The flip: PCI enumeration leaves the kernel (M19.3)

enumeratePci, addBars, pciConfigurationPtr, and the PciHeader struct are
deleted; the kernel seeds only the host bridge, and the ring-3 pci-bus
driver's reports are the sole source of PCI function nodes. The manager
matches PCI drivers from reported identity, deduped by registered device
id so a bus restart never double-spawns.

The flip did its job by exposing a latent SMP race: ring-3
device_register made the broker table concurrent for the first time, and
mmio_map read it lock-free — under load a torn resource length mapped
hpet's window wrong (its user fault) and underflowed r.len-1 into a
kernel integer-overflow panic. Fixed: the broker read in mmio_map (and
claim) runs under the big kernel lock, the arithmetic rejects
zero-length and wrapping windows cleanly, and pci-bus no longer registers
unimplemented size-0 BARs. driver-restart hammered 6x, suite 55/55.

README.md

@@ -2,12 +2,31 @@
 Codename: Shodan
 Version: 1
 
-A small operating system, written from scratch in Zig — a bootloader (`boot/`)
-and a microkernel (`system/kernel/`), sharing a neutral handoff contract (`system/boot-handoff.zig`).
-It boots x86-64 via UEFI, and so far has a framebuffer console, a physical frame
-allocator, its own paging with W^X permissions, interrupt/exception handling, a
-LAPIC timer, a kernel heap, a fixed-priority preemptive scheduler, and in-kernel IPC
-channels. See [`docs/`](docs/README.md) for how each piece works.
+A small resilient operating system, written from scratch in Zig.
+
+## Zen of DanOS:
+
+- Resilient Micro-Kernel Architecture.
+    - Every process run in an isolated user space not kernel space.
+    - Processes cannot take down the entire OS with it when they die or is killed
+- Stable public runtime library, private OS ABI.
+    - Keeps a stable runtime for user space processes between OS versions (great for backwards compatibility)
+    - Allows the underlying OS to be changed without effecting applications
+    - Provides a boundary to enable compatibility between OS's e.g. POSIX, MUSL etc
+- Drivers are just isolated processes in user space.
+    - Thin binaries that can be restarted like applications.
+    - Useful during driver development.
+    - Drivers can claim MMIO / ports
+    - Driver resources (e.g. IRQ/Port/MMIO) claims are automatically cleaned up if the driver dies or is killed
+    - Drivers can also hook into the process lifecyle to clean up or reset hardware
+- No legacy to deal with
+    - Zig code uses a clean coding style (Zen of Zig)
+    - Favor reading code over writing code.
+    - No magic numbers.
+    - No shortend names unless its for ABI compatibility or acronyms
+- Inter-Process Communication (IPC)
+  - Publish and subscribe to Asynchronous Messages
+  - Talk to services and processes synchronously
 
 ## Prerequisites
 
@@ -60,7 +79,7 @@ straight into CI.
 
 ## Documentation
 
-Design notes explaining the *why* behind the code live in
+Design notes explaining *why* behind the code live in
 [`docs/`](docs/README.md) — start with [`docs/README.md`](docs/README.md).
 
 ## Logo

build.zig

@@ -212,6 +212,13 @@ pub fn build(b: *std.Build) void {
         },
     });
 
+    // The device-manager protocol: hello + (M18.2) tree reports, exposed as its
+    // own module like the other protocol modules. Imported through the runtime.
+    const device_manager_protocol_module = b.addModule("device-manager-protocol", .{
+        .root_source_file = b.path("system/services/device-manager/device-manager-protocol.zig"),
+    });
+    runtime_module.addImport("device-manager-protocol", device_manager_protocol_module);
+
     // Typed volatile MMIO register access + memory-ordering barriers, for drivers on
     // top of an mmio_map grant. Depends only on `builtin` (arch-conditional barriers);
     // no target set, so it inherits each driver's. See library/mmio/mmio.zig.
@@ -332,6 +339,25 @@ pub fn build(b: *std.Build) void {
     const ps2_keyboard_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "ps2-keyboard", "system/drivers/ps2-bus/keyboard.zig");
     const ps2_mouse_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "ps2-mouse", "system/drivers/ps2-bus/mouse.zig");
     const usb_xhci_bus_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "usb-xhci-bus", "system/drivers/usb-xhci-bus/usb-xhci-bus.zig");
+    const pci_bus_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "pci-bus", "system/drivers/pci-bus/pci-bus.zig");
+    // A test fixture, not a real driver: hellos to the device manager, then faults —
+    // what the driver-restart scenario drives the crash-loop cap with.
+    const crash_test_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "crash-test", "system/services/crash-test/crash-test.zig");
+    const device_list_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "device-list", "system/services/device-list/device-list.zig");
+    // The discovery service: one swappable process per firmware
+    // (docs/m19-m20-plan.md decision 7), bundled under the neutral ramdisk name
+    // "discovery" so the device manager never learns which firmware it is on.
+    // x86 boots describe hardware with ACPI; the Raspberry Pis hand over a
+    // flattened device tree — the aarch64 target flips the default when it
+    // lands (docs/arm.md). Both are placeholders until M20.1 (acpi) and the
+    // ARM bring-up (fdt).
+    const Discovery = enum { acpi, fdt };
+    const discovery = b.option(Discovery, "discovery", "Which discovery service fills the ramdisk's 'discovery' slot (default: acpi)") orelse Discovery.acpi;
+    const discovery_source: []const u8 = switch (discovery) {
+        .acpi => "system/services/acpi/acpi.zig",
+        .fdt => "system/services/fdt/fdt.zig",
+    };
+    const discovery_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "discovery", discovery_source);
     const device_manager_exe = addUserBinary(b, kernel_target, runtime_module, posix_module, mmio_module, xkeyboard_config_module, acpi_ids_module, "device-manager", "system/services/device-manager/device-manager.zig");
     // The input service and its exercisers: the fan-out server, a hardware-free synthetic
     // source, and a subscriber that doubles as the `input` test's oracle. See docs/input.md.
@@ -363,6 +389,14 @@ pub fn build(b: *std.Build) void {
     mk_run.addFileArg(ps2_mouse_exe.getEmittedBin());
     mk_run.addArg("usb-xhci-bus");
     mk_run.addFileArg(usb_xhci_bus_exe.getEmittedBin());
+    mk_run.addArg("pci-bus");
+    mk_run.addFileArg(pci_bus_exe.getEmittedBin());
+    mk_run.addArg("crash-test");
+    mk_run.addFileArg(crash_test_exe.getEmittedBin());
+    mk_run.addArg("device-list");
+    mk_run.addFileArg(device_list_exe.getEmittedBin());
+    mk_run.addArg("discovery");
+    mk_run.addFileArg(discovery_exe.getEmittedBin());
     mk_run.addArg("device-manager");
     mk_run.addFileArg(device_manager_exe.getEmittedBin());
     mk_run.addArg("input");

docs/README.md

@@ -45,20 +45,22 @@ rather than restate it. Roughly in the order things happen at runtime:
     until its hardware interrupts it**. The claim is the capability; `irq_ack` is the
     unmask.
 14. **[driver-model.md](driver-model.md) — buses, classes and host controllers.** How
-    real driver stacks factor into three shapes, how families share code, and the
-    proposed ABI for the three primitives still missing (capability passing, DMA +
-    memory barriers, MSI).
+    real driver stacks factor into three shapes and how families share code. The
+    three primitives it proposed are long since built (M13 capability passing,
+    M14 DMA + barriers, M15 MSI), and the driver *contract* on top of them —
+    hello, supervision, restart — is built too (device-manager.md, M18).
 15. **[process-management.md](process-management.md) — process management.** The
     microkernel's `ps`/`kill`/SIGCHLD: enumerate as a table snapshot, the
     supervision link as the kill authority, and child-exit notifications over the
     same endpoints IRQs arrive on.
-16. **[process-lifecycle.md](process-lifecycle.md) — the process lifecycle.** Design:
-    signals over IPC as the one lifecycle vocabulary every process speaks — the
+16. **[process-lifecycle.md](process-lifecycle.md) — the process lifecycle.** Built
+    (M17): signals over IPC as the one lifecycle vocabulary every process speaks — the
     POSIX.1-1990 words with message delivery instead of stack hijack, the stable
     `runtime.process` interface, exit reasons, published exit events any stateful
     service can subscribe to (the VFS releasing dead clients' handles), and the two
     iron rules (cleanup is the kernel's job; kill is not a signal).
-17. **[device-manager.md](device-manager.md) — the device manager.** Design: the
+17. **[device-manager.md](device-manager.md) — the device manager.** Built (M18,
+    through the app surface): the
     tree, the matcher, and the supervisor. Tree structure lives in the manager,
     authority stays in the kernel; bus drivers report what they see; drivers are
     restarted through the lifecycle vocabulary — the plan that turns

docs/device-manager.md

@@ -1,6 +1,16 @@
 # The device manager
 
-**Status: design.** The primitives this builds on are real ([process-management.md](process-management.md):
+**Status: the protocol and supervision are built** (M18.1, 2026-07-13): `hello`
+with its deadline, supervised spawn, restart with backoff, and the crash-loop
+cap are in — usb-xhci-bus is the first conforming driver, and the
+`driver-restart` scenario proves fault → backoff → re-claim → cap end to end.
+Tree reports are built too (M18.2, 2026-07-13): the xHCI driver scans its
+root-hub ports and reports each connected device (`child_added`); the manager
+mirrors them and prunes a dead reporter's children, and the `usb-report`
+scenario proves report → prune → respawn → re-report. The application surface is built (M18.3, 2026-07-13):
+`enumerate` and `subscribe` over IPC, with `device-list` as the first client —
+the manager is now the one answer to "what devices exist" for applications.
+The primitives underneath are real ([process-management.md](process-management.md):
 spawn/supervise/kill/exit-notification; [driver-model.md](driver-model.md): the device
 table as a capability system; [drivers.md](drivers.md): claim/map/IRQ), and the first
 per-device driver spawn works (the device manager matches the xHCI controller by PCI

docs/discovery.md

@@ -167,3 +167,13 @@ free; discovery on x86 is partly about *finding* what ARM just tells you.
 - [ipc.md](ipc.md) — the channels that interrupts-as-messages and the device manager
   will ride on.
 - [vision.md](vision.md) — why drivers belong in isolated user space at all.
+
+## Update (M19.3, 2026-07-13): PCI enumeration left the kernel
+
+The kernel now seeds only the `pci_host_bridge` node (ECAM window, MMIO
+apertures derived from the memory map's holes, bus range, and the 16-bit I/O
+window). The per-function walk moved to the ring-3 `pci-bus` driver
+([device-manager.md](device-manager.md)): it claims the bridge, repeats the
+ECAM scan through its mmio grant, and `device_register`s what it finds, which
+the device manager mirrors and matches. The ACPI namespace walk follows in M20;
+the static tables (MADT, HPET, MCFG, FADT + `\\_S5`) stay kernel-side.

