tid), .exit_endpoint = endpoint, .stack_base = base, .stack_size = config.stack_size };
+        return .{ .tid = @intCast(tid), .stack_base = base, .stack_size = config.stack_size };
     }
 
     /// Block until this thread finishes, then reclaim its stack. Mirrors
     /// `std.Thread.join`. The exit endpoint is private to this thread, so the first
     /// child-exit notification on it is this thread's.
     pub fn join(self: Thread) void {
-        var receive: [0]u8 = undefined;
-        while (true) {
-            const got = ipc.replyWait(self.exit_endpoint, &.{}, &receive, null);
-            if (got.isChildExit() and got.childProcessId() == self.tid) break;
-        }
-        _ = system.munmap(self.stack_base, self.stack_size);
+        _ = sc.systemCall1(.thread_join, self.tid); // block until the thread has exited
+        _ = system.munmap(self.stack_base, self.stack_size); // reclaim its (now-vacated) stack
     }
 
     /// Relinquish the right to join: never wait for or reclaim this thread. Its stack is
@@ -119,17 +128,37 @@ pub const Thread = struct {
         /// the value already differs (safe against spurious returns, as in std): the
         /// caller re-checks its condition in a loop.
         pub fn wait(ptr: *const std.atomic.Value(u32), expect: u32) void {
-            _ = futexWait(@intFromPtr(ptr), expect, 0);
+            if (comptime on_danos) {
+                _ = futexWait(@intFromPtr(ptr), expect, 0);
+            } else {
+                // Host unit-test mock: spin+yield until the value changes (`wake` is a
+                // no-op — the callers re-check their condition in a loop anyway). Correct,
+                // if busy; fine for the state-machine tests.
+                while (ptr.load(.acquire) == expect) std.Thread.yield() catch {};
+            }
         }
 
         /// As `wait`, but returns `error.Timeout` if `timeout_ns` elapses first.
         pub fn timedWait(ptr: *const std.atomic.Value(u32), expect: u32, timeout_ns: u64) error{Timeout}!void {
-            if (futexWait(@intFromPtr(ptr), expect, timeout_ns) == abi.futex_timed_out) return error.Timeout;
+            if (comptime on_danos) {
+                if (futexWait(@intFromPtr(ptr), expect, timeout_ns) == abi.futex_timed_out) return error.Timeout;
+            } else {
+                var spins: u64 = 0;
+                const limit = timeout_ns / 1000 + 1;
+                while (ptr.load(.acquire) == expect) : (spins += 1) {
+                    if (spins >= limit) return error.Timeout;
+                    std.Thread.yield() catch {};
+                }
+            }
         }
 
         /// Wake up to `max_waiters` threads blocked on `ptr`.
         pub fn wake(ptr: *const std.atomic.Value(u32), max_waiters: u32) void {
-            _ = futexWake(@intFromPtr(ptr), max_waiters);
+            if (comptime on_danos) {
+                _ = futexWake(@intFromPtr(ptr), max_waiters);
+            } else {
+                // host mock: spin-waiters re-check their condition, so no wake is needed.
+            }
         }
     };
 
@@ -230,10 +259,101 @@ pub const Thread = struct {
             s.cond.signal();
         }
     };
+
+    /// A reader/writer lock, `std.Thread.RwLock`-shaped: many concurrent readers OR one
+    /// exclusive writer. Reader-preferring (a steady stream of readers can delay a writer),
+    /// built on `Mutex` + `Condition` over a signed state: `>0` = that many readers hold
+    /// it, `-1` = a writer holds it, `0` = free.
+    pub const RwLock = struct {
+        mutex: Mutex = .{},
+        cond: Condition = .{},
+        state: i64 = 0,
+
+        /// Acquire shared (read) access, blocking while a writer holds the lock.
+        pub fn lockShared(rw: *RwLock) void {
+            rw.mutex.lock();
+            defer rw.mutex.unlock();
+            while (rw.state < 0) rw.cond.wait(&rw.mutex);
+            rw.state += 1;
+        }
+
+        /// Try to acquire shared access without blocking.
+        pub fn tryLockShared(rw: *RwLock) bool {
+            rw.mutex.lock();
+            defer rw.mutex.unlock();
+            if (rw.state < 0) return false;
+            rw.state += 1;
+            return true;
+        }
+
+        /// Release shared access; wake a waiting writer once the last reader leaves.
+        pub fn unlockShared(rw: *RwLock) void {
+            rw.mutex.lock();
+            defer rw.mutex.unlock();
+            rw.state -= 1;
+            if (rw.state == 0) rw.cond.broadcast();
+        }
+
+        /// Acquire exclusive (write) access, blocking until no readers or writer remain.
+        pub fn lock(rw: *RwLock) void {
+            rw.mutex.lock();
+            defer rw.mutex.unlock();
+            while (rw.state != 0) rw.cond.wait(&rw.mutex);
+            rw.state = -1;
+        }
+
+        /// Try to acquire exclusive access without blocking.
+        pub fn tryLock(rw: *RwLock) bool {
+            rw.mutex.lock();
+            defer rw.mutex.unlock();
+            if (rw.state != 0) return false;
+            rw.state = -1;
+            return true;
+        }
+
+        /// Release exclusive access; wake all waiters (they re-check their condition).
+        pub fn unlock(rw: *RwLock) void {
+            rw.mutex.lock();
+            defer rw.mutex.unlock();
+            rw.state = 0;
+            rw.cond.broadcast();
+        }
+    };
+
+    /// A `std.Thread.WaitGroup`-shaped counter: `start` before spawning work, `finish` as
+    /// each unit completes, `wait` blocks until the count returns to zero.
+    pub const WaitGroup = struct {
+        mutex: Mutex = .{},
+        cond: Condition = .{},
+        counter: usize = 0,
+
+        /// Register one pending unit of work.
+        pub fn start(wg: *WaitGroup) void {
+            wg.mutex.lock();
+            defer wg.mutex.unlock();
+            wg.counter += 1;
+        }
+
+        /// Mark one unit done; wake waiters if that was the last.
+        pub fn finish(wg: *WaitGroup) void {
+            wg.mutex.lock();
+            defer wg.mutex.unlock();
+            wg.counter -= 1;
+            if (wg.counter == 0) wg.cond.broadcast();
+        }
+
+        /// Block until every started unit has finished.
+        pub fn wait(wg: *WaitGroup) void {
+            wg.mutex.lock();
+            defer wg.mutex.unlock();
+            while (wg.counter != 0) wg.cond.wait(&wg.mutex);
+        }
+    };
 };
 
 /// thread_spawn(entry, stack_top, arg, exit_endpoint) -> tid, or a wrapped error.
-fn threadSpawn(entry: usize, stack_top: usize, arg: usize, exit_endpoint: ipc.Handle) usize {
+fn threadSpawn(entry: usize, stack_top: usize, arg: usize) usize {
+    const exit_endpoint: usize = @intCast(abi.no_cap); // join uses thread_join, not an endpoint
     return sc.systemCall4(.thread_spawn, entry, stack_top, arg, exit_endpoint);
 }
 
@@ -249,6 +369,11 @@ fn exitThread() noreturn {
     unreachable;
 }
 
+/// Set the calling thread's FS base (its user TLS thread pointer).
+fn setThreadPointer(addr: usize) void {
+    _ = sc.systemCall1(.set_thread_pointer, addr);
+}
+
 /// futex_wait(addr, expect, timeout_ns) -> status (abi.futex_*).
 fn futexWait(addr: usize, expect: u32, timeout_ns: u64) usize {
     return sc.systemCall3(.futex_wait, addr, expect, timeout_ns);
@@ -258,3 +383,102 @@ fn futexWait(addr: usize, expect: u32, timeout_ns: u64) usize {
 fn futexWake(addr: usize, count: u32) usize {
     return sc.systemCall2(.futex_wake, addr, count);
 }
+
+// --- host unit tests (docs/threading-plan.md M11) ---------------------------
+//
+// These run under `zig build test` on the host: the `Futex` seam above uses
+// `std.Thread.Futex` off-danos, so the lock/condvar state machines can be exercised by
+// real host threads. They are never compiled into a danos binary (test blocks only build
+// under test), so their `std.Thread` use is fine even though `std.Thread` is unavailable
+// on the freestanding target.
+
+test "Mutex serialises concurrent increments across host threads" {
+    var m: Thread.Mutex = .{};
+    var counter: u64 = 0;
+    const workers = 8;
+    const per = 20_000;
+    const Ctx = struct {
+        m: *Thread.Mutex,
+        c: *u64,
+        fn run(ctx: @This()) void {
+            var i: usize = 0;
+            while (i < per) : (i += 1) {
+                ctx.m.lock();
+                ctx.c.* += 1;
+                ctx.m.unlock();
+            }
+        }
+    };
+    var handles: [workers]std.Thread = undefined;
+    for (&handles) |*h| h.* = try std.Thread.spawn(.{}, Ctx.run, .{Ctx{ .m = &m, .c = &counter }});
+    for (handles) |h| h.join();
+    try std.testing.expectEqual(@as(u64, workers * per), counter);
+}
+
+test "RwLock never lets a reader observe a half-written pair" {
+    var rw: Thread.RwLock = .{};
+    var a: u64 = 0;
+    var b: u64 = 0; // invariant while a lock is held: a == b
+    var stop = std.atomic.Value(bool).init(false);
+    var ok = std.atomic.Value(bool).init(true);
+
+    const Writer = struct {
+        rw: *Thread.RwLock,
+        a: *u64,
+        b: *u64,
+        stop: *std.atomic.Value(bool),
+        fn run(w: @This()) void {
+            var v: u64 = 1;
+            while (!w.stop.load(.acquire)) : (v +%= 1) {
+                w.rw.lock();
+                w.a.* = v; // update both halves under the exclusive lock...
+                w.b.* = v;
+                w.rw.unlock();
+            }
+        }
+    };
+    const Reader = struct {
+        rw: *Thread.RwLock,
+        a: *u64,
+        b: *u64,
+        ok: *std.atomic.Value(bool),
+        fn run(r: @This()) void {
+            var i: usize = 0;
+            while (i < 200_000) : (i += 1) {
+                r.rw.lockShared();
+                if (r.a.* != r.b.*) r.ok.store(false, .release); // ...so a reader must never see them differ
+                r.rw.unlockShared();
+            }
+        }
+    };
+
+    var writers: [2]std.Thread = undefined;
+    for (&writers) |*w| w.* = try std.Thread.spawn(.{}, Writer.run, .{Writer{ .rw = &rw, .a = &a, .b = &b, .stop = &stop }});
+    var readers: [4]std.Thread = undefined;
+    for (&readers) |*rd| rd.* = try std.Thread.spawn(.{}, Reader.run, .{Reader{ .rw = &rw, .a = &a, .b = &b, .ok = &ok }});
+    for (readers) |rd| rd.join();
+    stop.store(true, .release);
+    for (writers) |w| w.join();
+    try std.testing.expect(ok.load(.acquire));
+}
+
+test "WaitGroup blocks until every started unit finishes" {
+    var wg: Thread.WaitGroup = .{};
+    var done = std.atomic.Value(u32).init(0);
+    const n = 6;
+    const Ctx = struct {
+        wg: *Thread.WaitGroup,
+        done: *std.atomic.Value(u32),
+        fn run(c: @This()) void {
+            _ = c.done.fetchAdd(1, .monotonic);
+            c.wg.finish();
+        }
+    };
+    var i: usize = 0;
+    while (i < n) : (i += 1) wg.start();
+    var handles: [n]std.Thread = undefined;
+    for (&handles) |*h| h.* = try std.Thread.spawn(.{}, Ctx.run, .{Ctx{ .wg = &wg, .done = &done }});
+    wg.wait(); // must not return until all n finished
+    try std.testing.expectEqual(@as(u32, n), done.load(.acquire));
+    for (handles) |h| h.join();
+}
diff --git a/system/abi.zig b/system/abi.zig
index d536237..7ab350b 100644
--- a/system/abi.zig
+++ b/system/abi.zig
@@ -39,13 +39,13 @@ pub const SystemCall = enum(u64) {
     ipc_reply_wait = 10, // ipc_reply_wait(h, reply, len, receive, cap) -> receive_len (+badge in rdx)
     device_enumerate = 11, // device_enumerate(buffer, maximum) -> count: snapshot the device table
     device_claim = 12, // device_claim(id) -> ok: take exclusive ownership of a device
-    mmio_map = 13, // mmio_map(id, resource_index) -> vaddr: map a claimed device's MMIO into this AS
+    mmio_map = 13, // mmio_map(id, resource_index) -> virtual_address: map a claimed device's MMIO into this address space
     irq_bind = 14, // irq_bind(id, resource_index, endpoint): deliver a device IRQ as an IPC notification
     irq_ack = 15, // irq_ack(id, resource_index): re-arm a bound IRQ after servicing it
     device_register = 16, // device_register(parent_id, descriptor) -> id: publish a child of a device you claimed
     system_spawn = 17, // system_spawn(name_ptr, name_len, arguments_ptr, arguments_len, exit_endpoint) -> child process id: start a named initial-ramdisk binary as a new ring-3 process
-    dma_alloc = 18, // dma_alloc(len, flags) -> vaddr (rax), paddr (rdx): contiguous, pinned, uncacheable DMA memory
-    dma_free = 19, // dma_free(vaddr, len) -> 0: release a prior dma_alloc
+    dma_alloc = 18, // dma_alloc(len, flags) -> virtual_address (rax), physical_address (rdx): contiguous, pinned, uncacheable DMA memory
+    dma_free = 19, // dma_free(virtual_address, len) -> 0: release a prior dma_alloc
     msi_bind = 20, // msi_bind(device_id, endpoint) -> address (rax), data (rdx): a per-device MSI vector for a claimed device
     io_read = 21, // io_read(device_id, resource_index, offset, width) -> value: read a port in a claimed device's io_port resource
     io_write = 22, // io_write(device_id, resource_index, offset, width, value) -> 0: write a port in a claimed device's io_port resource
@@ -60,15 +60,17 @@ pub const SystemCall = enum(u64) {
     timer_bind = 31, // timer_bind(endpoint, ms) -> 0/-errno: one-shot timer — posts a notification when ms elapse
     klog_read = 32, // klog_read(offset, ptr, len) -> bytes copied: copy the kernel RAM log buffer out to a user buffer (for persisting the boot log to disk)
     wall_clock = 33, // wall_clock() -> Unix epoch seconds (UTC): the RTC wall-clock time, for filesystem timestamps (mtime). Monotonic time is `clock`.
-    shm_create = 34, // shm_create(len) -> vaddr (rax), handle (rdx): a shareable, zeroed, cacheable RAM region mapped into this AS; the handle is a capability passed to another process as an ipc_call send_cap (docs/display-v2.md)
-    shm_map = 35, // shm_map(cap) -> vaddr: map the shared region named by a received capability into this AS (the same physical pages the creator sees)
-    shm_physical = 36, // shm_physical(cap) -> paddr: the guest-physical base of a shared region held by capability, so a driver can program it into a device (e.g. virtio-gpu attach_backing); the pages are contiguous (docs/display-v2.md)
+    shm_create = 34, // shm_create(len) -> virtual_address (rax), handle (rdx): a shareable, zeroed, cacheable RAM region mapped into this AS; the handle is a capability passed to another process as an ipc_call send_cap (docs/display-v2.md)
+    shm_map = 35, // shm_map(cap) -> virtual_address: map the shared region named by a received capability into this address space (the same physical pages the creator sees)
+    shm_physical = 36, // shm_physical(cap) -> physical_address: the guest-physical base of a shared region held by capability, so a driver can program it into a device (e.g. virtio-gpu attach_backing); the pages are contiguous (docs/display-v2.md)
     thread_spawn = 37, // thread_spawn(entry, stack_top, arg, exit_endpoint) -> tid: start a task sharing the caller's address space at `entry` on `stack_top`, `arg` in rdi; exit_endpoint (a handle, or no_cap) is notified when it ends — how join waits (docs/threading.md)
     thread_exit = 38, // thread_exit(): end the calling thread, dropping one reference to its address space (destroyed on the last)
     current_core = 39, // current_core() -> index: the dense 0-based index of the core the caller is running on (for parallelism/affinity introspection)
     futex_wait = 40, // futex_wait(addr, expected, timeout_ns) -> status: if *addr == expected, block until woken or the timeout; returns futex_woken/mismatch/timed_out (docs/threading.md)
     futex_wake = 41, // futex_wake(addr, count) -> woken: wake up to `count` tasks blocked in futex_wait on `addr` in this address space
     thread_self = 42, // thread_self() -> tid: the calling thread's kernel task id (runtime.Thread.getCurrentId)
+    thread_join = 43, // thread_join(tid) -> 0: block until the thread with id `tid` has exited (runtime.Thread.join; no per-thread IPC endpoint) (docs/threading.md)
+    set_thread_pointer = 44, // set_thread_pointer(addr) -> 0: set the caller's thread pointer (user-space TLS base; x86_64 IA32_FS_BASE, aarch64 TPIDR_EL0); restored per task across context switches (docs/threading-plan.md M10)
     _,
 };
 
diff --git a/system/kernel/architecture/x86_64/cpu.zig b/system/kernel/architecture/x86_64/cpu.zig
index 3625c2a..9bbd7a9 100644
--- a/system/kernel/architecture/x86_64/cpu.zig
+++ b/system/kernel/architecture/x86_64/cpu.zig
@@ -304,6 +304,17 @@ pub fn cpuLocal() usize {
     return pcpu.scheduler();
 }
 
+const ia32_fs_base = 0xC000_0100;
+
+/// Set the user-space TLS **thread pointer** — the arch-neutral name the generic scheduler
+/// calls (`architecture.setThreadPointer`). On x86_64 that is the FS-segment base
+/// (`IA32_FS_BASE`); an aarch64 port implements the same call against `TPIDR_EL0`. The
+/// kernel never touches FS, so this only affects the user task that runs next, which the
+/// scheduler restores per task across context switches (docs/threading-plan.md M10).
+pub fn setThreadPointer(base: u64) void {
+    io.wrmsr(ia32_fs_base, base);
+}
+
 // --- SMP: application-processor bring-up ----------------------------------
 
 /// Record the low (<1 MiB) frame reserved for the AP trampoline. Run once at boot.
diff --git a/system/kernel/architecture/x86_64/paging.zig b/system/kernel/architecture/x86_64/paging.zig
index b4b2213..5838185 100644
--- a/system/kernel/architecture/x86_64/paging.zig
+++ b/system/kernel/architecture/x86_64/paging.zig
@@ -509,7 +509,7 @@ pub fn unmapInto(pml4: u64, virtual: u64) void {
 /// any address space, not just the live one). Returns null if `virtual` is not
 /// mapped at any level. Stops at a 2 MiB huge-page leaf (the physmap uses them),
 /// resolving the offset within it. The foundation for cross-address-space copies
-/// and for munmap (which needs the frame behind a user vaddr to free it).
+/// and for munmap (which needs the frame behind a user virtual_address to free it).
 pub fn translateIn(pml4: u64, virtual: u64) ?u64 {
     const pml4e = tableAt(pml4)[(virtual >> 39) & 0x1FF];
     if (pml4e & present == 0) return null;
diff --git a/system/kernel/ipc-synchronous.zig b/system/kernel/ipc-synchronous.zig
index 70edee7..34476bf 100644
--- a/system/kernel/ipc-synchronous.zig
+++ b/system/kernel/ipc-synchronous.zig
@@ -355,7 +355,7 @@ pub fn replyWait(endpoint: *Endpoint, reply_ptr: u64, reply_len: u64, receive_pt
         me.ipc_client = null;
         const n = @min(reply_len, client.ipc_reply_cap);
         client.ipc_received_cap = abi.no_cap;
-        if (!copyAcross(me.aspace, reply_ptr, client.aspace, client.ipc_reply_ptr, n)) {
+        if (!copyAcross(me.address_space, reply_ptr, client.address_space, client.ipc_reply_ptr, n)) {
             client.ipc_status = -EFAULT;
         } else if (send_cap != abi.no_cap) {
             // Transfer the reply's capability into the client. A failure fails the
@@ -383,8 +383,8 @@ pub fn replyWait(endpoint: *Endpoint, reply_ptr: u64, reply_len: u64, receive_pt
         }
         if (popPost(endpoint)) |slot| {
             const n = @min(@as(usize, slot.length), receive_cap);
-            // Copy from the kernel-resident ring slot (source aspace 0) into the receiver.
-            if (!copyAcross(0, @intFromPtr(&slot.bytes), me.aspace, receive_ptr, n)) {
+            // Copy from the kernel-resident ring slot (source address_space 0) into the receiver.
+            if (!copyAcross(0, @intFromPtr(&slot.bytes), me.address_space, receive_ptr, n)) {
                 continue; // bad receive buffer: drop this message, keep serving
             }
             out_badge.* = slot.sender_id | notify_badge_bit | notify_message_bit;
@@ -392,7 +392,7 @@ pub fn replyWait(endpoint: *Endpoint, reply_ptr: u64, reply_len: u64, receive_pt
         }
         if (dequeueSender(endpoint)) |caller| {
             const n = @min(caller.ipc_send_len, receive_cap);
-            if (!copyAcross(caller.aspace, caller.ipc_send_ptr, me.aspace, receive_ptr, n)) {
+            if (!copyAcross(caller.address_space, caller.ipc_send_ptr, me.address_space, receive_ptr, n)) {
                 caller.ipc_status = -EFAULT; // bad sender buffer: fail it, keep serving
                 scheduler.readyLocked(caller);
                 continue;
diff --git a/system/kernel/process.zig b/system/kernel/process.zig
index 4fbe9c0..266bdfe 100644
--- a/system/kernel/process.zig
+++ b/system/kernel/process.zig
@@ -58,8 +58,9 @@ pub const stack_top_virtual: u64 = stack_base_virtual + parameters.user_stack_pa
 
 /// The mmap grant arena: where `mmap` hands out fresh user pages, above the image
 /// and stack but still inside PML4[224] (so no kernel mapping is widened). Each
-/// process bump-allocates from `heap_arena_base` upward via `Task.heap_next`; a
-/// 1 GiB window is far more than any user heap needs today.
+/// process bump-allocates from `heap_arena_base` upward via a per-address-space cursor
+/// (`scheduler.addressSpaceMmapNextPtr`, shared by its threads); a 1 GiB window is far more
+/// than any user heap needs today.
 pub const heap_arena_base: u64 = 0x0000_7000_1000_0000;
 pub const heap_arena_end: u64 = heap_arena_base + (1 << 30);
 
@@ -69,8 +70,8 @@ pub const user_half_end: u64 = 0x0000_8000_0000_0000;
 
 /// The MMIO-grant arena: where `mmio_map` places device windows, in PML4[226] —
 /// a user-exclusive region distinct from code/stack/heap (PML4[224]), so mapping
-/// device pages user-accessible widens no kernel mapping. Per-process cursor in
-/// `Task.device_map_next`.
+/// device pages user-accessible widens no kernel mapping. Per-address-space cursor
+/// (`scheduler.addressSpaceDeviceMapNextPtr`).
 pub const device_arena_base: u64 = 0x0000_7100_0000_0000;
 pub const device_arena_end: u64 = device_arena_base + (4 << 30);
 
@@ -163,7 +164,7 @@ fn fail(state: *architecture.CpuState) void {
 
 fn system_call(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    const user = t.aspace != 0;
+    const user = t.address_space != 0;
     if (user) {
         // A condemned process (process_kill caught it running) dies at its next
         // kernel entry — before it can spawn, claim, or message anything else.
@@ -230,6 +231,8 @@ fn system_call(state: *architecture.CpuState) void {
         .thread_spawn => systemThreadSpawn(state),
         .current_core => systemCurrentCore(state),
         .thread_self => systemThreadSelf(state),
+        .thread_join => systemThreadJoin(state),
+        .set_thread_pointer => systemSetThreadPointer(state),
         .futex_wait => systemFutexWait(state),
         .futex_wake => systemFutexWake(state),
         .thread_exit => {
@@ -314,7 +317,7 @@ fn systemIpcReplyWait(state: *architecture.CpuState) void {
 fn systemIpcSend(state: *architecture.CpuState) void {
     const me = scheduler.current();
     const endpoint = ipc.resolveHandle(me, architecture.systemCallArg(state, 0)) orelse return failErr(state, ipc.EBADF);
-    const r = ipc.send(endpoint, me.aspace, architecture.systemCallArg(state, 1), architecture.systemCallArg(state, 2), me.id);
+    const r = ipc.send(endpoint, me.address_space, architecture.systemCallArg(state, 1), architecture.systemCallArg(state, 2), me.id);
     architecture.setSystemCallResult(state, @bitCast(r));
 }
 
@@ -324,7 +327,7 @@ fn systemDeviceEnumerate(state: *architecture.CpuState) void {
     const buffer_ptr = architecture.systemCallArg(state, 0);
     const maximum = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0 or buffer_ptr >= user_half_end) return fail(state);
+    if (t.address_space == 0 or buffer_ptr >= user_half_end) return fail(state);
     const sz = @sizeOf(device_abi.DeviceDescriptor);
     const cap = @min(maximum, (user_half_end - buffer_ptr) / sz); // clamp to the user half
     const out: [*]device_abi.DeviceDescriptor = @ptrFromInt(buffer_ptr);
@@ -348,14 +351,14 @@ fn systemDeviceClaim(state: *architecture.CpuState) void {
     } else fail(state);
 }
 
-/// mmio_map(device_id, resource_index) -> vaddr: map a claimed device's MMIO window into
+/// mmio_map(device_id, resource_index) -> virtual_address: map a claimed device's MMIO window into
 /// this address space (strong-uncacheable) and return the register base address.
 /// The claim is the capability — a process can only map hardware it owns.
 fn systemMmioMap(state: *architecture.CpuState) void {
     const device_id = architecture.systemCallArg(state, 0);
     const resource_index = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     // Read the broker table under the lock: ring-3 device_register (M19) now
     // mutates it concurrently on other cores, so a lock-free read here could
     // see a torn resource (and a torn length used to panic the arithmetic
@@ -373,18 +376,23 @@ fn systemMmioMap(state: *architecture.CpuState) void {
     if (r.len == 0) return fail(state);
     if (@addWithOverflow(r.start, r.len)[1] != 0) return fail(state);
 
-    if (t.device_map_next == 0) t.device_map_next = device_arena_base;
     const first = r.start & ~@as(u64, page_size - 1);
     const last = (r.start + r.len - 1) & ~@as(u64, page_size - 1);
     const pages = (last - first) / page_size + 1;
-    const base_v = t.device_map_next;
-    if (base_v + pages * page_size > device_arena_end) return fail(state);
-
     // A framebuffer resource asks (via its flag) to be mapped write-combining rather
     // than the strong-uncacheable default that register MMIO needs.
     const write_combining = (r.flags & device_abi.resource_flag_write_combining) != 0;
-    architecture.mapUserDeviceInto(t.aspace, base_v, r.start, r.len, write_combining);
-    t.device_map_next = base_v + pages * page_size;
+
+    // Per-address-space cursor + shared page tables → serialize under the big lock,
+    // same as mmap (docs/threading-plan.md M7).
+    const flags = sync.enter();
+    defer sync.leave(flags);
+    const cursor = scheduler.addressSpaceDeviceMapNextPtr(t.address_space) orelse return fail(state);
+    if (cursor.* == 0) cursor.* = device_arena_base; // seed the arena lazily
+    const base_v = cursor.*;
+    if (base_v + pages * page_size > device_arena_end) return fail(state);
+    architecture.mapUserDeviceInto(t.address_space, base_v, r.start, r.len, write_combining);
+    cursor.* = base_v + pages * page_size;
     architecture.setSystemCallResult(state, base_v + (r.start & (page_size - 1))); // register base
 }
 
@@ -413,7 +421,7 @@ pub fn resolveIoPort(t: *scheduler.Task, device_id: u64, resource_index: u64, of
 /// is fine. See docs/drivers.md.
 fn systemIoRead(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const width = architecture.systemCallArg(state, 3);
     const port = resolveIoPort(t, architecture.systemCallArg(state, 0), architecture.systemCallArg(state, 1), architecture.systemCallArg(state, 2), width) orelse return fail(state);
     architecture.setSystemCallResult(state, architecture.pioRead(@intCast(width), port));
@@ -424,14 +432,14 @@ fn systemIoRead(state: *architecture.CpuState) void {
 /// gate as `io_read`.
 fn systemIoWrite(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const width = architecture.systemCallArg(state, 3);
     const port = resolveIoPort(t, architecture.systemCallArg(state, 0), architecture.systemCallArg(state, 1), architecture.systemCallArg(state, 2), width) orelse return fail(state);
     architecture.pioWrite(@intCast(width), port, @intCast(architecture.systemCallArg(state, 4)));
     architecture.setSystemCallResult(state, 0);
 }
 