docs/drivers.md

@@ -362,3 +362,23 @@ the first DMA driver to protect and test against) and these smaller items:
 - **Interrupt priority / threaded IRQ latency.** `notifyFromIsr` enqueues the woken
   driver but doesn't preempt (`wakeLocked` deliberately leaves that to the caller), so
   a woken driver waits for the next scheduling point.
+
+## The driver contract (M17–M18)
+
+Claiming and mapping is half of being a danos driver; the other half is the
+**lifecycle and protocol contract**, and the runtime makes it nearly free:
+
+- Build on `runtime.service.run` — one replyWait loop folding protocol
+  requests, signals, and notifications into callbacks. The harness answers the
+  universal zero-length ping and turns `terminate` into a clean exit for you
+  ([process-lifecycle.md](process-lifecycle.md)).
+- A driver spawned with an assignment (its device id as argv[1]) sends the
+  versioned `hello` to the device manager inside the deadline, and a **bus**
+  driver reports what it discovers with `child_added`
+  ([device-manager.md](device-manager.md); usb-xhci-bus is the reference
+  implementation).
+- Crash freely — that is the design. The kernel releases your claims, IRQ
+  bindings, and MSI vectors at death; the manager reads your exit reason,
+  prunes what you reported, restarts you with backoff, and your fresh instance
+  re-claims and re-reports. Never depend on your own cleanup running
+  (iron rule 1).

docs/ipc.md

@@ -103,3 +103,22 @@ This is what makes a user-space driver possible at all, and it's the subject of
   the oldest (discrete messages, not a coalescing level like the notification ring).
 - **A bounded reply.** `MSG_MAX` is 256 bytes and the copy runs under the big kernel
   lock; a bulk transfer wants shared pages, not a copy.
+
+## Lifecycle conventions over IPC (M17)
+
+Three conventions from [process-lifecycle.md](process-lifecycle.md) ride the
+notification mechanism:
+
+- **Signals** arrive as notifications on the endpoint a process nominated with
+  `signal_bind` (`runtime.process.bindSignals`): badge = the signal bit plus the
+  coalesced pending mask (`runtime.process.signalsFrom` decodes). Statements,
+  never questions; no payload, no reply.
+- **One-shot timers** (`timer_bind`, `runtime.system.timerOnce`) land as a
+  timer-bit notification — the timed wait: a service arms a deadline and keeps
+  serving, instead of blocking in sleep.
+- **The universal ping**: a **zero-length request is the liveness probe**,
+  answered with a zero-length reply by the service harness itself
+  (`runtime.service.run`). No protocol's requests start at length zero, so the
+  encoding cannot collide, and a wedged service simply fails to answer — which
+  is the diagnosis. Deep health ("can I reach my hardware?") stays a per-service
+  protocol message.

docs/m17-m18-plan.md

@@ -1,5 +1,9 @@
 # M17–M18 execution plan: process lifecycle + device manager
 
+**Archived — completed 2026-07-13** (every item checked; suite ended 54/54).
+Kept as the record of how M17–M18 landed; the successor is
+[m19-m20-plan.md](m19-m20-plan.md).
+
 The operational plan for building [process-lifecycle.md](process-lifecycle.md)
 (M17) and [device-manager.md](device-manager.md) increments 5–7 (M18). Design is
 settled in those documents; this file is the build order — one phase at a time,
@@ -21,6 +25,51 @@ run the existing QEMU suite green before any new work starts.
 
 **Numbering note:** continues the milestone sequence (driver track ended at M16).
 
+## Status
+
+The loop marks a phase `[x]` in the same commit that lands it. A phase is marked
+only when its definition of green holds.
+
+- [x] **Phase 0** — baseline: docs committed, feat/usb merged to main, pushed;
+      `usb-xhci-libary.zig` renamed to `usb-xhci-library.zig`; existing QEMU
+      suite green from the worktree (48/48, 2026-07-12).
+- [x] **M17.1** — kernel releases claims/MSI on death (claims: `releaseAllOwnedBy`
+      in the reap; MSI was already swept by `irq.releaseOwner`; `claim-release`
+      test; suite 49/49)
+- [x] **M17.2** — exit reasons (`ExitReason` recorded at exit/fault/kill before
+      the notification; `process_exit_reason` supervisor-gated;
+      `runtime.process.exitReason`; kernel + ring-3 assertions; suite 49/49)
+- [x] **M17.3** — published exit events + VFS subscriber (`process_subscribe`,
+      bounded ref-counted table, publish on every death;
+      `runtime.process.subscribeExits`; VFS handles carry owners and are swept on
+      the owner's death; `vfs-client-death` test; suite 50/50)
+- [x] **M17.4** — signals, timer notifications, `runtime.process`, the service
+      harness (signal_bind/process_signal + coalescing pending mask; timer_bind
+      on the tick; bindSignals/signalsFrom/sendSignal/stop + timerOnce;
+      runtime.service.run with the zero-length ping; VFS converted; `signals`
+      scenario; suite 51/51)
+- [x] **merge** `feat/process-lifecycle` → main, push (merged 2026-07-13)
+- [x] **M18.1** — device-manager protocol: hello + restart policy
+      (device-manager-protocol module; the manager as a harness service:
+      supervised spawns, hello deadline via timer sweep, restart with
+      300/600/1200ms backoff, exit reasons deciding restart-vs-stopped,
+      crash-loop cap; usb-xhci-bus first conforming driver; crash-test fixture
+      re-proving claim release each respawn; `driver-restart` scenario;
+      maximum_tasks 16→32 — the sweep was overflowing the pool; suite 52/52)
+- [x] **merge** `feat/device-manager` → main, push (merged 2026-07-13)
+- [x] **M18.2** — xHCI port scan + tree reports (child_added/child_removed in
+      the protocol; the manager's child mirror with death-pruning; xHCI maps the
+      register BAR — resource 0 is ECAM — reads CAPLENGTH/HCSPARAMS1, scans
+      PORTSC, reports connected ports with speed-class identity; `usb-report`
+      scenario proves report → prune → respawn → re-report; suite 53/53)
+- [x] **M18.3** — app surface: enumerate/subscribe over IPC (subscriber
+      endpoint rides as the call's capability; events are the same structs the
+      buses send); device-list first client; protocol capped at the kernel's
+      IPC MESSAGE_MAXIMUM (256); the startUserTask debug print removed — it
+      sheared concurrent serial lines and was the scenario-flake root cause;
+      `device-list` scenario; suite 54/54)
+- [x] **merge** `feat/usb-xhci-bus` → main, push (merged 2026-07-13) — **plan complete**
+
 ---
 