-/// dma_alloc(len, flags) -> vaddr (rax), paddr (rdx): grant `len` bytes (rounded up to
+/// dma_alloc(len, flags) -> virtual_address (rax), physical_address (rdx): grant `len` bytes (rounded up to
 /// whole pages) of DMA-capable memory — physically contiguous, zeroed, pinned, and
 /// strong-uncacheable (coherent) — mapping it into the caller's DMA arena and handing
 /// back both the virtual address to touch and the physical address to program into the
@@ -443,7 +451,7 @@ fn systemDmaAlloc(state: *architecture.CpuState) void {
     const len = architecture.systemCallArg(state, 0);
     const flags = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0 or len == 0) return fail(state);
+    if (t.address_space == 0 or len == 0) return fail(state);
 
     const pages: usize = @intCast((len + page_size - 1) / page_size);
     const max_phys: u64 = if (flags & abi.dma_below_4g != 0) (@as(u64, 4) << 30) else ~@as(u64, 0);
@@ -459,14 +467,14 @@ fn systemDmaAlloc(state: *architecture.CpuState) void {
     // Zero through the physmap (the frames aren't mapped in the caller yet), then map.
     const kernel_view: [*]u8 = @ptrFromInt(boot_handoff.physicalToVirtual(phys));
     @memset(kernel_view[0 .. pages * page_size], 0);
-    architecture.mapUserDmaInto(t.aspace, base_v, phys, pages * page_size);
+    architecture.mapUserDmaInto(t.address_space, base_v, phys, pages * page_size);
 
     t.dma_map_next = base_v + pages * page_size;
     architecture.setSystemCallResult(state, base_v); // virtual address for the CPU
     architecture.setSystemCallResult2(state, phys); // physical address for the device
 }
 
-/// dma_free(vaddr, len) -> 0: release a prior `dma_alloc`. Bounded to the DMA arena so
+/// dma_free(virtual_address, len) -> 0: release a prior `dma_alloc`. Bounded to the DMA arena so
 /// it can never unmap-and-free the caller's stack, heap, or an MMIO grant; only pages
 /// actually mapped are freed (an unmapped hole is skipped). Teardown also reclaims any
 /// DMA pages left mapped at exit (they carry no `device_grant`, so `freeSubtree` frees
@@ -475,21 +483,21 @@ fn systemDmaFree(state: *architecture.CpuState) void {
     const base_v = architecture.systemCallArg(state, 0);
     const len = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const pages: usize = @intCast((len + page_size - 1) / page_size);
     if (base_v < dma_arena_base or base_v + pages * page_size > dma_arena_end) return fail(state);
 
     for (0..pages) |i| {
         const va = base_v + i * page_size;
-        if (architecture.translate(t.aspace, va)) |phys| {
-            architecture.unmapUserPageInto(t.aspace, va);
+        if (architecture.translate(t.address_space, va)) |phys| {
+            architecture.unmapUserPageInto(t.address_space, va);
             pmm.free(phys);
         }
     }
     architecture.setSystemCallResult(state, 0);
 }
 
-/// shm_create(len) -> vaddr (rax), handle (rdx): grant `len` bytes (rounded up to whole
+/// shm_create(len) -> virtual_address (rax), handle (rdx): grant `len` bytes (rounded up to whole
 /// pages) of **shareable, zeroed, cacheable** RAM — contiguous frames mapped into the
 /// caller's shm arena — and hand back the virtual address plus a capability handle. Unlike
 /// `dma_alloc` the memory is write-back cacheable (for CPU compositing, not device DMA) and
@@ -500,7 +508,7 @@ fn systemDmaFree(state: *architecture.CpuState) void {
 fn systemShmCreate(state: *architecture.CpuState) void {
     const len = architecture.systemCallArg(state, 0);
     const t = scheduler.current();
-    if (t.aspace == 0 or len == 0) return fail(state);
+    if (t.address_space == 0 or len == 0) return fail(state);
 
     const pages: usize = @intCast((len + page_size - 1) / page_size);
     if (pages == 0 or pages > maximum_shm_pages) return fail(state);
@@ -525,20 +533,20 @@ fn systemShmCreate(state: *architecture.CpuState) void {
         return fail(state);
     }
 
-    architecture.mapUserSharedInto(t.aspace, base_v, phys, pages * page_size);
+    architecture.mapUserSharedInto(t.address_space, base_v, phys, pages * page_size);
     t.shm_map_next = base_v + pages * page_size;
-    architecture.setSystemCallResult(state, base_v); // vaddr for the CPU
+    architecture.setSystemCallResult(state, base_v); // virtual_address for the CPU
     architecture.setSystemCallResult2(state, @intCast(handle)); // capability handle to pass on
 }
 
-/// shm_map(cap) -> vaddr: map the shared region named by a capability handle the caller
+/// shm_map(cap) -> virtual_address: map the shared region named by a capability handle the caller
 /// received (via an `ipc_call` send_cap) into its shm arena — the same physical frames the
 /// creator sees — returning the virtual address. The handle already holds a reference (taken
 /// when the capability was shared), so this only adds a mapping; it never bumps the refcount.
 fn systemShmMap(state: *architecture.CpuState) void {
     const cap = architecture.systemCallArg(state, 0);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
 
     const shm = ipc.resolveShm(t, cap) orelse return fail(state); // not an shm handle we hold
     if (t.shm_map_next == 0) t.shm_map_next = shm_arena_base;
@@ -546,12 +554,12 @@ fn systemShmMap(state: *architecture.CpuState) void {
     const size = shm.pages * page_size;
     if (base_v + size > shm_arena_end) return fail(state);
 
-    architecture.mapUserSharedInto(t.aspace, base_v, shm.phys, size);
+    architecture.mapUserSharedInto(t.address_space, base_v, shm.phys, size);
     t.shm_map_next = base_v + size;
     architecture.setSystemCallResult(state, base_v);
 }
 
-/// shm_physical(cap) -> paddr: the guest-physical base of a shared region the caller holds a
+/// shm_physical(cap) -> physical_address: the guest-physical base of a shared region the caller holds a
 /// capability for. The frames are contiguous (allocated by `allocContiguous`), so a single
 /// physical base + length describes the whole region — which is exactly what a driver needs
 /// to hand a shm surface to a device (virtio-gpu `attach_backing`). Only a holder of the
@@ -559,7 +567,7 @@ fn systemShmMap(state: *architecture.CpuState) void {
 fn systemShmPhysical(state: *architecture.CpuState) void {
     const cap = architecture.systemCallArg(state, 0);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const shm = ipc.resolveShm(t, cap) orelse return fail(state); // not an shm handle we hold
     architecture.setSystemCallResult(state, shm.phys);
 }
@@ -580,10 +588,10 @@ fn systemDeviceRegister(state: *architecture.CpuState) void {
     const parent_id = architecture.systemCallArg(state, 0);
     const descriptor_ptr = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
 
     var descriptor: device_abi.DeviceDescriptor = undefined;
-    if (!ipc.copyFromUser(t.aspace, descriptor_ptr, std.mem.asBytes(&descriptor))) return fail(state);
+    if (!ipc.copyFromUser(t.address_space, descriptor_ptr, std.mem.asBytes(&descriptor))) return fail(state);
 
     // Under the big kernel lock: the broker's table is also mutated by the
     // death sweep (releaseAllOwnedBy) and read by enumerate on other cores —
@@ -667,7 +675,7 @@ fn systemThreadSpawn(state: *architecture.CpuState) void {
     const arg = architecture.systemCallArg(state, 2);
     const exit_handle = architecture.systemCallArg(state, 3);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state); // kernel tasks own no address space to share
+    if (t.address_space == 0) return fail(state); // kernel tasks own no address space to share
     if (entry == 0 or entry >= user_half_end) return fail(state);
     if (stack_top == 0 or stack_top > user_half_end) return fail(state);
     // The endpoint the thread notifies on exit (how join waits), or none.
@@ -675,17 +683,17 @@ fn systemThreadSpawn(state: *architecture.CpuState) void {
         null
     else
         ipc.resolveHandle(t, exit_handle) orelse return failErr(state, ipc.EBADF);
-    const tid = spawnThreadSupervised(t.aspace, entry, stack_top, arg, t.priority, t.id, exit_endpoint) orelse return fail(state);
+    const tid = spawnThreadSupervised(t.address_space, entry, stack_top, arg, t.priority, t.id, exit_endpoint) orelse return fail(state);
     architecture.setSystemCallResult(state, tid);
 }
 
-/// Spawn a thread sharing `aspace`, taking the exit-endpoint reference under the **same**
+/// Spawn a thread sharing `address_space`, taking the exit-endpoint reference under the **same**
 /// lock as the spawn (as `spawnProcessSupervised` does), so the thread cannot die before
 /// its reference exists. Returns the new thread id, or null on resource exhaustion.
-fn spawnThreadSupervised(aspace: u64, entry: u64, stack_top: u64, arg: u64, priority: scheduler.Priority, supervisor: u32, exit_endpoint: ?*ipc.Endpoint) ?u32 {
+fn spawnThreadSupervised(address_space: u64, entry: u64, stack_top: u64, arg: u64, priority: scheduler.Priority, supervisor: u32, exit_endpoint: ?*ipc.Endpoint) ?u32 {
     const flags = sync.enter();
     defer sync.leave(flags);
-    const tid = scheduler.spawnUserLocked(aspace, entry, stack_top, arg, priority, "thread", supervisor, if (exit_endpoint) |e| @ptrCast(e) else null) orelse return null;
+    const tid = scheduler.spawnUserLocked(address_space, entry, stack_top, arg, priority, "thread", supervisor, if (exit_endpoint) |e| @ptrCast(e) else null) orelse return null;
     if (exit_endpoint) |endpoint| endpoint.refcount += 1; // the thread holds it birth-to-death
     return tid;
 }
@@ -700,6 +708,34 @@ fn systemThreadSelf(state: *architecture.CpuState) void {
     architecture.setSystemCallResult(state, scheduler.currentId());
 }
 
+/// set_thread_pointer(addr) -> 0: set the caller's user-space TLS thread pointer. The
+/// arch layer maps it to IA32_FS_BASE on x86_64, `TPIDR_EL0` on aarch64; the kernel
+/// never reads it, and the scheduler restores it per task across context switches
+/// (docs/threading-plan.md M10). `addr` must be a user-half address.
+fn systemSetThreadPointer(state: *architecture.CpuState) void {
+    const addr = architecture.systemCallArg(state, 0);
+    const t = scheduler.current();
+    if (t.address_space == 0) return fail(state); // kernel tasks have no user TLS
+    if (addr >= user_half_end) return fail(state);
+    const flags = sync.enter();
+    scheduler.setThreadPointerLocked(addr);
+    sync.leave(flags);
+    architecture.setSystemCallResult(state, 0);
+}
+
+/// thread_join(tid) -> 0: block until the thread with id `tid` has exited (docs/threading-
+/// plan.md M9). Needs no per-thread IPC endpoint. The compare-and-block is one critical
+/// section, so an exit cannot slip between "is it alive?" and the block.
+fn systemThreadJoin(state: *architecture.CpuState) void {
+    const tid: u32 = @truncate(architecture.systemCallArg(state, 0));
+    const t = scheduler.current();
+    if (t.address_space == 0) return fail(state); // kernel tasks don't join
+    const flags = sync.enter();
+    scheduler.joinThreadLocked(tid);
+    sync.leave(flags);
+    architecture.setSystemCallResult(state, 0);
+}
+
 /// futex_wait(addr, expected, timeout_ns) -> status (docs/threading.md): if the 4-byte
 /// user word at `addr` still equals `expected`, block until a futex_wake on `addr` or
 /// (if timeout_ns > 0) the deadline. The compare and the block are one critical section,
@@ -710,12 +746,12 @@ fn systemFutexWait(state: *architecture.CpuState) void {
     const expected: u32 = @truncate(architecture.systemCallArg(state, 1));
     const timeout_ns = architecture.systemCallArg(state, 2);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     if (addr == 0 or (addr & 3) != 0 or addr + 4 > user_half_end) return fail(state);
 
     const flags = sync.enter();
     var word_bytes: [4]u8 = undefined;
-    if (!ipc.copyFromUser(t.aspace, addr, &word_bytes)) {
+    if (!ipc.copyFromUser(t.address_space, addr, &word_bytes)) {
         sync.leave(flags);
         return fail(state);
     }
@@ -739,10 +775,10 @@ fn systemFutexWake(state: *architecture.CpuState) void {
     const addr = architecture.systemCallArg(state, 0);
     const count: u32 = @truncate(architecture.systemCallArg(state, 1));
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     if (addr == 0 or (addr & 3) != 0 or addr + 4 > user_half_end) return fail(state);
     const flags = sync.enter();
-    const woken = scheduler.futexWakeLocked(t.aspace, addr, count);
+    const woken = scheduler.futexWakeLocked(t.address_space, addr, count);
     sync.leave(flags);
     architecture.setSystemCallResult(state, woken);
 }
@@ -757,7 +793,7 @@ fn systemProcessEnumerate(state: *architecture.CpuState) void {
     const buffer_ptr = architecture.systemCallArg(state, 0);
     const maximum = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0 or buffer_ptr >= user_half_end) return fail(state);
+    if (t.address_space == 0 or buffer_ptr >= user_half_end) return fail(state);
     const sz = @sizeOf(abi.ProcessDescriptor);
     const cap = @min(maximum, (user_half_end - buffer_ptr) / sz); // clamp to the user half
     const out: [*]abi.ProcessDescriptor = @ptrFromInt(buffer_ptr);
@@ -770,7 +806,7 @@ fn systemProcessEnumerate(state: *architecture.CpuState) void {
 /// cannot be a weapon (ids are never reused, so a stale one just misses).
 fn systemProcessKill(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const id = architecture.systemCallArg(state, 0);
     if (id > std.math.maxInt(u32)) return failErr(state, ipc.ESRCH);
     const r = killProcess(t.id, @intCast(id));
@@ -903,7 +939,7 @@ pub fn killProcess(caller_id: u32, target_id: u32) i64 {
     const flags = sync.enter();
     defer sync.leave(flags);
     const target = scheduler.taskByIdLocked(target_id) orelse return -ipc.ESRCH;
-    if (target.aspace == 0) return -ipc.ESRCH; // kernel tasks are not processes
+    if (target.address_space == 0) return -ipc.ESRCH; // kernel tasks are not processes
     if (target.supervisor != caller_id) return -ipc.EPERM;
     target.exit_reason = .killed;
     if (target.state == .running) {
@@ -973,7 +1009,7 @@ var exit_subscribers: [exit_subscriber_capacity]?ExitSubscriber = .{null} ** exi
 /// secret between cooperating processes. -ENOSPC when the table is full.
 fn systemProcessSubscribe(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 0)) orelse return failErr(state, ipc.EBADF);
     const flags = sync.enter();
     defer sync.leave(flags);
@@ -993,7 +1029,7 @@ fn systemProcessSubscribe(state: *architecture.CpuState) void {
 /// delivered immediately on bind, coalesced into one notification.
 fn systemSignalBind(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 0)) orelse return failErr(state, ipc.EBADF);
     const flags = sync.enter();
     defer sync.leave(flags);
@@ -1013,7 +1049,7 @@ fn systemSignalBind(state: *architecture.CpuState) void {
 /// targets accumulate the signal in their pending mask.
 fn systemProcessSignal(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const id = architecture.systemCallArg(state, 0);
     const signal = architecture.systemCallArg(state, 1);
     if (id > std.math.maxInt(u32)) return failErr(state, ipc.ESRCH);
@@ -1021,7 +1057,7 @@ fn systemProcessSignal(state: *architecture.CpuState) void {
     const flags = sync.enter();
     defer sync.leave(flags);
     const target = scheduler.taskByIdLocked(@intCast(id)) orelse return failErr(state, ipc.ESRCH);
-    if (target.aspace == 0) return failErr(state, ipc.ESRCH);
+    if (target.address_space == 0) return failErr(state, ipc.ESRCH);
     if (target.supervisor != t.id and target.id != t.id) return failErr(state, ipc.EPERM);
     target.pending_signals |= @as(u32, 1) << @intCast(signal);
     if (target.signal_endpoint) |raw| {
@@ -1058,7 +1094,7 @@ fn timerSweepLocked() void {
 /// timer_bind(endpoint, ms): arm a one-shot timer. -ENOSPC when the table is full.
 fn systemTimerBind(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 0)) orelse return failErr(state, ipc.EBADF);
     const ms = architecture.systemCallArg(state, 1);
     const flags = sync.enter();
@@ -1075,7 +1111,7 @@ fn systemTimerBind(state: *architecture.CpuState) void {
 
 fn systemProcessExitReason(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const id = architecture.systemCallArg(state, 0);
     if (id > std.math.maxInt(u32)) return failErr(state, ipc.ESRCH);
     const r = exitReasonOf(t.id, @intCast(id));
@@ -1101,7 +1137,7 @@ fn ownedGsi(t: *scheduler.Task, device_id: u64, resource_index: u64) ?u32 {
 /// IPC_ReplyWait and is woken by the ISR; see system/kernel/irq.zig for the cycle.
 fn systemIrqBind(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const gsi = ownedGsi(t, architecture.systemCallArg(state, 0), architecture.systemCallArg(state, 1)) orelse
         return fail(state);
     const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 2)) orelse return fail(state);
@@ -1121,7 +1157,7 @@ fn systemIrqBind(state: *architecture.CpuState) void {
 fn systemMsiBind(state: *architecture.CpuState) void {
     const device_id = architecture.systemCallArg(state, 0);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const owner = devices_broker.ownerOf(device_id) orelse return fail(state);
     if (owner != t.id) return fail(state); // not claimed by this process
     const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 1)) orelse return failErr(state, ipc.EBADF);
@@ -1140,7 +1176,7 @@ fn systemMsiBind(state: *architecture.CpuState) void {
 /// more arrives until the driver says it has serviced the hardware.
 fn systemIrqAck(state: *architecture.CpuState) void {
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state);
+    if (t.address_space == 0) return fail(state);
     const gsi = ownedGsi(t, architecture.systemCallArg(state, 0), architecture.systemCallArg(state, 1)) orelse
         return fail(state);
 
@@ -1228,37 +1264,52 @@ fn systemKlogRead(state: *architecture.CpuState) void {
 fn systemMmap(state: *architecture.CpuState) void {
     const len = architecture.systemCallArg(state, 0);
     const t = scheduler.current();
-    if (t.aspace == 0) return fail(state); // not a user process — nothing to map into
+    if (t.address_space == 0) return fail(state); // not a user process — nothing to map into
     const pages = (len + page_size - 1) / page_size;
     if (pages == 0 or pages > maximum_mmap_pages) return fail(state);
 
-    if (t.heap_next == 0) t.heap_next = heap_arena_base; // seed the arena lazily
-    const base = t.heap_next;
-    if (base + pages * page_size > heap_arena_end) return fail(state); // arena exhausted
+    // Reserve a disjoint range under a *brief* lock (the cursor is shared by every thread
+    // in this address space). The mapping below then takes the lock **per page**, not for
+    // the whole grant: the big lock is held with interrupts disabled, so pinning it across
+    // a multi-MiB memset+map would freeze every other core on its next tick — which timed
+    // the `affinity` scenario out (docs/threading-plan.md M7).
+    const base = reserve: {
+        const flags = sync.enter();
+        defer sync.leave(flags);
+        const cursor = scheduler.addressSpaceMmapNextPtr(t.address_space) orelse return fail(state);
+        if (cursor.* == 0) cursor.* = heap_arena_base; // seed the arena lazily
+        const b = cursor.*;
+        if (b + pages * page_size > heap_arena_end) return fail(state); // arena exhausted
+        cursor.* = b + pages * page_size; // reserve now, so concurrent grants can't overlap
+        break :reserve b;
+    };
 
-    // Map page by page. On mid-way frame exhaustion, roll back the pages already mapped
-    // (unmap + free) so no partial grant leaks into the address space — the same
-    // all-or-nothing guarantee as before, but without a fixed scratch array, so the
-    // per-call size can be a multi-MiB framebuffer.
+    // Map the reserved range page by page, each page under a short-held lock (the range is
+    // already reserved, so pages can't overlap another thread's; the lock only serializes
+    // the shared page-table walk). On mid-way frame exhaustion, roll back the mapped pages
+    // so no partial grant leaks — the reserved-but-unmapped tail of the arena is left
+    // fallow (a rare, bounded address-space leak, not a memory leak).
     var mapped: usize = 0;
     while (mapped < pages) : (mapped += 1) {
+        const flags = sync.enter();
         const frame = pmm.alloc() orelse {
             var i: usize = 0;
             while (i < mapped) : (i += 1) {
                 const va = base + i * page_size;
-                if (architecture.translate(t.aspace, va)) |physical| {
-                    architecture.unmapUserPageInto(t.aspace, va);
+                if (architecture.translate(t.address_space, va)) |physical| {
+                    architecture.unmapUserPageInto(t.address_space, va);
                     pmm.free(physical);
                 }
             }
+            sync.leave(flags);
             return fail(state);
         };
         const destination: [*]u8 = @ptrFromInt(boot_handoff.physicalToVirtual(frame));
         @memset(destination[0..page_size], 0); // hand out zeroed memory
-        architecture.mapUserPageInto(t.aspace, base + mapped * page_size, frame, true, false); // RW + NX
+        architecture.mapUserPageInto(t.address_space, base + mapped * page_size, frame, true, false); // RW + NX
+        sync.leave(flags);
     }
-    t.heap_next = base + pages * page_size;
-    architecture.setSystemCallResult(state, base);
+    architecture.setSystemCallResult(state, base); // the cursor was already advanced at reserve
 }
 
 /// munmap(base, len): release a range previously handed out by `mmap`. Unmaps
@@ -1270,14 +1321,14 @@ fn systemMunmap(state: *architecture.CpuState) void {
     const base = architecture.systemCallArg(state, 0);
     const len = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
-    if (t.aspace == 0 or base % page_size != 0) return fail(state);
+    if (t.address_space == 0 or base % page_size != 0) return fail(state);
     const pages = (len + page_size - 1) / page_size;
     if (base < heap_arena_base or base + pages * page_size > heap_arena_end) return fail(state);
 
     for (0..pages) |i| {
         const va = base + i * page_size;
-        if (architecture.translate(t.aspace, va)) |physical| {
-            architecture.unmapUserPageInto(t.aspace, va);
+        if (architecture.translate(t.address_space, va)) |physical| {
+            architecture.unmapUserPageInto(t.address_space, va);
             pmm.free(physical);
         }
     }
@@ -1341,7 +1392,7 @@ const maximum_segments = 16;
 const maximum_pages = 256; // 1 MiB loader budget; the user region caps at 2 MiB anyway
 
 const Segment = struct {
-    vaddr: u64,
+    virtual_address: u64,
     memsz: u64,
     filesz: u64,
     off: u64,
@@ -1389,7 +1440,7 @@ fn parseSegments(image: []const u8, segs: *[maximum_segments]Segment) InitError!
         if (w and x) return error.BadSegment; // W^X, even for init
 
         const seg = Segment{
-            .vaddr = phdr.p_vaddr,
+            .virtual_address = phdr.p_vaddr,
             .memsz = phdr.p_memsz,
             .filesz = phdr.p_filesz,
             .off = phdr.p_offset,
@@ -1398,9 +1449,9 @@ fn parseSegments(image: []const u8, segs: *[maximum_segments]Segment) InitError!
         };
         // No overlap with any earlier segment (page-granular, since mapping is).
         for (segs[0..count]) |other| {
-            const a_end = seg.vaddr + seg.pages() * page_size;
-            const b_end = other.vaddr + other.pages() * page_size;
-            if (seg.vaddr < b_end and other.vaddr < a_end) return error.BadSegment;
+            const a_end = seg.virtual_address + seg.pages() * page_size;
+            const b_end = other.virtual_address + other.pages() * page_size;
+            if (seg.virtual_address < b_end and other.virtual_address < a_end) return error.BadSegment;
         }
         total_pages += seg.pages();
         if (total_pages > maximum_pages) return error.ProgramTooBig;
@@ -1411,17 +1462,17 @@ fn parseSegments(image: []const u8, segs: *[maximum_segments]Segment) InitError!
 
     // The entry point must land inside an executable segment.
     for (segs[0..count]) |seg| {
-        if (seg.executable and ehdr.e_entry >= seg.vaddr and ehdr.e_entry < seg.vaddr + seg.memsz)
+        if (seg.executable and ehdr.e_entry >= seg.virtual_address and ehdr.e_entry < seg.virtual_address + seg.memsz)
             return .{ .count = count, .entry = ehdr.e_entry };
     }
     return error.BadEntry;
 }
 
-/// Load one page of a segment into address space `aspace`: a fresh frame, zeroed
+/// Load one page of a segment into address space `address_space`: a fresh frame, zeroed
 /// and filled through the physmap, mapped user-accessible with the segment's W^X.
 /// On a later failure the whole address space is torn down, which frees every
 /// frame mapped into it — so no per-page rollback list is needed here.
-fn loadPageInto(aspace: u64, image: []const u8, seg: Segment, page_index: u64) InitError!void {
+fn loadPageInto(address_space: u64, image: []const u8, seg: Segment, page_index: u64) InitError!void {
     const frame = pmm.alloc() orelse return error.OutOfMemory;
     const destination: [*]u8 = @ptrFromInt(boot_handoff.physicalToVirtual(frame));
     @memset(destination[0..page_size], 0);
@@ -1430,7 +1481,7 @@ fn loadPageInto(aspace: u64, image: []const u8, seg: Segment, page_index: u64) I
         const n = @min(page_size, seg.filesz - page_off);
         @memcpy(destination[0..n], image[seg.off + page_off ..][0..n]);
     }
-    architecture.mapUserPageInto(aspace, seg.vaddr + page_off, frame, seg.writable, seg.executable);
+    architecture.mapUserPageInto(address_space, seg.virtual_address + page_off, frame, seg.writable, seg.executable);
 }
 
 /// Build the System V AMD64 process-entry block at the top of a process's stack
@@ -1517,11 +1568,11 @@ pub fn spawnProcessSupervised(image: []const u8, priority: u3, argv: []const []c
     const flags = sync.enter();
     defer sync.leave(flags);
 
-    const aspace = architecture.createAddressSpace() orelse return error.OutOfMemory;
-    errdefer architecture.destroyAddressSpace(aspace);
+    const address_space = architecture.createAddressSpace() orelse return error.OutOfMemory;
+    errdefer architecture.destroyAddressSpace(address_space);
 
     for (segs[0..parsed.count]) |seg| {
-        for (0..seg.pages()) |i| try loadPageInto(aspace, image, seg, i);
+        for (0..seg.pages()) |i| try loadPageInto(address_space, image, seg, i);
     }
 
     // The stack: `user_stack_pages` zeroed pages below stack_top_virtual, RW + NX.
@@ -1535,10 +1586,10 @@ pub fn spawnProcessSupervised(image: []const u8, priority: u3, argv: []const []c
         const page_virtual = stack_base_virtual + i * page_size;
         if (i == parameters.user_stack_pages - 1)
             user_sp = buildEntryStack(stack_page, page_virtual, argv);
-        architecture.mapUserPageInto(aspace, page_virtual, stack_frame, true, false); // RW + NX
+        architecture.mapUserPageInto(address_space, page_virtual, stack_frame, true, false); // RW + NX
     }
 
-    const child = scheduler.spawnUserLocked(aspace, parsed.entry, user_sp, 0, priority, argv[0], supervisor, if (exit_endpoint) |endpoint| @ptrCast(endpoint) else null) orelse
+    const child = scheduler.spawnUserLocked(address_space, parsed.entry, user_sp, 0, priority, argv[0], supervisor, if (exit_endpoint) |endpoint| @ptrCast(endpoint) else null) orelse
         return error.OutOfMemory;
     // The child holds a reference to its exit endpoint from birth to death. Taken
     // only now, after nothing can fail; the lock is still held, so the child
diff --git a/system/kernel/scheduler.zig b/system/kernel/scheduler.zig
index 4a91ef8..b6ede31 100644
--- a/system/kernel/scheduler.zig
+++ b/system/kernel/scheduler.zig
@@ -31,7 +31,10 @@ const number_priorities = 8;
 const stack_size = parameters.kernel_stack_size; // each task's kernel stack
 const maximum_tasks = parameters.maximum_tasks; // maximum tasks alive at once (static pool)
 