 ## M17.1 — the kernel releases a dead process's claims

docs/m19-m20-plan.md

@@ -0,0 +1,204 @@
+# M19–M20 execution plan: discovery migration
+
+The operational plan for [device-manager.md](device-manager.md)'s increment 8:
+discovery leaves the kernel — a **pci-bus driver** (M19) and an **acpi service**
+(M20), with the kernel's device enumeration retired behind them. Same rules as
+[m17-m18-plan.md](m17-m18-plan.md): one phase at a time, each green before the
+next; this file is the build order and the checklist.
+
+**Definition of green, every phase:** `zig build` clean, `zig build test` clean,
+`python3 test/qemu_test.py` passes (existing scenarios plus the phase's new
+one), and the relevant design doc updated. Commit per green phase (no co-author
+trailers). The full suite is the regression net — the existing
+`driver-restart` / `usb-report` / `device-list` / `input` scenarios must stay
+green *through* the migration, which is the whole point: the system must not be
+able to tell who enumerated it.
+
+**Workflow:** dedicated worktree; branches off `main` — `feat/pci-bus`
+(M19.0–19.3), `feat/acpi-service` (M20.1–20.3); auto-merge to main when a
+branch is green; keep branches; push everything.
+
+## Settled decisions (2026-07-13 — veto before the loop starts)
+
+1. **What "retiring the kernel scan" means.** The kernel keeps, forever, the
+   parses it needs before user space exists: RSDP/XSDT location, MADT (SMP),
+   the HPET table (the tick), FADT + the AML `\_S5` evaluation (poweroff — the
+   power tests prove it), and MCFG (the host bridge node). What retires is
+   **device enumeration**: the ECAM function walk (M19.3) and the DSDT/SSDT
+   namespace walk that builds device nodes (M20.3). The AML module stays a
+   shared build module compiled into both the kernel (for `\_S5`) and the acpi
+   service (for everything else) — same source, two builds, no fork.
+2. **Bridge apertures come from the firmware memory map, not AML.** Registered
+   PCI functions carry BAR resources, and containment demands the bridge own
+   windows that cover them. The apertures are derived kernel-side from the
+   boot memory map's MMIO holes (regions that are neither RAM nor tables) —
+   mechanical, AML-free, and available at boot regardless of what later moved
+   to user space. (The bridge today carries only ECAM + bus range; this is the
+   prerequisite M19.0 exists for.)
+3. **`device_register` becomes idempotent on exact match.** A re-registration
+   with identical (parent, class, resources) returns the existing id instead
+   of appending. The kernel table has no unregister, so without this a
+   restarted registering bus would duplicate its children on every respawn —
+   idempotence makes restart-and-re-report safe for every future bus, not just
+   PCI.
+4. **The manager matches from reports.** `ChildAdded` gains a `device_id`
+   field (the kernel-registered id, `no_device` for unregistered leaves like
+   USB ports). After the M19.3 flip, PCI driver matching keys off reported
+   identity (the class triple) instead of the manager's boot-time snapshot —
+   the snapshot match remains only for what the kernel still seeds. One flip
+   phase changes both sides at once so no device is ever matched twice.
+5. **The acpi service's authority is one node.** The kernel publishes an
+   `acpi-tables` device: memory resources covering the table blobs plus a
+   broad `io_port` resource — the documented trust grant to exactly one
+   process (AML OperationRegions reach EC/PM ports; the claim-gated
+   io_read/io_write calls already exist). The service claims it, maps the
+   tables, and runs the shared AML module in ring 3 behind a `Hal` backed by
+   `mmio_map` + `io_read`/`io_write`.
+6. **Both new processes are protocol drivers** under the manager: hello,
+   supervision, restart with backoff — all inherited from M18.1 for free.
+   Registration idempotence (decision 3) is what makes their restarts sound.
+7. **Firmware neutrality is the contract** (2026-07-13). The generic layer is
+   everything at and above the device-manager protocol — descriptors,
+   containment, reports, matching, supervision — and none of it may become
+   x86-specific. Discovery is one swappable process per firmware: the acpi
+   service on x86; an **fdt service** on the Raspberry Pis (claims a
+   `devicetree-blob` node, reports children from the flattened device tree —
+   pure data, no bytecode, no port grant, strictly simpler than ACPI). The
+   manager owns the tree as *data* and touches no hardware, ever — AML runs in
+   a crashable, supervised discoverer precisely so a firmware-bytecode fault
+   can never take down the supervisor. Two consequences recorded now:
+   `DeviceDescriptor`'s 8-byte `hid` cannot hold an FDT `compatible` string
+   ("brcm,bcm2835-aux-uart") — identity widens before the fdt service exists;
+   and cross-firmware surfaces are named by **domain, not firmware** (M21
+   defines a *power* protocol, not an "ACPI events" protocol — PSCI/mailbox
+   sources feed the same subscribers on ARM). **Landed early (2026-07-13):**
+   both services exist as placeholders (system/services/acpi, system/services/
+   fdt) and the build's `-Ddiscovery=acpi|fdt` option fills the ramdisk's
+   neutral `discovery` slot — the manager will spawn "discovery" by that name
+   in M20.3 and never learn which firmware it is on.
+
+## Status
+
+- [x] **M19.0** — prerequisites (bridge apertures from the memory map's
+      *gaps* — the single-hole rule died on OVMF's flash at the top of 4 GiB,
+      caught by the new every-BAR-contained assert in `discovery`; idempotent
+      `device_register` proven in `bus`; `ChildAdded.device_id`;
+      m17-m18-plan.md archived; suite 54/54).
+- [x] **M19.1** — pci-bus driver, scan only (claims the bridge, maps ECAM
+      through its grant, brute-force walk with the multifunction rule; the
+      manager matches pci_host_bridge → pci-bus per device with the full
+      protocol contract; `pci-scan` builds its expected marker from the
+      kernel's own count — equivalence on the first run; suite 55/55).
+- [x] **M19.2** — register + report (BAR probe mirrored byte-for-byte from the
+      kernel's addBars so dedupe returns the kernel's node ids during
+      coexistence; the bridge gained the io_port aperture I/O BARs need;
+      reports carry the registered device_id; pci-scan drills a forced restart
+      and asserts the PCI node count never grows — plus harness hardening: a
+      failing case now preserves its serial as <case>-failed-serial.log, and
+      the heavy scenarios run at 150s; suite 55/55).
+- [x] **M19.3** — the flip: kernel `enumeratePci`/`addBars`/`PciHeader` all
+      deleted (bridge node stays); manager matches PCI drivers from reported
+      identity, deduped by registered id. Surfaced and fixed a real SMP race the
+      flip created — ring-3 device_register made the broker table concurrent, so
+      mmio_map's lock-free read intermittently tore hpet's resource length
+      (user fault) and overflowed `r.len-1` into a kernel panic; now the broker
+      read is under the big lock and the arithmetic is guarded, and pci-bus
+      skips size-0 BARs. discovery.md updated; suite 55/55 (driver-restart
+      hammered 6×).
+- [ ] **merge** `feat/pci-bus` → main, push.
+- [ ] **M20.1** — acpi service, parse only (fills the existing placeholder at
+      system/services/acpi/acpi.zig): kernel publishes `acpi-tables`
+      (decision 5) — memory over the table blobs, the broad io_port grant, and
+      **the SCI as an irq resource** (from the FADT; unused until M21 but free
+      to record now). The service claims it, maps the blobs, runs the shared
+      AML module in ring 3, logs the namespace device count and `_HID`s.
+      Scenario `acpi-parse`: user-space count equals the kernel walk's count.
+- [ ] **M20.2** — register + report: namespace devices with `_HID` + `_CRS`
+      resources registered under `acpi-tables` (its io_port + the memory-map
+      holes give containment), reported to the manager. Spawn-from-reports for
+      ACPI matches stays off. Scenario: the reported set includes the PS/2
+      keyboard and mouse nodes with their IRQ resources.
+- [ ] **M20.3** — the flip: kernel DSDT device-node building removed (static
+      tables + `\_S5` stay, decision 1); manager matches ACPI-hid drivers
+      (ps2-bus) from reports. The `input` and `device-manager` scenarios are
+      the assertion. discovery.md + acpi.md + device-manager.md updated;
+      device-manager.md increment 8 closed.
+- [ ] **merge** `feat/acpi-service` → main, push — **loop ends here**.
+
+---
+
+## Phase notes
+
+**M19.0 apertures:** the boot memory map already crosses the handoff
+([boot-handoff]), but discovery never sees it today — expect a small
+pass-through (kernel init hands the map to the platform layer) before the
+holes computation, which belongs where the bridge node is built
+(`parseMcfg`). Sanity-check on QEMU q35: the xHCI BAR (`0xc0000000`-region
+values seen in the M18 logs) must land inside a derived aperture, asserted in
+the kernel unit test.
+
+**M19.1 scanning without owning config access twice:** the driver reads config
+space through its ECAM mmio_map grant of the *bridge* window — the same bytes
+the kernel walk read. Vendor-id `0xFFFF` skip, header-type multifunction rule,
+no bridge recursion (matches the kernel's current single-segment walk).
+
+**M19.2 BAR sizing:** the classic size probe (write all-ones, read mask,
+restore) is deferred — the BARs' current programmed values and types are
+enough for containment-checked registration at bring-up; sizing lands with the
+first driver that needs to *move* a BAR. Log what is registered so the
+scenario can assert it.
+
+**M19.3 what the manager still seeds from the snapshot:** everything the
+kernel still enumerates (timers, ACPI nodes until M20.3). The PCI arm of
+`pciDriverFor` switches source; `driverFor` doesn't move until M20.3.
+
+**M20.1 spawn and identity (pre-settled 2026-07-13):** the manager spawns
+`discovery` by its neutral ramdisk name at startup, as an ordinary protocol
+driver (hello, supervision) — from M20.1 on, on every boot. For reporting ACPI
+devices, `ChildAdded` gains `hid: [8]u8` (EISA ids fit; zero = none):
+firmware *string* identity travels beside the numeric `identity` field until
+the FDT-driven widening replaces both (decision 7).
+
+**M20.1 Hal in ring 3:** `mapMmio` → `device.mmioMap` over the claimed
+acpi-tables node (plus a table-offset map for blobs); `pioRead`/`pioWrite` →
+`device.ioRead`/`ioWrite` against its io_port resource. The interpreter cannot
+tell it moved — that is the assertion of `acpi-parse`.
+
+**M20.2 containment for `_CRS`:** io ports fall inside the node's broad
+io_port resource; MMIO windows (HPET, LAPIC ranges some firmwares list) fall
+inside the memory-map holes added to the node in M20.1. Anything that doesn't
+fit is logged and skipped, loudly — bring-up honesty over silent drops.
+
+**M20.3 ps2 ordering:** ps2-bus binds nodes the acpi service now reports, so
+its spawn moves behind the report (the manager's matching handles this once
+the source flips); the `input` scenario proves the keyboard still types.
+
+**Explicitly out of scope:** PCI bridge recursion (single segment, flat bus
+walk stays); BAR reprogramming/sizing; disk/PCIe hotplug; interrupt routing
+changes (`_PRT` stays wherever it is today); the USB descriptor track;
+multi-segment ECAM; per-device power states (D-states, `_PSx`/`_PRx`,
+suspend/resume — a future *lifecycle-vocabulary* extension, since "suspend"
+has the shape of a signal every driver must answer, and it has no consumer
+until laptop sleep); CPU P/C-states.
+
+## M21 preview — ACPI events + system power (planned next, not in this loop)
+
+The acpi service grows the event side (settled direction 2026-07-13; detailed
+phases when M20 lands):
+
+- **21.1 SCI + fixed events**: irq_bind the SCI (the resource M20.1 already
+  records), read/clear PM1 status, publish the power-button event to
+  subscribers (the same pub/sub shape the manager uses).
+- **21.2 GPE + Notify**: Notify dispatch in the shared AML interpreter, GPE
+  block handling, `Notify(device, code)` published per reported node. The
+  acpi service is a **bus** here: battery (PNP0C0A), AC (ACPI0003), and lid
+  (PNP0C0D) nodes are reported children; small class drivers bind them and
+  speak an evaluate/subscribe protocol to the service — the xHCI split,
+  repeated. The embedded controller (`_Qxx` queries) rides this phase;
+  QEMU emulates no battery/EC, so those paths are interface-complete and
+  validated on real hardware (the laptop is the win condition), while the
+  plumbing is proven by the power button.
+- **21.3 the capstone**: QEMU `system_powerdown` → acpi service event → init
+  runs the M17 stop sequence over its children → kernel `\_S5` — orderly
+  shutdown as the scenario that proves lifecycle + events compose. (The
+  harness grows a QMP poke to inject the event.)