-const State = enum { free, ready, running, blocked };
+// `reaping` = the task has exited and is queued on its core's reap list; its slot must not
+// be reused (freeSlot skips it) until the reaper has freed its kernel stack and set it
+// `free` (docs/threading-plan.md M8/M9).
+const State = enum { free, ready, running, blocked, reaping };
 
 pub const Task = struct {
     id: u32 = 0,
@@ -75,21 +78,25 @@ pub const Task = struct {
     ipc_wait_endpoint: ?*anyopaque = null,
     // Physical root of this task's address space, or 0 for a kernel task (which
     // runs on the shared kernel page tables). A user task carries its own.
-    aspace: u64 = 0,
+    address_space: u64 = 0,
     user_ip: u64 = 0, // user-mode entry point (user task only)
     user_sp: u64 = 0, // user-mode stack pointer (user task only)
-    user_arg: u64 = 0, // value delivered in the user's rdi at first entry: 0 for a
-    // process (its _start ignores it), the closure pointer for a thread (docs/threading.md)
+    user_arg: u64 = 0, // value delivered in the user's first argument register at first entry
+    // (rdi on x86_64, via architecture.jumpToUserArg): 0 for a process (its _start ignores
+    // it), the closure pointer for a thread (docs/threading.md)
     // The user address this task is blocked on in futex_wait (0 = not futex-waiting).
     // Cleared to 0 by futexWakeLocked as the "woken, not timed out" signal (docs/threading.md).
     futex_addr: u64 = 0,
-    // Next free virtual address in this task's mmap grant arena (0 = uninitialised;
-    // process.zig lazily seeds it to the arena base on the first mmap). Bumped up
-    // as the user heap grows; user task only.
-    heap_next: u64 = 0,
-    // Next free virtual address in this task's MMIO-grant arena (PML4[226]; 0 =
-    // uninitialised, process.zig seeds it on the first mmio_map). User task only.
-    device_map_next: u64 = 0,
+    // The task id this task is blocked in `thread_join` on (0 = not joining). Woken by
+    // `wakeJoinersLocked` when that task exits (docs/threading-plan.md M9).
+    join_target: u32 = 0,
+    // This task's user-space TLS thread pointer — 0 until set via `set_thread_pointer`.
+    // Architecture-neutral: the arch layer maps it to the FS base on x86_64, `TPIDR_EL0` on
+    // aarch64. Restored on every context switch to this task (docs/threading-plan.md M10).
+    thread_pointer: u64 = 0,
+    // The mmap / MMIO grant-arena cursors moved from Task to the per-address-space object
+    // (`AddressSpaceRef`, below) so threads sharing one address space hand out disjoint grants
+    // — see addressSpaceMmapNextPtr / addressSpaceDeviceMapNextPtr (docs/threading-plan.md M7).
     // --- synchronous IPC (ipc_sync.zig) ---
     // Per-process handle table: a small-int handle names a kernel capability object.
     // Each entry tags its `kind` (an IPC endpoint or a shared-memory object) so the
@@ -100,9 +107,9 @@ pub const Task = struct {
     // receive, cleared when it replies). A client, while blocked in Call, records
     // its message + reply buffers here and its result lands in `ipc_status`.
     ipc_client: ?*Task = null,
-    ipc_send_ptr: u64 = 0, // client: outgoing message (vaddr in this task's AS)
+    ipc_send_ptr: u64 = 0, // client: outgoing message (virtual_address in this task's address space)
     ipc_send_len: u64 = 0,
-    ipc_reply_ptr: u64 = 0, // client: reply buffer (vaddr)
+    ipc_reply_ptr: u64 = 0, // client: reply buffer (virtual_address)
     ipc_reply_cap: u64 = 0,
     ipc_status: i64 = 0, // client: reply length / -errno, written by the replier
     dma_map_next: u64 = 0, // bump pointer into this task's DMA arena (0 = unseeded)
@@ -148,17 +155,57 @@ var tasks = [_]Task{.{}} ** maximum_tasks;
 /// only when the **last** task on an address space exits. All access is under the big
 /// kernel lock. There can be no more live address spaces than tasks, so the table is
 /// sized to the task pool and never overflows in practice.
-const AspaceRef = struct { root: u64 = 0, count: u32 = 0 };
-var aspace_refs = [_]AspaceRef{.{}} ** maximum_tasks;
-var aspace_destroy_count: u64 = 0;
+// The per-address-space kernel object: a reference count plus the grant-arena cursors.
+// One live entry per address space; threads sharing an address space share this entry,
+// so their mmap/mmio grants bump one cursor and never overlap (docs/threading-plan.md M7).
+// `mmap_next`/`device_map_next` are 0 until process.zig seeds them to the arena base.
+const AddressSpaceRef = struct { root: u64 = 0, count: u32 = 0, mmap_next: u64 = 0, device_map_next: u64 = 0 };
+var address_space_refs = [_]AddressSpaceRef{.{}} ** maximum_tasks;
+var address_space_destroy_count: u64 = 0;
+
+/// Total bytes of task **kernel** stacks currently allocated from the kernel heap —
+/// incremented when a task is created, decremented when the reaper frees a dead task's
+/// stack. A test-observable proof that the reaper reclaims every stack (docs/threading-
+/// plan.md M8): with no live tasks beyond the baseline, this returns to its baseline.
+var live_stack_bytes: usize = 0;
+
+/// Test-observable: bytes of task kernel stacks currently live (see `live_stack_bytes`).
+pub fn liveStackBytes() usize {
+    return live_stack_bytes;
+}
+
+/// Free a dead task's kernel stack and drop it from `live_stack_bytes`. The task must be
+/// off that stack already (killed while not running, or reaped after it switched away).
+/// Caller holds the kernel lock.
+fn reapStackLocked(t: *Task) void {
+    if (t.stack.len == 0) return; // boot/idle tasks run on a static stack — nothing to free
+    live_stack_bytes -= t.stack.len;
+    heap.allocator().free(t.stack);
+    t.stack = &.{};
+    t.kstack_top = 0;
+}
+
+/// Free every `.reaping` task queued on this core's reap list and mark each `.free` (now
+/// its slot may be reused). The tasks are all off their stacks (they switched away), and
+/// the caller holds the lock, so freeing is safe (docs/threading-plan.md M8/M9).
+fn drainReapListLocked(pc: *PerCpu) void {
+    var node = pc.reap_list;
+    pc.reap_list = null;
+    while (node) |t| {
+        node = t.next; // save the link before we clear it
+        t.next = null;
+        reapStackLocked(t);
+        t.state = .free; // reusable only now, after the stack is freed
+    }
+}
 
 /// Take a reference to address space `root` (0 = a kernel task, which owns none).
 /// Returns false only if the ref table is full — bounded by `maximum_tasks`, so in
 /// practice it never is. Caller holds the kernel lock.
-fn retainAspace(root: u64) bool {
+fn retainAddressSpace(root: u64) bool {
     if (root == 0) return true;
-    var free: ?*AspaceRef = null;
-    for (&aspace_refs) |*entry| {
+    var free: ?*AddressSpaceRef = null;
+    for (&address_space_refs) |*entry| {
         if (entry.count != 0 and entry.root == root) {
             entry.count += 1;
             return true;
@@ -173,34 +220,54 @@ fn retainAspace(root: u64) bool {
 /// Drop a reference to `root`; destroy the address space when the **last** one drops.
 /// A `root` with no entry — never retained, e.g. a hand-built test space — is
 /// destroyed directly, preserving the pre-refcount behaviour. Caller holds the lock.
-fn releaseAspace(root: u64) void {
+fn releaseAddressSpace(root: u64) void {
     if (root == 0) return;
-    for (&aspace_refs) |*entry| {
+    for (&address_space_refs) |*entry| {
         if (entry.count == 0 or entry.root != root) continue;
         entry.count -= 1;
         if (entry.count == 0) {
             entry.root = 0;
             architecture.destroyAddressSpace(root);
-            aspace_destroy_count += 1;
+            address_space_destroy_count += 1;
         }
         return;
     }
     architecture.destroyAddressSpace(root);
-    aspace_destroy_count += 1;
+    address_space_destroy_count += 1;
 }
 
 /// Test-observable: how many address spaces are live (entries with a nonzero count).
-pub fn liveAspaceCount() u32 {
+pub fn liveAddressSpaceCount() u32 {
     var live: u32 = 0;
-    for (&aspace_refs) |*entry| {
+    for (&address_space_refs) |*entry| {
         if (entry.count != 0) live += 1;
     }
     return live;
 }
 
 /// Test-observable: total address-space destructions since boot.
-pub fn aspaceDestroyCount() u64 {
-    return aspace_destroy_count;
+pub fn addressSpaceDestroyCount() u64 {
+    return address_space_destroy_count;
+}
+
+/// Pointer to the mmap grant-arena cursor for address space `root`, so the mmap syscall
+/// can read-and-bump it. Per-address-space (not per-task), so sibling threads get
+/// disjoint grants. **Caller holds the kernel lock** (the entry is stable while held).
+/// Null only if `root` was never retained — which can't happen for a live user task.
+pub fn addressSpaceMmapNextPtr(root: u64) ?*u64 {
+    for (&address_space_refs) |*entry| {
+        if (entry.count != 0 and entry.root == root) return &entry.mmap_next;
+    }
+    return null;
+}
+
+/// Pointer to the MMIO grant-arena cursor for address space `root` (see
+/// `addressSpaceMmapNextPtr`). Caller holds the kernel lock.
+pub fn addressSpaceDeviceMapNextPtr(root: u64) ?*u64 {
+    for (&address_space_refs) |*entry| {
+        if (entry.count != 0 and entry.root == root) return &entry.device_map_next;
+    }
+    return null;
 }
 var next_id: u32 = 1;
 
@@ -221,11 +288,18 @@ pub const PerCpu = struct {
     hw_id: u32 = 0, // the core's hardware id (Local APIC id on x86_64)
     index: u32 = 0, // dense 0-based core index
     online: bool = false, // has this core finished bring-up?
-    loaded_aspace: u64 = 0, // the address-space root currently loaded on this core
+    loaded_address_space: u64 = 0, // the address-space root currently loaded on this core
+    loaded_thread_pointer: u64 = 0, // the TLS thread pointer currently loaded on this core (docs/threading-plan.md M10)
     // Tasks pinned to this core (affinity == index), per priority level + bitmap.
     pinned_head: [number_priorities]?*Task = .{null} ** number_priorities,
     pinned_tail: [number_priorities]?*Task = .{null} ** number_priorities,
     pinned_bitmap: u8 = 0,
+    // Tasks that ended while running on THIS core: they could not free the kernel stack
+    // they were standing on, so each pushed itself onto this list (`.reaping` state, linked
+    // via `Task.next`) and switched away. The next task to run on this core — or the timer
+    // tick — frees their stacks from its own stack, safely (docs/threading-plan.md M8). A
+    // *list* (not one slot) so a second death before the first is drained can't lose it.
+    reap_list: ?*Task = null,
 };
 
 const maximum_cpus = parameters.maximum_cpus;
@@ -257,7 +331,7 @@ var preemption_enabled = true;
 /// boot, before interrupts are enabled — so no lock is needed here.
 pub fn init(boot_priority: Priority) void {
     const pc = &cpus[0];
-    pc.* = .{ .index = 0, .online = true, .loaded_aspace = architecture.kernelPageTable() };
+    pc.* = .{ .index = 0, .online = true, .loaded_address_space = architecture.kernelPageTable() };
     architecture.setCpuLocal(0, @intFromPtr(pc));
     tasks[0] = .{ .id = 0, .state = .running, .priority = boot_priority };
     pc.current = &tasks[0];
@@ -296,7 +370,7 @@ pub fn secondaryMain() callconv(.c) noreturn {
     pc.current = t;
     pc.idle = t;
     pc.online = true;
-    pc.loaded_aspace = architecture.kernelPageTable(); // the AP adopted the kernel tables at bring-up
+    pc.loaded_address_space = architecture.kernelPageTable(); // the AP adopted the kernel tables at bring-up
     sync.leave(flags);
 
     architecture.enableInterrupts(); // the timer now preempts this idle context into work
@@ -382,7 +456,7 @@ pub fn spawnOn(entry: *const fn () void, priority: Priority, cpu: u32) bool {
     return ok;
 }
 
-/// Spawn a **user** task: a task with its own address space (`aspace`) that starts
+/// Spawn a **user** task: a task with its own address space (`address_space`) that starts
 /// in user mode at `entry` on `user_sp`, recorded under `name` (its argv[0]).
 /// `supervisor` is the id of the spawning process (0 = the kernel) — the kill
 /// authority — and `exit_endpoint` (an *ipc.Endpoint whose reference the caller
@@ -391,23 +465,24 @@ pub fn spawnOn(entry: *const fn () void, priority: Priority, cpu: u32) bool {
 /// lands in `user_task_trampoline`.
 /// Returns the new process id, or null (creating nothing) if the table is full or
 /// out of memory.
-/// **Caller must hold the kernel lock** (the loader that builds `aspace` holds it
+/// **Caller must hold the kernel lock** (the loader that builds `address_space` holds it
 /// across the whole spawn, so the address space and the task appear atomically).
-pub fn spawnUserLocked(aspace: u64, entry: u64, user_sp: u64, user_arg: u64, priority: Priority, task_name: []const u8, supervisor: u32, exit_endpoint: ?*anyopaque) ?u32 {
+pub fn spawnUserLocked(address_space: u64, entry: u64, user_sp: u64, user_arg: u64, priority: Priority, task_name: []const u8, supervisor: u32, exit_endpoint: ?*anyopaque) ?u32 {
     const t = freeSlot() orelse return null;
     const stack = heap.allocator().alloc(u8, stack_size) catch return null;
     // Take this task's reference to the address space before we commit the slot, so a
-    // failure here leaves nothing to unwind (the caller still owns the raw `aspace`).
-    if (!retainAspace(aspace)) {
+    // failure here leaves nothing to unwind (the caller still owns the raw `address_space`).
+    if (!retainAddressSpace(address_space)) {
         heap.allocator().free(stack);
         return null;
     }
+    live_stack_bytes += stack.len; // the reaper drops this when the task dies (M8)
     t.* = .{
         .id = next_id,
         .state = .ready,
         .priority = priority,
         .stack = stack,
-        .aspace = aspace,
+        .address_space = address_space,
         .user_ip = entry,
         .user_sp = user_sp,
         .user_arg = user_arg,
@@ -445,6 +520,7 @@ fn startUserTask() void {
 fn create(entry: *const fn () void, priority: Priority, affinity: ?u32) *Task {
     const t = freeSlot() orelse @panic("sched: task table full");
     const stack = heap.allocator().alloc(u8, stack_size) catch @panic("sched: no memory for task stack");
+    live_stack_bytes += stack.len; // the reaper drops this when the task dies (M8)
     t.* = .{ .id = next_id, .state = .ready, .priority = priority, .stack = stack, .affinity = affinity };
     next_id += 1;
     const top = @intFromPtr(stack.ptr) + stack.len;
@@ -485,7 +561,7 @@ fn schedule() void {
 /// Make `next` this core's running task: publish its kernel stack (TSS.rsp0, so a
 /// user-mode interrupt lands on a good stack) and its address space (only when
 /// it differs from what's loaded — every page-table switch is a full TLB flush),
-/// then switch registers/stacks. Kernel tasks (aspace == 0, no kstack_top used
+/// then switch registers/stacks. Kernel tasks (address_space == 0, no kstack_top used
 /// from user mode) resolve to the shared kernel page tables and skip the kernel-
 /// stack write, so this is a no-op beyond the register switch for a pure-kernel
 /// workload. The big kernel lock is held and interrupts are off throughout, so no
@@ -493,12 +569,25 @@ fn schedule() void {
 /// `save_sp` receives the outgoing task's stack pointer.
 fn switchTo(pc: *PerCpu, save_sp: *usize, next: *Task) void {
     if (next.kstack_top != 0) architecture.setKernelStack(pc.index, next.kstack_top);
-    const want = if (next.aspace != 0) next.aspace else architecture.kernelPageTable();
-    if (want != pc.loaded_aspace) {
+    const want = if (next.address_space != 0) next.address_space else architecture.kernelPageTable();
+    if (want != pc.loaded_address_space) {
         architecture.loadPageTable(want);
-        pc.loaded_aspace = want;
+        pc.loaded_address_space = want;
+    }
+    // Restore the next task's user TLS thread pointer — only on change, the same
+    // conditional-load discipline as CR3 above (docs/threading-plan.md M10).
+    if (next.thread_pointer != pc.loaded_thread_pointer) {
+        architecture.setThreadPointer(next.thread_pointer);
+        pc.loaded_thread_pointer = next.thread_pointer;
     }
     architecture.switchContext(save_sp, next.sp);
+    // Resumed now (switchContext returned into our own switchTo frame). Re-fetch the core
+    // via thisCpu(): the `pc` parameter is from *our* earlier switchTo call, so it names
+    // the core we last ran on — stale if we migrated. switchContext only swaps stacks on
+    // the current core, so thisCpu() is the core the just-dead task died on. If a task
+    // died switching to us, free its kernel stack: we're on ours so it's safe, and the big
+    // lock is still held so its slot can't have been reused (docs/threading-plan.md M8).
+    drainReapListLocked(thisCpu());
 }
 
 /// Voluntarily give up the CPU to the next ready task.
@@ -546,12 +635,12 @@ pub fn futexWaitLocked(addr: u64, timeout_ms: u64) FutexResult {
 }
 
 /// Wake up to `count` tasks blocked in `futex_wait` on `addr` in address space
-/// `aspace`. Precondition: the big kernel lock is held. Returns how many woke.
-pub fn futexWakeLocked(aspace: u64, addr: u64, count: u32) u32 {
+/// `address_space`. Precondition: the big kernel lock is held. Returns how many woke.
+pub fn futexWakeLocked(address_space: u64, addr: u64, count: u32) u32 {
     var woken: u32 = 0;
     for (&tasks) |*t| {
         if (woken >= count) break;
-        if (t.state == .blocked and t.aspace == aspace and t.futex_addr == addr) {
+        if (t.state == .blocked and t.address_space == address_space and t.futex_addr == addr) {
             t.futex_addr = 0; // the "woken, not timed out" signal to futexWaitLocked
             t.wake_at = 0;
             t.state = .ready;
@@ -562,6 +651,55 @@ pub fn futexWakeLocked(aspace: u64, addr: u64, count: u32) u32 {
     return woken;
 }
 
+// --- thread join (docs/threading-plan.md M9) --------------------------------
+//
+// join needs no per-thread IPC endpoint: `thread_join(tid)` blocks the caller until the
+// task with id `tid` has exited, and the exit paths wake any joiner. The caller only ever
+// reclaims the joined thread's *user* stack (which the thread vacated the moment it entered
+// the kernel to exit), so waking at exit time — not reap time — is safe.
+
+/// True if a task with id `tid` is still live (has not exited). Caller holds the lock.
+fn aliveTid(tid: u32) bool {
+    for (&tasks) |*t| {
+        if (t.id == tid and t.state != .free and t.state != .reaping) return true;
+    }
+    return false;
+}
+
+/// Block the current task until the task with id `tid` exits (or return at once if it
+/// already has / never existed). **Precondition:** the big kernel lock is held; returns
+/// with it still held. Woken by `wakeJoinersLocked`.
+pub fn joinThreadLocked(tid: u32) void {
+    while (aliveTid(tid)) {
+        const t = current();
+        t.join_target = tid;
+        t.state = .blocked;
+        schedule(); // woken when the joined task exits; lock handed off across the switch
+        t.join_target = 0;
+    }
+}
+
+/// Set the calling task's user TLS thread pointer and load it now. Persisted on the Task so
+/// context switches restore it (docs/threading-plan.md M10). Caller holds the kernel lock.
+pub fn setThreadPointerLocked(addr: u64) void {
+    const pc = thisCpu();
+    pc.current.thread_pointer = addr;
+    architecture.setThreadPointer(addr);
+    pc.loaded_thread_pointer = addr;
+}
+
+/// Wake every task blocked in `thread_join` on `tid` — called from the exit paths once the
+/// exiting task's state is `.free`. Caller holds the lock.
+fn wakeJoinersLocked(tid: u32) void {
+    for (&tasks) |*t| {
+        if (t.state == .blocked and t.join_target == tid) {
+            t.join_target = 0;
+            t.state = .ready;
+            enqueue(t);
+        }
+    }
+}
+
 // --- event-based blocking -------------------------------------------------
 //
 // A WaitQueue is a set of tasks blocked waiting for something (a resource, a
@@ -663,7 +801,7 @@ fn removeFrom(head: *[number_priorities]?*Task, tail: *[number_priorities]?*Task
 /// Precondition: the big kernel lock is held.
 pub fn taskByIdLocked(id: u32) ?*Task {
     for (&tasks) |*t| {
-        if (t.state != .free and t.id == id) return t;
+        if (t.state != .free and t.state != .reaping and t.id == id) return t;
     }
     return null;
 }
@@ -673,7 +811,7 @@ pub fn taskByIdLocked(id: u32) ?*Task {
 /// Precondition: the big kernel lock is held.
 pub fn forgetIpcClientLocked(t: *Task) void {
     for (&tasks) |*other| {
-        if (other.state != .free and other.ipc_client == t) other.ipc_client = null;
+        if (other.state != .free and other.state != .reaping and other.ipc_client == t) other.ipc_client = null;
     }
 }
 
@@ -765,7 +903,11 @@ pub var reap_task_hook: ?*const fn (*Task) void = null;
 fn reapKillPendingLocked() void {
     const pc = thisCpu();
     const cur = pc.current;
-    if (cur.kill_pending and cur.aspace != 0 and !cur.in_system_call) {
+    // Safety net: if a dying task switched to a *fresh* task (which enters via
+    // task_trampoline, not switchTo's tail), its stack is still queued here. The dying
+    // task switched away before this tick, so it is off its stack — drain now (M8).
+    drainReapListLocked(pc);
+    if (cur.kill_pending and cur.address_space != 0 and !cur.in_system_call) {
         if (terminate_current_hook) |hook| hook(); // noreturn
     }
     if (reap_task_hook) |hook| {
@@ -806,7 +948,11 @@ pub fn setPreemption(enabled: bool) void {
 pub fn exit() noreturn {
     _ = sync.enter();
     const pc = thisCpu();
-    pc.current.state = .free;
+    // Queue this task for reaping: `.reaping` keeps its slot out of freeSlot until its
+    // stack is freed; `next` links it on the core's reap list (M8/M9).
+    pc.current.state = .reaping;
+    pc.current.next = pc.reap_list;
+    pc.reap_list = pc.current;
     const next = dequeueHighest(pc) orelse @panic("sched: no task left to run");
     next.state = .running;
     pc.current = next;
@@ -831,17 +977,21 @@ pub fn exitUser() noreturn {
 pub fn exitUserLocked() noreturn {
     const pc = thisCpu();
     const dying = pc.current;
-    const as = dying.aspace;
+    const as = dying.address_space;
     if (as != 0) {
         const kroot = architecture.kernelPageTable();
         architecture.loadPageTable(kroot); // off the process tables before freeing them
-        pc.loaded_aspace = kroot;
-        releaseAspace(as); // destroys only when this was the last task on the space
+        pc.loaded_address_space = kroot;
+        releaseAddressSpace(as); // destroys only when this was the last task on the space
     }
-    dying.state = .free;
-    dying.aspace = 0;
+    dying.state = .reaping; // dead but its slot stays reserved until the stack is freed
+    wakeJoinersLocked(dying.id); // let any thread_join(dying.id) return (M9)
+    dying.address_space = 0;
     dying.kill_pending = false;
     dying.in_system_call = false;
+    // Queue for reaping: the task we switch to (or the next tick) frees this stack (M8/M9).
+    dying.next = pc.reap_list;
+    pc.reap_list = dying;
     const next = dequeueHighest(pc) orelse @panic("sched: no task left to run");
     next.state = .running;
     pc.current = next;
@@ -857,12 +1007,14 @@ pub fn exitUserLocked() noreturn {
 /// task isn't running). The kernel stack is leaked, as in `exitUser` (no reaper
 /// yet). Precondition: the big kernel lock is held.
 pub fn destroyTaskLocked(t: *Task) void {
-    if (t.aspace != 0) releaseAspace(t.aspace); // destroys only on the last reference
-    t.aspace = 0;
+    if (t.address_space != 0) releaseAddressSpace(t.address_space); // destroys only on the last reference
+    reapStackLocked(t); // safe to free now: `t` is not running on any core (M8)
+    t.address_space = 0;
     t.kill_pending = false;
     t.in_system_call = false;
     t.wake_at = 0;
     t.state = .free;
+    wakeJoinersLocked(t.id); // a killed thread's joiners must return too (M9)
 }
 
 /// Snapshot the task table into `out` (up to its length), returning the total
@@ -876,7 +1028,7 @@ pub fn enumerate(out: []abi.ProcessDescriptor) u64 {
     defer sync.leave(flags);
     var total: u64 = 0;
     for (&tasks) |*t| {
-        if (t.state == .free) continue;
+        if (t.state == .free or t.state == .reaping) continue; // reaping = already exited
         if (total < out.len) {
             const d = &out[total];
             d.* = .{
@@ -886,7 +1038,7 @@ pub fn enumerate(out: []abi.ProcessDescriptor) u64 {
                     .ready => .ready,
                     .running => .running,
                     .blocked => .blocked,
-                    .free => unreachable,
+                    .free, .reaping => unreachable,
                 })),
                 .priority = t.priority,
                 .name_length = t.name_length,
@@ -900,7 +1052,7 @@ pub fn enumerate(out: []abi.ProcessDescriptor) u64 {
 
 /// Whether the running task is a user process (has its own address space).
 pub fn currentIsUserProcess() bool {
-    return current().aspace != 0;
+    return current().address_space != 0;
 }
 
 pub fn currentId() u32 {
diff --git a/system/kernel/tests.zig b/system/kernel/tests.zig
index d1c0f61..f6126c5 100644
--- a/system/kernel/tests.zig
+++ b/system/kernel/tests.zig
@@ -139,8 +139,8 @@ pub fn run(case: []const u8, boot_information: *const BootInformation) void {
         userPfTest();
     } else if (eql(case, "fault-recovery")) {
         faultRecoveryTest(boot_information);
-    } else if (eql(case, "aspace-refcount")) {
-        aspaceRefcountTest(boot_information);
+    } else if (eql(case, "address-space-refcount")) {
+        addressSpaceRefcountTest(boot_information);
     } else if (eql(case, "thread-spawn")) {
         threadSpawnTest(boot_information);
     } else if (eql(case, "thread-join")) {
@@ -151,6 +151,14 @@ pub fn run(case: []const u8, boot_information: *const BootInformation) void {
         threadMutexTest(boot_information);
     } else if (eql(case, "thread-id")) {
         threadIdTest(boot_information);
+    } else if (eql(case, "thread-alloc")) {
+        threadAllocTest(boot_information);
+    } else if (eql(case, "task-reap")) {
+        taskReapTest(boot_information);
+    } else if (eql(case, "thread-tls")) {
+        threadTlsTest(boot_information);
+    } else if (eql(case, "thread-rwlock")) {
+        threadRwlockTest(boot_information);
     } else if (eql(case, "args")) {
         argsTest(boot_information);
     } else if (eql(case, "init")) {
@@ -779,11 +787,15 @@ fn affinityTest() void {
         return;
     }
 
-    var spins: u64 = 0;
-    while (spins < 3_000_000_000) spins +%= 1; // many time slices across the cores
+    // Let many time slices pass so the scheduler runs the pinned worker across ticks.
+    // Wait on the wall clock, not a raw iteration count: a fixed-count busy-loop's
+    // wall-time is a codegen lottery (the optimiser may elide or vectorise it), so an
+    // unrelated change elsewhere in this file could swing this test from ~4 s to ~50 s.
+    const run_until = architecture.millis() + 400;
+    while (architecture.millis() < run_until) {}
     affinity_running = false;
-    var settle: u64 = 0;
-    while (settle < 200_000_000) settle +%= 1; // let the worker see the flag and exit
+    const settle_until = architecture.millis() + 50;
+    while (architecture.millis() < settle_until) {} // let the worker see the flag and exit
 
     var others: u32 = 0;
     for (affinity_cores, 0..) |seen, c| {
@@ -908,12 +920,12 @@ fn userMemTest() void {
     log("DANOS-TEST-BEGIN: usermem\n", .{});
     const base_free = pmm.stats().free_frames;
 
-    const aspace = architecture.createAddressSpace() orelse {
+    const address_space = architecture.createAddressSpace() orelse {
         check("created a fresh address space", false);
         result();
         return;
     };
-    check("created a fresh address space", aspace != 0);
+    check("created a fresh address space", address_space != 0);
 