docs/process-lifecycle.md

@@ -1,6 +1,9 @@
 # Process lifecycle: signals over IPC
 
-**Status: design.** The primitives underneath are built ([process-management.md](process-management.md):
+**Status: increments 1–4 built** (2026-07-12): claim release on death, exit
+reasons, published exit events, and signals + one-shot timers + the service
+harness are all in — the interface below is as-built. The primitives underneath
+predate this design ([process-management.md](process-management.md):
 spawn, the supervision link, kill, child-exit notifications); this document designs
 the layer above them — the standard vocabulary a danos process speaks about its own
 life, and the stable `runtime.process` interface that carries it. Nothing here is
@@ -245,13 +248,16 @@ pub fn stop(id: u32, deadline_ms: u64) void { ... }
 /// process_enumerate.
 pub fn subscribeExits(endpoint: usize) bool { ... }
 
-/// How a process ended — from the exit notification. What restart policy reads.
-pub const ExitReason = enum {
+/// How a process ended — queried after the exit notification (the kernel records
+/// it first, so the two never race). What restart policy reads. (Built in M17.2.)
+pub const ExitReason = enum(u8) {
     exited, // returned from main / clean exit
-    aborted, // abort() — deliberate self-termination (SIGABRT's ghost)
+    aborted, // abort() — deliberate self-termination (SIGABRT's ghost; reserved)
     segmentation_fault, // SIGSEGV's ghost
     illegal_instruction, // SIGILL's ghost
     arithmetic_fault, // SIGFPE's ghost
+    protection_fault, // general protection fault
+    fault, // any other CPU exception
     killed, // process_kill
 };
 ```

docs/process-management.md

@@ -95,11 +95,18 @@ the architecture layer calls up into `tick`.
 
 ## Known gaps (bring-up honesty)
 
-- Device **claims** are not released on death (pre-existing: the fault path has
-  the same gap) — a killed driver's device stays claimed until reboot.
+- ~~Device claims are not released on death~~ Closed (M17.1): every path out of a
+  process releases its device claims alongside its IRQ and MSI bindings
+  (`releaseTaskResourcesLocked`), so a restarted driver can claim its hardware
+  again — the cleanup half of [process-lifecycle.md](process-lifecycle.md)'s iron
+  rule 1. The `claim-release` test proves the kill → release → re-claim cycle.
 - Kernel stacks of dead tasks are leaked, as on every exit path (no reaper yet).
-- There is no exit *status* in the notification, only the id; a supervisor that
-  needs the code can grow a wait-style call later.
+- ~~There is no exit status in the notification~~ Closed (M17.2): the kernel
+  records how every process ends — exited, a fault class, or killed — before it
+  posts the exit notification, and the supervisor reads it with
+  `process_exit_reason` (`runtime.process.exitReason`). This is the input to
+  restart policy ([process-lifecycle.md](process-lifecycle.md)); an exit *code*
+  for the clean case can still ride alongside later.
 - Enumerate writes through the caller's raw pointer under the bring-up trust
   model, like `device_enumerate` (an unmapped page is a self-DoS, not an
   isolation break).
@@ -109,4 +116,5 @@ the architecture layer calls up into `tick`.
 `process-list` (enumerate), `process-kill` (kernel-level kill paths, refusals,
 notifications), `supervision` (the whole user-side surface via the process-test
 service: spawn supervised → enumerate → kill blocked and spinning children →
-notifications → gone). See test/qemu_test.py.
+notifications → gone), `claim-release` (a killed claim-holder's device is
+claimable again). See test/qemu_test.py.

docs/resilience.md

@@ -1,10 +1,17 @@
 # Resilience: fault isolation and live restart
 
-Steps 1–2 of the ordering below are **built**: user-mode isolation, and fault →
-kill the process → keep the core (`onException` in `system/kernel/kernel.zig`; the
-`fault-recovery` test proves a crashing ring-3 process dies alone while the system
-keeps running). The supervisor notification and restart policy (steps 3+) are
-still design. This is the property danos is really chasing:
+Steps 1–4 of the ordering below are **built** (M17–M18, 2026-07-13): user-mode
+isolation; fault → kill the process → keep the core (`onException`; the
+`fault-recovery` test); the supervisor notification **with exit reasons**
+([process-lifecycle.md](process-lifecycle.md) — clean exit, fault class, or
+killed, recorded before the notice posts); and the **restart policy itself**
+([device-manager.md](device-manager.md)): the device manager supervises every
+driver, restarts crashes with backoff, caps crash loops, and re-claims work
+because the kernel releases a dead process's claims. The `driver-restart` and
+`usb-report` scenarios prove kill → release → respawn → re-claim → re-report
+end to end. What remains of this document's ladder is scope, not mechanism:
+more of the system moved into restartable processes (the discovery migration,
+[m19-m20-plan.md](m19-m20-plan.md), is the next rung). This is the property danos is really chasing:
 **if a part of the OS breaks, isolate it, and re-initialise it — without rebooting.**
 A crashed driver gets restarted; a wedged service gets killed and brought back. It's
 the reason the [microkernel](vision.md) shape was chosen, and it's a *separate* goal

library/runtime/ipc.zig

@@ -94,6 +94,15 @@ pub fn send(h: Handle, message: []const u8) bool {
 /// GSI. See `isNotification`.
 pub const notify_badge_bit: u64 = abi.notify_badge_bit;
 
+/// Set alongside `notify_badge_bit` when the notification is a **signal** — the
+/// lifecycle vocabulary of docs/process-lifecycle.md, delivered to the endpoint
+/// nominated with `process.bindSignals`. Decode with `process.signalsFrom`.
+pub const notify_signal_bit: u64 = abi.notify_signal_bit;
+
+/// Set alongside `notify_badge_bit` when the notification is a **one-shot timer**
+/// landing (`system.timerOnce`).
+pub const notify_timer_bit: u64 = abi.notify_timer_bit;
+
 /// Set alongside `notify_badge_bit` when the notification is a **child-exit
 /// notice** — a process this one spawned (with an exit endpoint) has ended —
 /// rather than a device interrupt. The low bits carry the child's process id.
@@ -133,6 +142,17 @@ pub const Received = struct {
     }
 
     /// The task id of whoever posted a buffered message, meaningful only when
+    /// Whether this arrival is a signal notification — decode the set with
+    /// `process.signalsFrom(badge)`.
+    pub fn isSignal(self: Received) bool {
+        return self.isNotification() and self.badge & notify_signal_bit != 0;
+    }
+
+    /// Whether this arrival is a one-shot timer landing (`system.timerOnce`).
+    pub fn isTimer(self: Received) bool {
+        return self.isNotification() and self.badge & notify_timer_bit != 0;
+    }
+
     /// `isMessage`. (The badge's low bits, with the three high marker bits masked off.)
     pub fn senderTaskId(self: Received) u32 {
         return @intCast(self.badge & ~(notify_badge_bit | notify_exit_bit | notify_message_bit));

library/runtime/process.zig

@@ -1,8 +1,15 @@
-//! Process-level runtime types: what a user program receives at entry. Mirrors
+//! Process-level runtime types: what a user program receives at entry (`Init`,
+//! the argv contract) and the process end of the lifecycle
+//! (docs/process-lifecycle.md) — today the exit reason a supervisor reads to
+//! decide restart; signals and the stop sequence land here with M17.4. Mirrors
 //! the spirit of `std.process.Init.Minimal` in danos terms — std's `Args` holds
 //! no data on freestanding targets, so the type is danos's own.
 
 const std = @import("std");
+const abi = @import("abi");
+const sc = @import("system-call.zig");
+const ipc = @import("ipc.zig");
+const system = @import("system.zig");
 
 /// Everything a program receives at entry. Passed to
 /// `pub fn main(init: runtime.process.Init)`; programs that need nothing keep
@@ -45,3 +52,86 @@ pub const Arguments = struct {
         }
     };
 };
+
+/// How a process ended — what a supervisor's restart policy reads: a clean exit
+/// meant to stop, a fault wants a restart with backoff, killed means the
+/// supervisor did it itself (docs/process-lifecycle.md).
+pub const ExitReason = abi.ExitReason;
+
+/// How dead child `id` ended. Ask after the exit notification arrives — the
+/// kernel records the reason before it posts the notification, so this never
+/// races it. Returns null for an id that never lived, is still alive, was
+/// evicted from the kernel's bounded record, or is not this process's child
+/// (the same authority gate as `kill`).
+pub fn exitReason(id: u32) ?ExitReason {
+    const r = sc.systemCall1(.process_exit_reason, id);
+    if (r > ~@as(usize, 0) - 4095) return null; // a wrapped -errno
+    return @enumFromInt(r);
+}
+
+/// The signal vocabulary (docs/process-lifecycle.md): POSIX's concepts, danos's
+/// names, message delivery. A signal is a one-way coalescing statement — never a
+/// question (liveness is the zero-length ping call) and never kill (that is
+/// `system.kill`, unhandleable by definition).
+pub const Signal = abi.Signal;
+
+/// The coalesced set of signals one notification delivered: two pending
+/// terminates arrive as one. Decode a received badge with `signalsFrom`.
+pub const SignalSet = struct {
+    pending: u32,
+
+    pub fn has(set: SignalSet, signal: Signal) bool {
+        return set.pending & (@as(u32, 1) << @intFromEnum(signal)) != 0;
+    }
+};
+
+/// Nominate `endpoint` as this process's signal endpoint. Signals posted while
+/// unbound have pended; they are delivered immediately on bind, coalesced.
+pub fn bindSignals(endpoint: usize) bool {
+    return sc.systemCall1(.signal_bind, endpoint) == 0;
+}
+
+/// Decode a received badge into the signals it delivered, or null if it is not
+/// a signal notification.