     // Grant three pages into the arena, mapped RW + NX (the mmap contract).
     const npages = 3;
@@ -922,7 +934,7 @@ fn userMemTest() void {
     var mapped: usize = 0;
     while (mapped < npages) : (mapped += 1) {
         frames[mapped] = pmm.alloc() orelse break;
-        architecture.mapUserPageInto(aspace, arena + mapped * abi.page_size, frames[mapped], true, false);
+        architecture.mapUserPageInto(address_space, arena + mapped * abi.page_size, frames[mapped], true, false);
     }
     check("granted three user pages", mapped == npages);
 
@@ -931,7 +943,7 @@ fn userMemTest() void {
     var rw_ok = true;
     for (0..npages) |i| {
         const va = arena + i * abi.page_size;
-        const physical = architecture.translate(aspace, va) orelse {
+        const physical = architecture.translate(address_space, va) orelse {
             translate_ok = false;
             continue;
         };
@@ -946,13 +958,13 @@ fn userMemTest() void {
     // Release them the way munmap does, then tear down the address space.
     for (0..npages) |i| {
         const va = arena + i * abi.page_size;
-        if (architecture.translate(aspace, va)) |physical| {
-            architecture.unmapUserPageInto(aspace, va);
+        if (architecture.translate(address_space, va)) |physical| {
+            architecture.unmapUserPageInto(address_space, va);
             pmm.free(physical);
         }
     }
-    check("munmap unmapped every grant", architecture.translate(aspace, arena) == null);
-    architecture.destroyAddressSpace(aspace);
+    check("munmap unmapped every grant", architecture.translate(address_space, arena) == null);
+    architecture.destroyAddressSpace(address_space);
 
     check("no frames leaked (free count restored)", pmm.stats().free_frames == base_free);
     result();
@@ -1120,12 +1132,12 @@ fn dmaTest() void {
 
     // Map the run into a fresh address space as coherent DMA and translate each page
     // back: the same physical run, in order — proving contiguity and the mapping.
-    const aspace = architecture.createAddressSpace().?;
-    architecture.mapUserDmaInto(aspace, process.dma_arena_base, phys, frames * abi.page_size);
+    const address_space = architecture.createAddressSpace().?;
+    architecture.mapUserDmaInto(address_space, process.dma_arena_base, phys, frames * abi.page_size);
     var mapped_ok = true;
     for (0..frames) |i| {
         const va = process.dma_arena_base + i * abi.page_size;
-        const got = architecture.translate(aspace, va) orelse {
+        const got = architecture.translate(address_space, va) orelse {
             mapped_ok = false;
             break;
         };
@@ -1135,7 +1147,7 @@ fn dmaTest() void {
 
     // Teardown must reclaim the DMA RAM (the leaves carry no device_grant, so
     // freeSubtree frees them as ordinary frames) — a driver that just dies leaks none.
-    architecture.destroyAddressSpace(aspace);
+    architecture.destroyAddressSpace(address_space);
     for (0..2) |i| pmm.free(low + i * abi.page_size);
     check("no frames leaked after DMA teardown", pmm.stats().free_frames == base_free);
     result();
@@ -1367,9 +1379,9 @@ fn spawnFaultingProcess() ?u32 {
     const flags = sync.enter();
     defer sync.leave(flags);
 
-    const aspace = architecture.createAddressSpace() orelse return null;
+    const address_space = architecture.createAddressSpace() orelse return null;
     const code_frame = pmm.alloc() orelse {
-        architecture.destroyAddressSpace(aspace);
+        architecture.destroyAddressSpace(address_space);
         return null;
     };
     // Fill through the physmap (the user mapping is read-only); pad with int3 so a
@@ -1377,17 +1389,17 @@ fn spawnFaultingProcess() ?u32 {
     const code: [*]u8 = @ptrFromInt(boot_handoff.physicalToVirtual(code_frame));
     @memset(code[0..abi.page_size], 0xCC);
     @memcpy(code[0..blob.len], blob);
-    architecture.mapUserPageInto(aspace, process.code_virtual, code_frame, false, true); // RO + X
+    architecture.mapUserPageInto(address_space, process.code_virtual, code_frame, false, true); // RO + X
 
     const stack_frame = pmm.alloc() orelse {
-        architecture.destroyAddressSpace(aspace); // frees code_frame too — it's mapped
+        architecture.destroyAddressSpace(address_space); // frees code_frame too — it's mapped
         return null;
     };
-    architecture.mapUserPageInto(aspace, process.stack_base_virtual, stack_frame, true, false); // RW + NX
+    architecture.mapUserPageInto(address_space, process.stack_base_virtual, stack_frame, true, false); // RW + NX
 
     // Supervised by the calling test task, so exitReasonOf can read the verdict.
-    const id = scheduler.spawnUserLocked(aspace, process.code_virtual, process.stack_base_virtual + abi.page_size, 0, 4, "fault-probe", scheduler.currentId(), null) orelse {
-        architecture.destroyAddressSpace(aspace);
+    const id = scheduler.spawnUserLocked(address_space, process.code_virtual, process.stack_base_virtual + abi.page_size, 0, 4, "fault-probe", scheduler.currentId(), null) orelse {
+        architecture.destroyAddressSpace(address_space);
         return null;
     };
     return id;
@@ -1448,11 +1460,11 @@ fn faultRecoveryTest(boot_information: *const BootInformation) void {
 /// address spaces returns to baseline while destructions advance by exactly that many.
 /// This is the foundation threads (shared address spaces) build on: the refactor must be
 /// invisible while every space still has exactly one task.
-fn aspaceRefcountTest(boot_information: *const BootInformation) void {
+fn addressSpaceRefcountTest(boot_information: *const BootInformation) void {
     _ = boot_information;
-    log("DANOS-TEST-BEGIN: aspace-refcount\n", .{});
-    const base_live = scheduler.liveAspaceCount();
-    const base_destroyed = scheduler.aspaceDestroyCount();
+    log("DANOS-TEST-BEGIN: address-space-refcount\n", .{});
+    const base_live = scheduler.liveAddressSpaceCount();
+    const base_destroyed = scheduler.addressSpaceDestroyCount();
     const rounds: u32 = 5;
     var killed: u32 = 0;
     var round: u32 = 0;
@@ -1468,11 +1480,11 @@ fn aspaceRefcountTest(boot_information: *const BootInformation) void {
         if (process.fault_kill_count >= 1) killed += 1;
     }
     check("all probes spawned and were killed", killed == rounds);
-    check("live address-space count returned to baseline", scheduler.liveAspaceCount() == base_live);
-    check("each address space destroyed exactly once", scheduler.aspaceDestroyCount() == base_destroyed + rounds);
-    if (killed == rounds and scheduler.liveAspaceCount() == base_live and
-        scheduler.aspaceDestroyCount() == base_destroyed + rounds)
-        log("aspace-refcount: spaces released to baseline ok\n", .{});
+    check("live address-space count returned to baseline", scheduler.liveAddressSpaceCount() == base_live);
+    check("each address space destroyed exactly once", scheduler.addressSpaceDestroyCount() == base_destroyed + rounds);
+    if (killed == rounds and scheduler.liveAddressSpaceCount() == base_live and
+        scheduler.addressSpaceDestroyCount() == base_destroyed + rounds)
+        log("address-space-refcount: spaces released to baseline ok\n", .{});
     result();
 }
 
@@ -1497,7 +1509,7 @@ fn threadSpawnTest(boot_information: *const BootInformation) void {
     check("thread-test spawned", spawnNamed(rd, "thread-test"));
 
     // Wait for the service's verdict marker (it polls shared memory the worker wrote).
-    const ok_marker = "thread-test: child ran in shared aspace ok";
+    const ok_marker = "thread-test: child ran in shared address space ok";
     const fail_marker = "thread-test: FAIL";
     scheduler.setPriority(1);
     const deadline = architecture.millis() + 12000;
@@ -1689,6 +1701,172 @@ fn threadIdTest(boot_information: *const BootInformation) void {
     result();
 }
 
+/// Thread-safe allocation (docs/threading-plan.md M7): `thread-test` in alloc mode runs N
+/// threads that each do many `alloc`/fill/verify/`free` cycles of varied sizes on the
+/// shared runtime heap. If the heap lock or the per-address-space mmap arena were unsafe,
+/// two threads' blocks would overlap and a thread would read another's pattern; the
+/// verdict marker is emitted only when every thread completes with every block intact.
+fn threadAllocTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: thread-alloc\n", .{});
+    if (boot_information.initial_ramdisk_len == 0) {
+        check("bootloader handed over an initial_ramdisk", false);
+        result();
+        return;
+    }
+    const image = @as([*]const u8, @ptrFromInt(boot_handoff.physicalToVirtual(boot_information.initial_ramdisk_base)))[0..boot_information.initial_ramdisk_len];
+    const rd = initial_ramdisk.Reader.init(image) orelse {
+        check("initial_ramdisk image is valid", false);
+        result();
+        return;
+    };
+
+    var started = false;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "thread-test")) continue;
+        started = if (process.spawnProcess(item.blob, 4, &.{ "thread-test", "alloc" })) true else |_| false;
+        break;
+    }
+    check("thread-test (alloc mode) spawned", started);
+
+    const ok_marker = "thread-alloc: ok";
+    const fail_marker = "thread-alloc: FAIL";
+    scheduler.setPriority(1);
+    const deadline = architecture.millis() + 20000;
+    while (architecture.millis() < deadline) {
+        if (bufferHas(ok_marker) or bufferHas(fail_marker)) break;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+
+    check("concurrent heap allocation stayed corruption-free (shared heap + per-address-space arena)", bufferHas(ok_marker) and !bufferHas(fail_marker));
+    result();
+}
+
+/// Per-thread TLS / FS base (docs/threading-plan.md M10): `thread-test` in tls mode has two
+/// threads each set their own FS base and write a unique marker to `%fs:8`, then — after
+/// both have written — read it back. If the FS base were not per-thread and restored across
+/// context switches, the second write would clobber the first and a thread would read the
+/// wrong marker. The verdict marker means both read their own value (no cross-talk).
+fn threadTlsTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: thread-tls\n", .{});
+    if (boot_information.initial_ramdisk_len == 0) {
+        check("bootloader handed over an initial_ramdisk", false);
+        result();
+        return;
+    }
+    const image = @as([*]const u8, @ptrFromInt(boot_handoff.physicalToVirtual(boot_information.initial_ramdisk_base)))[0..boot_information.initial_ramdisk_len];
+    const rd = initial_ramdisk.Reader.init(image) orelse {
+        check("initial_ramdisk image is valid", false);
+        result();
+        return;
+    };
+
+    var started = false;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "thread-test")) continue;
+        started = if (process.spawnProcess(item.blob, 4, &.{ "thread-test", "tls" })) true else |_| false;
+        break;
+    }
+    check("thread-test (tls mode) spawned", started);
+
+    const ok_marker = "thread-tls: ok";
+    const fail_marker = "thread-tls: FAIL";
+    scheduler.setPriority(1);
+    const deadline = architecture.millis() + 12000;
+    while (architecture.millis() < deadline) {
+        if (bufferHas(ok_marker) or bufferHas(fail_marker)) break;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+
+    check("each thread has its own FS-base TLS slot (no cross-talk across switches)", bufferHas(ok_marker) and !bufferHas(fail_marker));
+    result();
+}
+
+/// RwLock (docs/threading-plan.md M11): `thread-test` in rwlock mode runs writers that set
+/// two halves of a value under the exclusive lock and readers that check the halves match
+/// under the shared lock. If the reader/writer lock were wrong, a reader would observe a
+/// half-written value; zero violations across many reads → the lock holds.
+fn threadRwlockTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: thread-rwlock\n", .{});
+    if (boot_information.initial_ramdisk_len == 0) {
+        check("bootloader handed over an initial_ramdisk", false);
+        result();
+        return;
+    }
+    const image = @as([*]const u8, @ptrFromInt(boot_handoff.physicalToVirtual(boot_information.initial_ramdisk_base)))[0..boot_information.initial_ramdisk_len];
+    const rd = initial_ramdisk.Reader.init(image) orelse {
+        check("initial_ramdisk image is valid", false);
+        result();
+        return;
+    };
+
+    var started = false;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "thread-test")) continue;
+        started = if (process.spawnProcess(item.blob, 4, &.{ "thread-test", "rwlock" })) true else |_| false;
+        break;
+    }
+    check("thread-test (rwlock mode) spawned", started);
+
+    const ok_marker = "thread-rwlock: ok";
+    const fail_marker = "thread-rwlock: FAIL";
+    scheduler.setPriority(1);
+    const deadline = architecture.millis() + 20000;
+    while (architecture.millis() < deadline) {
+        if (bufferHas(ok_marker) or bufferHas(fail_marker)) break;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+
+    check("readers/writers over an RwLock never observed a half-written value", bufferHas(ok_marker) and !bufferHas(fail_marker));
+    result();
+}
+
+/// The task reaper (docs/threading-plan.md M8): a dead task's kernel stack used to be
+/// leaked ("no reaper yet"). Spawn and kill many ring-3 processes and confirm the total
+/// kernel-stack bytes return to baseline — every stack reclaimed, no leak. (Threads exit
+/// through the same exitUserLocked path, so this covers them too.)
+fn taskReapTest(boot_information: *const BootInformation) void {
+    _ = boot_information;
+    log("DANOS-TEST-BEGIN: task-reap\n", .{});
+    const base = scheduler.liveStackBytes();
+    const rounds: u32 = 12;
+    var killed: u32 = 0;
+    var round: u32 = 0;
+    while (round < rounds) : (round += 1) {
+        process.fault_kill_count = 0;
+        const probe = spawnFaultingProcess() orelse break;
+        _ = probe;
+        scheduler.setPriority(1);
+        const deadline = architecture.millis() + 5000;
+        while (process.fault_kill_count < 1 and architecture.millis() < deadline) scheduler.yield();
+        scheduler.setPriority(4);
+        if (process.fault_kill_count >= 1) killed += 1;
+    }
+    // Reaping is asynchronous — a dead task's stack is freed when its core next switches
+    // or ticks. Poll (bounded) until the live bytes return to baseline: a correct reaper
+    // gets there in a few ms; a genuine leak never does and this times out.
+    scheduler.setPriority(1);
+    const settle_deadline = architecture.millis() + 3000;
+    while (scheduler.liveStackBytes() != base and architecture.millis() < settle_deadline) scheduler.yield();
+    scheduler.setPriority(4);
+
+    const final = scheduler.liveStackBytes();
+    log("task-reap: base={d} final={d} killed={d}/{d}\n", .{ base, final, killed, rounds });
+    check("all probes spawned and were killed", killed == rounds);
+    check("kernel stacks reclaimed to baseline (no leak)", final == base);
+    if (killed == rounds and final == base)
+        log("task-reap: kernel stacks reclaimed to baseline ok\n", .{});
+    result();
+}
+
 /// The full PID-1 path: the bootloader read /system/services/init off the boot volume and
 /// handed it over; load it as a user ELF and spawn it as a real ring-3 process
 /// — the same call the normal boot path makes — then confirm it beats. init
@@ -3111,20 +3289,20 @@ fn ioPassTest() void {
     log("DANOS-TEST-BEGIN: iopass\n", .{});
     const base_free = pmm.stats().free_frames;
 
-    const aspace = architecture.createAddressSpace() orelse {
+    const address_space = architecture.createAddressSpace() orelse {
         check("created a fresh address space", false);
         result();
         return;
     };
     const frame = pmm.alloc() orelse {
-        architecture.destroyAddressSpace(aspace);
+        architecture.destroyAddressSpace(address_space);
         check("allocated a frame to grant", false);
         result();
         return;
     };
     // Map it the way mmio_map does (device grant, strong-uncacheable), then tear the space down.
-    architecture.mapUserDeviceInto(aspace, process.device_arena_base, frame, abi.page_size, false);
-    architecture.destroyAddressSpace(aspace);
+    architecture.mapUserDeviceInto(address_space, process.device_arena_base, frame, abi.page_size, false);
+    architecture.destroyAddressSpace(address_space);
 
     // The page tables were reclaimed; the device-granted frame must not have been.
     check("device-granted frame survived teardown (not reclaimed as RAM)", pmm.stats().free_frames == base_free - 1);
@@ -3176,26 +3354,26 @@ fn displayTest(boot_information: *const BootInformation) void {
     // space, and confirm the leaf's cache type. We never run this space (no CR3 load) —
     // we only read back the page-table entries — so aliasing the same physical page at
     // two cache types below is inert.
-    const aspace = architecture.createAddressSpace() orelse {
+    const address_space = architecture.createAddressSpace() orelse {
         check("created a fresh address space", false);
         result();
         return;
     };
-    defer architecture.destroyAddressSpace(aspace);
+    defer architecture.destroyAddressSpace(address_space);
 
     const page_base = fb.base & ~@as(u64, abi.page_size - 1);
-    architecture.mapUserDeviceInto(aspace, process.device_arena_base, page_base, abi.page_size, true);
+    architecture.mapUserDeviceInto(address_space, process.device_arena_base, page_base, abi.page_size, true);
     check(
         "the framebuffer maps write-combining (PAT entry 4: PAT bit set, PCD/PWT clear)",
-        architecture.userLeafIsWriteCombining(aspace, process.device_arena_base) == true,
+        architecture.userLeafIsWriteCombining(address_space, process.device_arena_base) == true,
     );
 
     // Regression guard: the strong-uncacheable default is still that, so WC is a real
     // choice the flag makes, not the only behaviour.
-    architecture.mapUserDeviceInto(aspace, process.device_arena_base + abi.page_size, page_base, abi.page_size, false);
+    architecture.mapUserDeviceInto(address_space, process.device_arena_base + abi.page_size, page_base, abi.page_size, false);
     check(
         "a register window still maps strong-uncacheable",
-        architecture.userLeafIsWriteCombining(aspace, process.device_arena_base + abi.page_size) == false,
+        architecture.userLeafIsWriteCombining(address_space, process.device_arena_base + abi.page_size) == false,
     );
 
     log("display: mapped {d}x{d} pitch {d} (write-combining)\n", .{ fb.width, fb.height, fb.pitch });
diff --git a/system/services/thread-test/thread-test.zig b/system/services/thread-test/thread-test.zig
index bd15643..50a72c4 100644
--- a/system/services/thread-test/thread-test.zig
+++ b/system/services/thread-test/thread-test.zig
@@ -40,7 +40,7 @@ fn runSpawnMode() void {
         runtime.system.yield();
     }
     if (spawn_done.load(.acquire) == 1 and shared_value == sentinel) {
-        write("thread-test: child ran in shared aspace ok\n");
+        write("thread-test: child ran in shared address space ok\n");
     } else {
         write("thread-test: FAIL worker did not update shared memory\n");
     }
@@ -74,6 +74,8 @@ fn detachWorker() void {
     detach_done.store(1, .release);
 }
 
+fn noopWorker() void {}
+
 fn runJoinMode() void {
     write("thread-test: join mode starting\n");
 
@@ -114,7 +116,19 @@ fn runJoinMode() void {
         return;
     }
 
-    write("thread-test: join ok\n"); // the M3 verdict marker
+    // Prove join needs no per-thread kernel endpoint (M9): many spawn+join cycles. Under
+    // the old per-thread-endpoint scheme these leaked handles and would exhaust the
+    // 16-slot handle table well before 40; here they all succeed.
+    var cycle: u32 = 0;
+    while (cycle < 40) : (cycle += 1) {
+        const th = runtime.Thread.spawn(.{}, noopWorker, .{}) catch {
+            write("thread-test: FAIL spawn exhausted across join cycles (endpoint leak?)\n");
+            return;
+        };
+        th.join();
+    }
+
+    write("thread-test: join ok\n"); // the M3/M9 verdict marker
 }
 
 // --- M4: futex mode ---------------------------------------------------------
@@ -294,6 +308,167 @@ fn runIdMode() void {
     write("thread-id: ok\n"); // the M6 verdict marker
 }
 
+// --- M7: alloc mode (concurrent heap allocation) ----------------------------
+
+const alloc_threads: u32 = 4;
+const allocs_per_thread: u32 = 500;
+var allocs_clean = std.atomic.Value(u32).init(0);
+
+fn allocWorker(seed: u32) void {
+    const gpa = runtime.allocator();
+    var rng: u32 = seed | 1;
+    var round: u32 = 0;
+    while (round < allocs_per_thread) : (round += 1) {
+        rng = rng *% 1664525 +% 1013904223; // cheap LCG for varied sizes
+        const size: usize = 16 + (rng % 4080); // 16..4095 bytes
+        const buf = gpa.alloc(u8, size) catch return; // OOM: don't count this thread clean
+        const pattern: u8 = @truncate(seed +% round);
+        @memset(buf, pattern);
+        // Nothing else should touch our block; if a concurrent allocation overlapped it,
+        // one of us would read the other's pattern here.
+        var ok = true;
+        for (buf) |b| {
+            if (b != pattern) ok = false;
+        }
+        gpa.free(buf);
+        if (!ok) return; // corruption — leave without counting clean
+    }
+    _ = allocs_clean.fetchAdd(1, .monotonic);
+}
+
+fn runAllocMode() void {
+    write("thread-alloc: starting\n");
+    var threads: [alloc_threads]runtime.Thread = undefined;
+    var n: u32 = 0;
+    while (n < alloc_threads) : (n += 1) {
+        threads[n] = runtime.Thread.spawn(.{}, allocWorker, .{n +% 1}) catch {
+            write("thread-alloc: FAIL spawn\n");
+            return;
+        };
+    }
+    for (threads[0..alloc_threads]) |t| t.join();
+
+    // Every thread must have completed all rounds with each block intact — proof the
+    // shared heap and the per-address_space mmap arena are safe under concurrent allocation.
+    if (allocs_clean.load(.acquire) != alloc_threads) {
+        write("thread-alloc: FAIL corruption or OOM under concurrent allocation\n");
+        return;
+    }
+    write("thread-alloc: ok\n"); // the M7 verdict marker
+}
+
+// --- M10: tls mode (per-thread FS base storage) -----------------------------
+
+fn writeTlsSlot(value: u64) void {
+    asm volatile ("movq %[v], %%fs:8"
+        :
+        : [v] "r" (value),
+        : .{ .memory = true });
+}
+
+fn readTlsSlot() u64 {
+    return asm volatile ("movq %%fs:8, %[out]"
+        : [out] "=r" (-> u64),
+        :
+        : .{ .memory = true });
+}
+
+var tls_written = std.atomic.Value(u32).init(0);
+var tls_ok = std.atomic.Value(u32).init(0);
+
+fn tlsWorker(marker: u64) void {
+    writeTlsSlot(marker);
+    _ = tls_written.fetchAdd(1, .release);
+    // Wait until both threads have written their own slot. If the FS base were shared, the
+    // second write would clobber the first, and the read below would return the wrong
+    // marker — cross-talk. A per-thread FS base keeps each thread's slot private.
+    var spins: usize = 0;
+    while (tls_written.load(.acquire) < 2 and spins < 50_000_000) : (spins += 1) {
+        runtime.system.yield();
+    }
+    if (readTlsSlot() == marker and runtime.Thread.getCurrentId() != 0) {
+        _ = tls_ok.fetchAdd(1, .monotonic);
+    }
+}
+
+fn runTlsMode() void {
+    write("thread-tls: starting\n");
+    const t0 = runtime.Thread.spawn(.{}, tlsWorker, .{@as(u64, 0xAAAA_0000)}) catch {
+        write("thread-tls: FAIL spawn\n");
+        return;
+    };
+    const t1 = runtime.Thread.spawn(.{}, tlsWorker, .{@as(u64, 0xBBBB_0000)}) catch {
+        write("thread-tls: FAIL spawn\n");
+        return;
+    };
+    t0.join();
+    t1.join();
+    if (tls_ok.load(.acquire) == 2) {
+        write("thread-tls: ok\n"); // the M10 verdict marker
+    } else {
+        write("thread-tls: FAIL cross-talk (FS base not per-thread)\n");
+    }
+}
+
+// --- M11: rwlock mode (readers/writers over an RwLock) ----------------------
+
+const RwLock = runtime.Thread.RwLock;
+
+var rwlock = RwLock{};
+var rw_a: u64 = 0;
+var rw_b: u64 = 0; // invariant while any lock is held: rw_a == rw_b
+var rw_stop = std.atomic.Value(u32).init(0);
+var rw_violations = std.atomic.Value(u32).init(0);
+var rw_reads = std.atomic.Value(u64).init(0);
+
+fn rwWriter() void {
+    var v: u64 = 1;
+    while (rw_stop.load(.acquire) == 0) : (v +%= 1) {
+        rwlock.lock(); // exclusive: no reader may observe the gap between the two writes
+        rw_a = v;
+        rw_b = v;
+        rwlock.unlock();
+    }
+}
+
+fn rwReader() void {
+    const reads: u64 = 50_000;
+    var i: u64 = 0;
+    while (i < reads) : (i += 1) {
+        rwlock.lockShared();
+        if (rw_a != rw_b) _ = rw_violations.fetchAdd(1, .monotonic); // saw a half-write!
+        rwlock.unlockShared();
+    }
+    _ = rw_reads.fetchAdd(reads, .monotonic);
+}
+
+fn runRwlockMode() void {
+    write("thread-rwlock: starting\n");
+    var writers: [2]runtime.Thread = undefined;
+    var readers: [3]runtime.Thread = undefined;
+    for (&writers) |*w| {
+        w.* = runtime.Thread.spawn(.{}, rwWriter, .{}) catch {
+            write("thread-rwlock: FAIL spawn\n");
+            return;
+        };
+    }
+    for (&readers) |*r| {
+        r.* = runtime.Thread.spawn(.{}, rwReader, .{}) catch {
+            write("thread-rwlock: FAIL spawn\n");
+            return;
+        };
+    }
+    for (readers) |r| r.join();
+    rw_stop.store(1, .release); // readers done → stop the writers
+    for (writers) |w| w.join();
+
+    if (rw_violations.load(.acquire) == 0 and rw_reads.load(.acquire) > 0) {
+        write("thread-rwlock: ok\n"); // the M11 verdict marker
+    } else {
+        write("thread-rwlock: FAIL reader observed a half-written value\n");
+    }
+}
+
 pub fn main(init: runtime.process.Init) void {
     const mode = init.arguments.get(1) orelse "spawn";
     if (std.mem.eql(u8, mode, "join")) {
@@ -304,6 +479,12 @@ pub fn main(init: runtime.process.Init) void {
         runMutexMode();
     } else if (std.mem.eql(u8, mode, "id")) {
         runIdMode();
+    } else if (std.mem.eql(u8, mode, "alloc")) {
+        runAllocMode();
+    } else if (std.mem.eql(u8, mode, "tls")) {
+        runTlsMode();
+    } else if (std.mem.eql(u8, mode, "rwlock")) {
+        runRwlockMode();
     } else {
         runSpawnMode();
     }
diff --git a/test/qemu_test.py b/test/qemu_test.py
index f5753ec..ecca6b2 100644
--- a/test/qemu_test.py
+++ b/test/qemu_test.py
@@ -295,8 +295,8 @@ CASES = [
      "fail":   r"DANOS-TEST-RESULT: FAIL"},
 
     # docs/threading-plan.md M1: address-space refcount — spaces destroyed exactly
-    # once per process, no leak/double-free (the foundation shared-aspace threads need).
-    {"name": "aspace-refcount",
+    # once per process, no leak/double-free (the foundation shared-address-space threads need).
+    {"name": "address-space-refcount",
      "timeout": 60,
      "expect": r"DANOS-TEST-RESULT: PASS",
      "fail":   r"DANOS-TEST-RESULT: FAIL"},
@@ -339,6 +339,38 @@ CASES = [
      "timeout": 60,
      "expect": r"DANOS-TEST-RESULT: PASS",
      "fail":   r"DANOS-TEST-RESULT: FAIL"},
+
+    # docs/threading-plan.md M7: thread-safe allocation — N threads hammer the shared heap
+    # (per-aspace mmap arena + locked free list) with no cross-block corruption.
+    {"name": "thread-alloc",
+     "smp":    4,
+     "timeout": 60,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+
+    # docs/threading-plan.md M8: the task reaper — spawn+kill many processes; total kernel
+    # stack bytes return to baseline (every dead task's stack reclaimed, no leak).
+    {"name": "task-reap",
+     "smp":    4,
+     "timeout": 60,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+
+    # docs/threading-plan.md M10: per-thread fs.base — two threads keep private %fs:8 TLS
+    # slots across context switches (no cross-talk).
+    {"name": "thread-tls",
+     "smp":    4,
+     "timeout": 60,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+
+    # docs/threading-plan.md M11: RwLock — readers/writers across cores; a reader never
+    # observes a half-written value (writers hold it exclusively).
+    {"name": "thread-rwlock",
+     "smp":    4,
+     "timeout": 60,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
     # Process arguments: argv arrives on the SysV entry stack (argv[0] = the spawned
     # name, argv[1..] = the system_spawn argument blob) and echoes back intact.
     {"name": "args",