+pub fn signalsFrom(badge: u64) ?SignalSet {
+    if (badge & abi.notify_badge_bit == 0 or badge & abi.notify_signal_bit == 0) return null;
+    return .{ .pending = @truncate(badge & ~(abi.notify_badge_bit | abi.notify_signal_bit)) };
+}
+
+/// Post `signal` to child `id` (or to yourself). Supervisor-gated, like kill;
+/// non-blocking, always — a statement, not a conversation.
+pub fn sendSignal(id: u32, signal: Signal) bool {
+    return sc.systemCall2(.process_signal, id, @intFromEnum(signal)) == 0;
+}
+
+/// The standard stop sequence (docs/process-lifecycle.md): terminate, wait up to
+/// `deadline_ms` for the exit notification on `exit_endpoint` (the endpoint the
+/// child was spawned with), then kill. Any *other* notifications arriving on
+/// that endpoint while stopping are consumed and dropped — a supervisor with
+/// concurrent traffic implements the same sequence inside its own event loop
+/// (arm `system.timerOnce`, keep serving) instead of calling this.
+pub fn stop(id: u32, deadline_ms: u64, exit_endpoint: usize) void {
+    _ = sendSignal(id, .terminate);
+    _ = system.timerOnce(exit_endpoint, deadline_ms);
+    var receive: [8]u8 = undefined;
+    while (true) {
+        const got = ipc.replyWait(exit_endpoint, &.{}, &receive, null);
+        if (got.isChildExit() and got.childProcessId() == id) return;
+        if (got.isTimer()) break; // the deadline passed first — escalate
+    }
+    _ = system.kill(id);
+    while (true) {
+        const got = ipc.replyWait(exit_endpoint, &.{}, &receive, null);
+        if (got.isChildExit() and got.childProcessId() == id) return;
+    }
+}
+
+/// Subscribe `endpoint` to published exit events: every process death posts an
+/// asynchronous notification with the same badge encoding as a supervisor's exit
+/// notice (decode with `ipc.Received.isChildExit`/`childProcessId`). For stateful
+/// services: release what the dead client held — file handles, subscriptions —
+/// because a service must never depend on clients cleaning up after themselves
+/// (docs/process-lifecycle.md). Ungated, like `system.processes`.
+pub fn subscribeExits(endpoint: usize) bool {
+    return sc.systemCall1(.process_subscribe, endpoint) == 0;
+}

library/runtime/runtime.zig

@@ -17,6 +17,9 @@ pub const ipc = @import("ipc.zig");
 pub const start = @import("start.zig");
 /// The VFS wire protocol (shared with the VFS server).
 pub const vfs_protocol = @import("vfs-protocol");
+
+/// The device-manager protocol: hello + tree reports (docs/device-manager.md).
+pub const device_manager_protocol = @import("device-manager-protocol");
 /// Keyboard-event listening (subscribe/next) and broadcasting (publish), over the input
 /// service. See library/runtime/input.zig and system/services/input/.
 pub const input = @import("input.zig");
@@ -35,5 +38,9 @@ pub const panic = start.panic;
 /// Process entry types: the `Init` handed to `main`, and its `Arguments`.
 pub const process = @import("process.zig");
 
+/// The service harness: one replyWait loop folding requests, signals, and
+/// notifications into callbacks (docs/process-lifecycle.md).
+pub const service = @import("service.zig");
+
 /// The heap as a `std.mem.Allocator`, for Zig `std` containers in user code.
 pub const allocator = heap.allocator;

library/runtime/service.zig

@@ -0,0 +1,83 @@
+//! The service harness (docs/process-lifecycle.md): one replyWait loop that
+//! folds protocol requests, signals, and subscribed notifications into
+//! callbacks — so the lifecycle contract ("answers ping, exits on terminate")
+//! is satisfied by construction and a service author writes domain logic only.
+//! Nothing is asynchronous inside the process: a callback runs at a point the
+//! loop chose, never on a hijacked stack — the whole reason signals are
+//! messages.
+//!
+//! The liveness probe: a **zero-length request is the universal ping**, answered
+//! with a zero-length reply by the harness itself. No protocol's requests start
+//! at length zero, so the encoding cannot collide, and there is nothing for a
+//! service author to implement — a wedged service simply fails to answer, which
+//! is the diagnosis (see docs/ipc.md).
+
+const abi = @import("abi");
+const ipc = @import("ipc.zig");
+const process = @import("process.zig");
+
+pub const Callbacks = struct {
+    /// Called once with the service's endpoint before the loop starts — the
+    /// place to subscribe to exit events, bind IRQs, or announce readiness.
+    /// Return false to abort startup (the process exits).
+    init: ?*const fn (endpoint: ipc.Handle) bool = null,
+    /// One protocol request from `sender` (a task id): write the reply into
+    /// `reply`, return its length. `capability` is the handle the request
+    /// carried, if any (M13 cap passing — how a subscriber hands over its
+    /// endpoint). The zero-length ping never reaches this.
+    on_message: *const fn (message: []const u8, reply: []u8, sender: u32, capability: ?ipc.Handle) usize,
+    /// A notification that is not a signal — a subscribed exit event, a bound
+    /// IRQ, a timer landing. The raw badge; decode with the ipc helpers.
+    on_notification: ?*const fn (badge: u64) void = null,
+    /// The reload signal. Default: ignored.
+    on_reload: ?*const fn () void = null,
+    /// The terminate signal, called before the loop returns. The clean exit is
+    /// the return itself — never put *necessary* work here (iron rule 1: a kill
+    /// arrives with no warning; this is for graceful extras only).
+    on_terminate: ?*const fn () void = null,
+    /// Publish the endpoint under a well-known service id at startup.
+    service: ?abi.ServiceId = null,
+};
+
+/// Run the service: create and (optionally) register the endpoint, bind signals
+/// to it, call `init`, then serve until `terminate` arrives — at which point the
+/// loop returns and main's return is the clean exit the supervisor reads as
+/// `ExitReason.exited`. `maximum_message` sizes the receive and reply buffers
+/// (a service passes its protocol's message maximum).
+pub fn run(comptime maximum_message: usize, callbacks: Callbacks) void {
+    const endpoint = ipc.createIpcEndpoint() orelse return;
+    if (callbacks.service) |id| {
+        if (!ipc.register(id, endpoint)) return;
+    }
+    _ = process.bindSignals(endpoint);
+    if (callbacks.init) |initialise| {
+        if (!initialise(endpoint)) return;
+    }
+
+    var reply_buffer: [maximum_message]u8 = undefined;
+    var reply_len: usize = 0;
+    var receive: [maximum_message]u8 = undefined;
+    while (true) {
+        const got = ipc.replyWait(endpoint, reply_buffer[0..reply_len], &receive, null);
+        if (got.isNotification()) {
+            reply_len = 0; // nothing owed for a notification
+            if (process.signalsFrom(got.badge)) |signals| {
+                if (signals.has(.reload)) {
+                    if (callbacks.on_reload) |onReload| onReload();
+                }
+                if (signals.has(.terminate)) {
+                    if (callbacks.on_terminate) |onTerminate| onTerminate();
+                    return; // the loop's return IS the clean exit
+                }
+                continue;
+            }
+            if (callbacks.on_notification) |onNotification| onNotification(got.badge);
+            continue;
+        }
+        if (got.len == 0) {
+            reply_len = 0; // the universal ping: a zero-length reply, from the harness
+            continue;
+        }
+        reply_len = callbacks.on_message(receive[0..got.len], &reply_buffer, got.senderTaskId(), got.cap);
+    }
+}

library/runtime/system.zig

@@ -32,6 +32,15 @@ pub fn sleep(ms: usize) void {
     _ = sc.systemCall1(.sleep, ms);
 }
 
+/// Arm a one-shot timer: after `ms` milliseconds the kernel posts a timer
+/// notification (`ipc.Received.isTimer`) to `endpoint`. The timed wait of
+/// docs/process-lifecycle.md — a service arms a deadline and keeps serving,
+/// instead of blocking in sleep; what stop-sequence escalation, hello deadlines,
+/// and restart backoff are built from.
+pub fn timerOnce(endpoint: usize, ms: u64) bool {
+    return sc.systemCall2(.timer_bind, endpoint, ms) == 0;
+}
+
 /// Monotonic nanoseconds since boot — a time source for timeouts and short delays. It
 /// only ever moves forward. This is *not* wall-clock time (no date, no timezone — that
 /// is a user-space service layered on top). Deadline pattern for a bounded poll loop:

system/abi.zig

@@ -53,9 +53,31 @@ pub const SystemCall = enum(u64) {
     process_enumerate = 24, // process_enumerate(buffer, maximum) -> total: snapshot the task table
     process_kill = 25, // process_kill(id) -> 0/-errno: end a process this process spawned
     ipc_send = 26, // ipc_send(handle, message_ptr, message_len) -> 0/-errno: post a payload to an endpoint's async queue without blocking
+    process_exit_reason = 27, // process_exit_reason(id) -> ExitReason/-errno: how a dead child ended (its supervisor only)
+    process_subscribe = 28, // process_subscribe(endpoint) -> 0/-errno: subscribe to published exit events — every death posts a notification
+    signal_bind = 29, // signal_bind(endpoint) -> 0/-errno: nominate the endpoint this process's signals arrive on
+    process_signal = 30, // process_signal(id, signal) -> 0/-errno: post a signal to a child (or to yourself)
+    timer_bind = 31, // timer_bind(endpoint, ms) -> 0/-errno: one-shot timer — posts a notification when ms elapse
     _,
 };
 
+/// How a process ended — recorded by the kernel at death, queried by the
+/// supervisor with `process_exit_reason`, and the input to its restart decision
+/// (docs/process-lifecycle.md): a clean exit meant to stop, a fault wants a
+/// restart with backoff, killed means the supervisor did it itself. The faults
+/// mirror the CPU exceptions a ring-3 process can die of; they are exit reasons,
+/// never delivered to the faulting process (recovery is restart, not a handler).
+pub const ExitReason = enum(u8) {
+    exited = 0, // returned from main / called exit
+    aborted = 1, // deliberate self-termination (reserved: no abort path yet)
+    segmentation_fault = 2, // page fault
+    illegal_instruction = 3, // invalid opcode
+    arithmetic_fault = 4, // divide error, x87 or SIMD fault
+    protection_fault = 5, // general protection fault
+    fault = 6, // any other CPU exception
+    killed = 7, // process_kill
+};
+
 /// The x86 MSI message address base (`0xFEE0_0000`): a device raises an MSI by writing
 /// `data` to this address, which the Local APIC turns into an interrupt at the vector
 /// in `data`. The kernel returns the concrete (address, data) from `msi_bind`; this is
@@ -93,6 +115,35 @@ pub const notify_exit_bit: u64 = 1 << 62;
 /// broadcasts where a rendezvous is the wrong shape (the input service is the first user).
 pub const notify_message_bit: u64 = 1 << 61;
 
+/// Set (alongside `notify_badge_bit`) in the badge of a **signal notification** —
+/// the process-lifecycle vocabulary of docs/process-lifecycle.md, delivered to the
+/// endpoint the process nominated with `signal_bind`. The low bits carry the
+/// coalesced pending mask (bit positions = `Signal` values): signals are
+/// statements, not questions, and two pending terminates are one terminate.
+pub const notify_signal_bit: u64 = 1 << 60;
+
+/// Set (alongside `notify_badge_bit`) in the badge of a **timer notification** —
+/// a one-shot `timer_bind` deadline landing. No payload bits: what to do when the
+/// deadline fires is whatever the receiver armed it for (a stop-sequence
+/// escalation, a restart backoff, an alarm).
+pub const notify_timer_bit: u64 = 1 << 59;
+
+/// The signal vocabulary (docs/process-lifecycle.md): POSIX's concepts, danos's
+/// names, message delivery. The value is the bit position in the pending mask — a
+/// private kernel/runtime detail, free to change while they ship together. Kill
+/// is not here (it is `process_kill`, unhandleable by definition); faults are not
+/// here (they are `ExitReason`s — recovery is restart, not a handler); liveness is
+/// not here (a question, asked as the zero-length ping call, not a statement).
+pub const Signal = enum(u5) {
+    terminate = 0, // finish up and exit (the polite half of the stop sequence)
+    reload = 1, // re-read configuration / re-scan
+    interrupt = 2, // interactive interrupt (no sender until a console exists)
+    quit = 3, // as interrupt, by convention more final
+    alarm = 4, // a timer the process armed for itself (unbuilt: no consumer yet)
+    user_1 = 5, // service-defined
+    user_2 = 6, // service-defined
+};
+
 /// Capacity of `ProcessDescriptor.name` — matches the longest name `system_spawn`
 /// accepts, so a process's recorded name (its argv[0]) is never truncated.
 pub const maximum_process_name = 64;
@@ -125,6 +176,7 @@ pub const ServiceId = enum(u32) {
     vfs = 1,
     input = 2,
     ps2_bus = 3, // the 8042 owner; child device drivers attach here for raw bytes
+    device_manager = 4, // the tree, the matcher, the supervisor (docs/device-manager.md)
     _,
 };
 

system/devices/acpi.zig

@@ -360,32 +360,14 @@ const Hpet = extern struct {
     page_protection: u8,
 };
 
-// --- PCI configuration-space header (first 64 bytes, common fields) ---------
-
-const PciHeader = extern struct {
-    vendor_id: u16 align(1),
-    device_id: u16 align(1),
-    command: u16 align(1),
-    status: u16 align(1),
-    revision_id: u8,
-    prog_if: u8,
-    subclass: u8,
-    class_code: u8,
-    cache_line_size: u8,
-    latency_timer: u8,
-    /// bit 7 set => multi-function device.
-    header_type: u8,
-    bist: u8,
-    // 0x10 onward (BARs, etc.) depends on header_type; read separately.
-};
-
 // --- Entry point ------------------------------------------------------------
 
 /// Discover hardware from the ACPI tables rooted at `rsdp_physical` and populate
 /// `device_tree`. `hal` provides MMIO mapping (for PCIe ECAM) and port I/O. Also parses the
 /// FADT and the AML sleep-state (`_Sx`) packages into `power_information` for the power service.
-pub fn discover(rsdp_physical: u64, device_tree: *DeviceTree, hal: Hal) !void {
+pub fn discover(rsdp_physical: u64, memory_regions: []const boot_handoff.MemoryRegion, device_tree: *DeviceTree, hal: Hal) !void {
     if (rsdp_physical == 0) return error.NoRsdp;
+    boot_memory_regions = memory_regions;
 
     // Start clean so a re-run doesn't accumulate stale state.
     power_information = .{};
@@ -451,7 +433,7 @@ fn handleTable(device_tree: *DeviceTree, hal: Hal, sdt_physical: u64) !void {
     if (std.mem.eql(u8, &sig, &APIC)) {
         try parseMadt(device_tree, header);
     } else if (std.mem.eql(u8, &sig, &MCFG)) {
-        try parseMcfg(device_tree, hal, header);
+        try parseMcfg(device_tree, header);
     } else if (std.mem.eql(u8, &sig, &HPET)) {
         try parseHpet(device_tree, hal, header);
     } else if (std.mem.eql(u8, &sig, &FACP)) {
@@ -536,7 +518,7 @@ fn parseMadt(device_tree: *DeviceTree, header: *const SystemDescriptorTableHeade
 }
 
 /// MCFG -> a pci_host_bridge per ECAM segment, then a PCI enumeration underneath.
-fn parseMcfg(device_tree: *DeviceTree, hal: Hal, header: *const SystemDescriptorTableHeader) !void {
+fn parseMcfg(device_tree: *DeviceTree, header: *const SystemDescriptorTableHeader) !void {
     const total: usize = header.length;
     const base: [*]const u8 = @ptrCast(header);
 
@@ -551,109 +533,85 @@ fn parseMcfg(device_tree: *DeviceTree, hal: Hal, header: *const SystemDescriptor
         // ECAM window: 1 MiB of configuration space per bus.
         _ = bridge.addResource(.memory, alloc.base_address, bus_count << 20);
         _ = bridge.addResource(.bus_range, alloc.start_bus, bus_count);
+        addBridgeApertures(bridge);
+        // The bridge decodes the whole 16-bit I/O space toward its bus — the
+        // window functions' I/O BARs must register-contain within (M19.2).
+        _ = bridge.addResource(.io_port, 0, 1 << 16);
 
-        try enumeratePci(device_tree, bridge, hal, alloc.*);
+        // The function walk itself retired to ring 3 (M19.3): the pci-bus
+        // driver claims this bridge, repeats the scan through its ECAM grant,
+        // and device_registers what it finds — the kernel seeds only the
+        // bridge. The scan's equivalence was proven before the hand-off
+        // (pci-scan), and the walk's history is in git if archaeology calls.
     }
 }
 
-/// Brute-force scan the ECAM window's bus range for present PCI functions. No
-/// bridge recursion yet: on the ECAM path the host bridge decodes every bus in
-/// the window, so scanning the declared range finds everything QEMU exposes.
-fn enumeratePci(
-    device_tree: *DeviceTree,
-    bridge: *device_model.Device,
-    hal: Hal,
-    alloc: McfgAllocation,
-) !void {
-    var bus: u16 = alloc.start_bus;
-    while (bus <= alloc.end_bus) : (bus += 1) {
-        var device: u8 = 0;
-        while (device < 32) : (device += 1) {
-            const h0: *align(1) const PciHeader = @ptrCast(pciConfigurationPtr(alloc, hal, @intCast(bus), device, 0));
-            if (h0.vendor_id == 0xFFFF) continue; // no function 0 => slot empty
+/// The boot memory map, stored at discover() entry for the aperture derivation
+/// below (and, in M20, for the acpi-tables node's containment windows).
+var boot_memory_regions: []const boot_handoff.MemoryRegion = &.{};
 
-            const funcs: u8 = if (h0.header_type & 0x80 != 0) 8 else 1;
-            var function: u8 = 0;
-            while (function < funcs) : (function += 1) {
-                const configuration = pciConfigurationPtr(alloc, hal, @intCast(bus), device, function);
-                const h: *align(1) const PciHeader = @ptrCast(configuration);
-                if (h.vendor_id == 0xFFFF) continue;
-
-                var nb: [24]u8 = undefined;
-                const nm = std.fmt.bufPrint(&nb, "{s}:{x:0>2}:{x:0>2}.{d}", .{
-                    bridge.name(), bus, device, function,
-                }) catch "pcidev";
-                const node = try device_tree.addChild(bridge, .pci_device, nm);
-                // Resource 0 is the function's own 4 KiB ECAM configuration space. A
-                // claimed PCI driver mmio_maps this to reach its command register,
-                // BARs, and — the point — its capability list (MSI/MSI-X, PCIe
-                // extended caps), without any new syscall. Physical address per the
-                // ECAM formula (same as pciConfigurationPtr).
-                const config_physical = alloc.base_address +
-                    (@as(u64, @as(u8, @intCast(bus)) - alloc.start_bus) << 20) +
-                    (@as(u64, device) << 15) + (@as(u64, function) << 12);
-                _ = node.addResource(.memory, config_physical, abi.page_size);
-                node.ids.pci_vendor = h.vendor_id;
-                node.ids.pci_device = h.device_id;
-                node.ids.pci_class = (@as(u24, h.class_code) << 16) |
-                    (@as(u24, h.subclass) << 8) | h.prog_if;
-                node.ids.pci_bdf = (@as(u16, @intCast(bus)) << 8) | (@as(u16, device) << 3) | function;
-
-                // BARs only exist in header type 0 (normal devices), not bridges.
-                if (h.header_type & 0x7F == 0) addBars(node, configuration);
+/// The bridge's MMIO apertures, derived from the boot memory map's holes
+/// (docs/m19-m20-plan.md decision 2): registered PCI functions carry BAR
+/// resources, and `device_register` containment demands the bridge own windows
+/// that cover them. Everything the firmware described is "not hole"; the low
+/// aperture runs from the end of the described space below 4 GiB up to the
+/// I/O-APIC region, the high one from 4 GiB (or the end of RAM above it) to
+/// the 46-bit line. Coarse, mechanical, and AML-free — available at boot no
+/// matter what later moved to user space.
+fn addBridgeApertures(bridge: *device_model.Device) void {
+    // Below 4 GiB the described regions are sparse (RAM low, firmware flash
+    // and tables high), so the holes are the *gaps between* them — a single
+    // "after the last region" rule dies on OVMF's flash at the very top.
+    // Sort-merge the described ranges, then keep the three largest gaps
+    // (resource slots are bounded at 8 per device; ECAM + bus range + 3 + the
+    // high aperture fits). Above 4 GiB one aperture runs from the end of the
+    // described space to the 46-bit line.
+    const Range = struct { base: u64, end: u64 };
+    var below: [64]Range = undefined;
+    var below_count: usize = 0;
+    var high_end: u64 = 1 << 32;
+    for (boot_memory_regions) |region| {
+        const end = region.base + region.pages * 4096;
+        // Above 4 GiB only *usable RAM* blocks the aperture: OVMF describes
+        // its own 64-bit PCI window as a reserved region and then programs
+        // BARs inside it — honoring reserved there would exclude the very
+        // space BARs live in. Below 4 GiB every described region blocks (the
+        // kernel image, the tables, the ramdisk all live there). Bring-up
+        // trust: only the bridge's claimant can register into the aperture.
+        if (region.kind == .usable and end > high_end) high_end = end;
+        if (region.base >= (1 << 32) or below_count == below.len) continue;
+        below[below_count] = .{ .base = region.base, .end = @min(end, 1 << 32) };
+        below_count += 1;
+    }
+    // Insertion sort by base (the map is small and this runs once at boot).
+    for (1..below_count) |i| {
+        const key = below[i];
+        var j = i;
+        while (j > 0 and below[j - 1].base > key.base) : (j -= 1) below[j] = below[j - 1];
+        below[j] = key;
+    }
+    // Walk the sorted ranges, collecting inter-region gaps of at least 1 MiB.
+    var gaps: [3]Range = .{Range{ .base = 0, .end = 0 }} ** 3;
+    var cursor: u64 = 0;
+    var index: usize = 0;
+    while (index <= below_count) : (index += 1) {
+        const gap_end = if (index == below_count) (1 << 32) else below[index].base;
+        if (gap_end > cursor and gap_end - cursor >= (1 << 20)) {
+            // Keep the three largest, replacing the smallest kept so far.
+            var smallest: usize = 0;
+            for (gaps, 0..) |gap, gi| {
+                if (gap.end - gap.base < gaps[smallest].end - gaps[smallest].base) smallest = gi;
+            }
+            if (gap_end - cursor > gaps[smallest].end - gaps[smallest].base) {
+                gaps[smallest] = .{ .base = cursor, .end = gap_end };
             }
         }
+        if (index < below_count and below[index].end > cursor) cursor = below[index].end;
     }
-}
-
-/// Record and size the memory/IO windows named by a device's Base Address
-/// Registers. Sizing is the standard probe: disable decode, write all-ones, read
-/// back the writable (address) bits, restore. `size = ~mask + 1`.
-fn addBars(node: *device_model.Device, configuration: [*]align(1) u8) void {
-    // Stop the device decoding its BARs while we transiently write all-ones.
-    const command = rd(u16, configuration, 0x04);
-    wr(u16, configuration, 0x04, command & ~@as(u16, 0b11));
-
-    var i: usize = 0;
-    while (i < 6) : (i += 1) {
-        const off = 0x10 + i * 4;
-        const orig = rd(u32, configuration, off);
-        if (orig == 0) continue;
-
-        if (orig & 1 != 0) {
-            // I/O-space BAR (16-bit address space on x86).
-            wr(u32, configuration, off, 0xFFFF_FFFF);
-            const readback = rd(u32, configuration, off);
-            wr(u32, configuration, off, orig);
-            const mask = readback & 0xFFFF_FFFC;
-            const size: u32 = if (mask == 0) 0 else (~mask +% 1) & 0xFFFF;
-            _ = node.addResource(.io_port, orig & 0xFFFF_FFFC, size);
-        } else if ((orig >> 1) & 0x3 == 2) {
-            // 64-bit memory BAR: this BAR pair spans two configuration slots.
-            const orig_hi = rd(u32, configuration, off + 4);
-            wr(u32, configuration, off, 0xFFFF_FFFF);
-            wr(u32, configuration, off + 4, 0xFFFF_FFFF);
-            const lo = rd(u32, configuration, off);
-            const hi = rd(u32, configuration, off + 4);
-            wr(u32, configuration, off, orig);
-            wr(u32, configuration, off + 4, orig_hi);
-            const readback = (@as(u64, hi) << 32) | (lo & 0xFFFF_FFF0);
-            const size: u64 = if (readback == 0) 0 else ~readback +% 1;
-            const address = (@as(u64, orig_hi) << 32) | (orig & 0xFFFF_FFF0);
-            _ = node.addResource(.memory, address, size);
-            i += 1; // consumed the high half
-        } else {
-            // 32-bit memory BAR.
-            wr(u32, configuration, off, 0xFFFF_FFFF);
-            const readback = rd(u32, configuration, off);
-            wr(u32, configuration, off, orig);
-            const mask = readback & 0xFFFF_FFF0;
-            const size: u32 = if (mask == 0) 0 else ~mask +% 1;
-            _ = node.addResource(.memory, orig & 0xFFFF_FFF0, size);
-        }
+    for (gaps) |gap| {
+        if (gap.end > gap.base) _ = bridge.addResource(.memory, gap.base, gap.end - gap.base);
     }
-
-    wr(u16, configuration, 0x04, command); // restore decode
+    _ = bridge.addResource(.memory, high_end, (@as(u64, 1) << 46) - high_end);
 }
 
 /// HPET -> a timer node with its register block as an MMIO resource, plus the GSI
@@ -1186,16 +1144,6 @@ fn readCntRegister(base: [*]align(1) const u8, len: usize, xoff: usize, legacy_o
 
 /// The mapped configuration space of one PCI function (its 4 KiB ECAM page). Mapped
 /// writable so BAR sizing can probe it; reads and writes both go through here.
-fn pciConfigurationPtr(alloc: McfgAllocation, hal: Hal, bus: u8, device: u8, function: u8) [*]align(1) u8 {
-    const physical = alloc.base_address +
-        (@as(u64, bus - alloc.start_bus) << 20) +
-        (@as(u64, device) << 15) +
-        (@as(u64, function) << 12);
-    // Map the configuration page (writable, for BAR sizing) and use the virtual
-    // address the HAL hands back.
-    return @ptrFromInt(hal.mapMmio(physical, abi.page_size, true));
-}
-
 /// Read a little-endian integer at `off` from a (possibly unaligned) byte pointer.
 /// x86 is little-endian and native, so an unaligned load suffices.
 fn rd(comptime T: type, bytes: [*]align(1) const u8, off: usize) T {
@@ -1203,12 +1151,6 @@ fn rd(comptime T: type, bytes: [*]align(1) const u8, off: usize) T {
     return p.*;
 }
 
-/// Write a little-endian integer at `off` through a (possibly unaligned) pointer.
-fn wr(comptime T: type, bytes: [*]align(1) u8, off: usize, value: T) void {
-    const p: *align(1) T = @ptrCast(bytes + off);
-    p.* = value;
-}
-
 // --- tests ------------------------------------------------------------------
 
 test "eisaIdToStr decodes a packed EISA id" {

system/devices/platform.zig

@@ -72,7 +72,8 @@ pub fn discover(
     var device_tree = try DeviceTree.init(allocator);
 
     if (boot_information.acpi_rsdp != 0) {
-        try acpi.discover(boot_information.acpi_rsdp, &device_tree, hal);
+        const memory_regions = @as([*]const boot_handoff.MemoryRegion, @ptrFromInt(boot_handoff.physicalToVirtual(boot_information.memory_map.regions)))[0..boot_information.memory_map.len];
+        try acpi.discover(boot_information.acpi_rsdp, memory_regions, &device_tree, hal);
     } else {
         // No ACPI RSDP. A device-tree boot would parse its blob here; today that
         // path is a stub, so this reports the machine described itself no way we

system/drivers/pci-bus/pci-bus.zig

@@ -0,0 +1,253 @@
+//! /system/drivers/pci-bus — the PCI bus driver: enumeration moved out of ring 0
+//! (docs/m19-m20-plan.md, M19). The device manager matches the `pci_host_bridge`
+//! node and spawns one instance per bridge, the bridge's device id as argv[1] —
+//! the same per-device contract as usb-xhci-bus.
+//!
+//! M19.1 (this increment): claim the bridge, map its ECAM window (resource 0;
+//! the bus range and the MMIO apertures follow it), walk every
+//! bus/device/function config header, and log what the walk finds — ending
+//! with "pci-bus: N functions found", which the `pci-scan` scenario compares
+//! against the kernel's own enumeration. Registration and reports (M19.2), and
+//! the kernel walk's retirement (M19.3), build on this proven-equivalent scan.
+
+const std = @import("std");
+const runtime = @import("runtime");
+const protocol = runtime.device_manager_protocol;
+const device = runtime.device;
+
+fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
+    var line: [128]u8 = undefined;
+    _ = runtime.system.write(std.fmt.bufPrint(&line, fmt, arguments) catch return);
+}
+
+var bridge_id: u64 = protocol.no_device;
+var ecam_base: usize = 0;
+var ecam_physical: u64 = 0;
+var start_bus: u64 = 0;
+var bus_count: u64 = 0;
+var manager_handle: runtime.ipc.Handle = 0;
+
+/// One aligned 32-bit read from a function's configuration space.
+fn configRead(bus: u64, dev: u64, function: u64, offset: u64) u32 {
+    const address = ecam_base + (((bus - start_bus) << 20) | (dev << 15) | (function << 12) | offset);
+    const register: *volatile u32 = @ptrFromInt(address);
+    return register.*;
+}
+
+fn configWrite(bus: u64, dev: u64, function: u64, offset: u64, value: u32) void {
+    const address = ecam_base + (((bus - start_bus) << 20) | (dev << 15) | (function << 12) | offset);
+    const register: *volatile u32 = @ptrFromInt(address);
+    register.* = value;
+}
+
+fn configRead16(bus: u64, dev: u64, function: u64, offset: u64) u16 {
+    const word = configRead(bus, dev, function, offset & ~@as(u64, 3));
+    return @truncate(word >> @intCast((offset & 3) * 8));
+}
+
+fn configWrite16(bus: u64, dev: u64, function: u64, offset: u64, value: u16) void {
+    const aligned = offset & ~@as(u64, 3);
+    const shift: u5 = @intCast((offset & 3) * 8);
+    const word = configRead(bus, dev, function, aligned);
+    const mask = @as(u32, 0xFFFF) << shift;
+    configWrite(bus, dev, function, aligned, (word & ~mask) | (@as(u32, value) << shift));
+}
+
+/// Claim the bridge, map the ECAM, hello the manager, then scan.
+fn initialise(endpoint: runtime.ipc.Handle) bool {
+    _ = endpoint;
+    if (!device.claim(bridge_id)) {
+        writeLine("pci-bus: unable to claim bridge device {d}\n", .{bridge_id});
+        return false;
+    }
+    const buffer = runtime.allocator().alloc(device.DeviceDescriptor, 64) catch {
+        _ = runtime.system.write("pci-bus: out of memory\n");
+        return false;
+    };
+    const total = device.enumerate(buffer);
+    const descriptor = for (buffer[0..@min(total, buffer.len)]) |d| {
+        if (d.id == bridge_id) break d;
+    } else {
+        writeLine("pci-bus: device {d} not in the device tree\n", .{bridge_id});
+        return false;
+    };
+    // Resource 0 is the ECAM window (1 MiB of config space per bus); the bus
+    // range rides beside it. The MMIO apertures (M19.0) come after both.
+    if (descriptor.resource_count < 2 or descriptor.resources[0].kind != @intFromEnum(device.ResourceKind.memory)) {
+        _ = runtime.system.write("pci-bus: bridge has no ECAM window\n");
+        return false;
+    }
+    const bus_range = for (descriptor.resources[0..@intCast(descriptor.resource_count)]) |resource| {
+        if (resource.kind == @intFromEnum(device.ResourceKind.bus_range)) break resource;
+    } else {
+        _ = runtime.system.write("pci-bus: bridge has no bus range\n");
+        return false;
+    };
+    start_bus = bus_range.start;
+    bus_count = bus_range.len;
+    ecam_physical = descriptor.resources[0].start;
+    ecam_base = device.mmioMap(bridge_id, 0) orelse {
+        _ = runtime.system.write("pci-bus: ECAM mmio_map failed\n");
+        return false;
+    };
+
+    // The handshake, then the scan (reports join in M19.2).
+    var manager: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (manager == null and tries < 100) : (tries += 1) {
+        manager = runtime.ipc.lookup(.device_manager);
+        if (manager == null) runtime.system.sleep(20);
+    }
+    const h = manager orelse {
+        _ = runtime.system.write("pci-bus: no device manager to hello\n");
+        return false;
+    };
+    const hello = protocol.Hello{ .role = @intFromEnum(protocol.Role.bus), .device_id = bridge_id };
+    var reply: [protocol.message_maximum]u8 = undefined;
+    const n = runtime.ipc.call(h, std.mem.asBytes(&hello), &reply) catch {
+        _ = runtime.system.write("pci-bus: hello call failed\n");
+        return false;
+    };
+    if (n < protocol.reply_size or std.mem.bytesToValue(protocol.HelloReply, reply[0..protocol.reply_size]).status != 0) {
+        _ = runtime.system.write("pci-bus: hello refused\n");
+        return false;
+    }
+    manager_handle = h;
+
+    scan();
+    return true;
+}
+
+/// The brute-force walk the kernel does today, from ring 3: every bus in the
+/// range, 32 devices, 8 functions; vendor id FFFFh means nothing decodes there,
+/// and only multifunction devices get their functions 1..7 probed.
+fn scan() void {
+    var found: u32 = 0;
+    var bus: u64 = start_bus;
+    while (bus < start_bus + bus_count) : (bus += 1) {
+        var dev: u64 = 0;
+        while (dev < 32) : (dev += 1) {
+            const first = configRead(bus, dev, 0, 0);
+            if (first & 0xFFFF == 0xFFFF) continue;
+            const multifunction = (configRead(bus, dev, 0, 0x0C) >> 16) & 0x80 != 0;
+            var function: u64 = 0;
+            while (function < 8) : (function += 1) {
+                if (function != 0 and !multifunction) break;
+                const vendor_device = configRead(bus, dev, function, 0);
+                if (vendor_device & 0xFFFF == 0xFFFF) continue;
+                const class_revision = configRead(bus, dev, function, 0x08);
+                found += 1;
+                writeLine("pci-bus: {d}:{d}.{d} class 0x{x:0>6}\n", .{ bus, dev, function, class_revision >> 8 });
+                registerAndReport(bus, dev, function, class_revision >> 8);
+            }
+        }
+    }
+    writeLine("pci-bus: {d} functions found\n", .{found});
+}
+
+/// Register one function under the bridge and report it to the manager. The
+/// descriptor mirrors the kernel's own recording byte for byte — config slice
+/// as resource 0, then the sized BARs — so during coexistence the idempotent
+/// device_register (M19.0) returns the kernel's existing node id rather than
+/// growing a duplicate, and the report carries the id drivers already use.
+fn registerAndReport(bus: u64, dev: u64, function: u64, class_triple: u32) void {
+    var descriptor = std.mem.zeroes(device.DeviceDescriptor);
+    descriptor.class = @intFromEnum(device.DeviceClass.pci_device);
+    descriptor.pci_class = class_triple;
+    descriptor.resources[0] = .{
+        .kind = @intFromEnum(device.ResourceKind.memory),
+        .start = ecam_physical + (((bus - start_bus) << 20) | (dev << 15) | (function << 12)),
+        .len = 4096,
+    };
+    descriptor.resource_count = 1;
+
+    // The standard BAR-sizing probe, exactly as the kernel does it: decode off,
+    // write all-ones, read the writable mask back, restore. Header type 0 only.
+    const header_type = (configRead(bus, dev, function, 0x0C) >> 16) & 0x7F;
+    if (header_type == 0) {
+        const command = configRead16(bus, dev, function, 0x04);
+        configWrite16(bus, dev, function, 0x04, command & ~@as(u16, 0b11));
+        var i: u64 = 0;
+        while (i < 6) : (i += 1) {
+            if (descriptor.resource_count >= 8) break;
+            const off = 0x10 + i * 4;
+            const original = configRead(bus, dev, function, off);
+            if (original == 0) continue;
+            const slot: usize = @intCast(descriptor.resource_count);
+            if (original & 1 != 0) {
+                configWrite(bus, dev, function, off, 0xFFFF_FFFF);
+                const readback = configRead(bus, dev, function, off);
+                configWrite(bus, dev, function, off, original);
+                const mask = readback & 0xFFFF_FFFC;
+                const size: u32 = if (mask == 0) 0 else (~mask +% 1) & 0xFFFF;
+                if (size == 0) continue; // unimplemented BAR — nothing to register
+                descriptor.resources[slot] = .{ .kind = @intFromEnum(device.ResourceKind.io_port), .start = original & 0xFFFF_FFFC, .len = size };
+                descriptor.resource_count += 1;
+            } else if ((original >> 1) & 0x3 == 2) {
+                const original_high = configRead(bus, dev, function, off + 4);
+                configWrite(bus, dev, function, off, 0xFFFF_FFFF);
+                configWrite(bus, dev, function, off + 4, 0xFFFF_FFFF);
+                const lo = configRead(bus, dev, function, off);
+                const hi = configRead(bus, dev, function, off + 4);
+                configWrite(bus, dev, function, off, original);
+                configWrite(bus, dev, function, off + 4, original_high);
+                const readback = (@as(u64, hi) << 32) | (lo & 0xFFFF_FFF0);
+                const size: u64 = if (readback == 0) 0 else ~readback +% 1;
+                i += 1; // consumed the high half regardless
+                if (size == 0) continue;
+                descriptor.resources[slot] = .{ .kind = @intFromEnum(device.ResourceKind.memory), .start = (@as(u64, original_high) << 32) | (original & 0xFFFF_FFF0), .len = size };
+                descriptor.resource_count += 1;
+            } else {
+                configWrite(bus, dev, function, off, 0xFFFF_FFFF);
+                const readback = configRead(bus, dev, function, off);
+                configWrite(bus, dev, function, off, original);
+                const mask = readback & 0xFFFF_FFF0;
+                const size: u32 = if (mask == 0) 0 else ~mask +% 1;
+                if (size == 0) continue;
+                descriptor.resources[slot] = .{ .kind = @intFromEnum(device.ResourceKind.memory), .start = original & 0xFFFF_FFF0, .len = size };
+                descriptor.resource_count += 1;
+            }
+        }
+        configWrite16(bus, dev, function, 0x04, command);
+    }
+
+    const registered = device.register(bridge_id, &descriptor) orelse {
+        writeLine("pci-bus: register refused for {d}:{d}.{d}\n", .{ bus, dev, function });
+        return;
+    };
+    const report = protocol.ChildAdded{
+        .parent = bridge_id,
+        .bus_address = (bus << 8) | (dev << 3) | function,
+        .identity = class_triple,
+        .device_id = registered,
+    };
+    var reply: [protocol.message_maximum]u8 = undefined;
+    _ = runtime.ipc.call(manager_handle, std.mem.asBytes(&report), &reply) catch {
+        writeLine("pci-bus: child report for {d}:{d}.{d} failed\n", .{ bus, dev, function });
+    };
+}
+
+fn onMessage(message: []const u8, reply: []u8, sender: u32, capability: ?runtime.ipc.Handle) usize {
+    _ = message;
+    _ = reply;
+    _ = sender;
+    _ = capability;
+    return 0;
+}
+
+pub fn main(init: runtime.process.Init) void {
+    const argument = init.arguments.get(1) orelse return; // bare (ramdisk sweep): stay silent
+    bridge_id = std.fmt.parseInt(u64, argument, 10) catch {
+        writeLine("pci-bus: malformed bridge device id '{s}'\n", .{argument});
+        return;
+    };
+    runtime.service.run(protocol.message_maximum, .{
+        .init = initialise,
+        .on_message = onMessage,
+    });
+}
+
+pub const panic = runtime.panic;
+comptime {
+    _ = &runtime.start._start; // pull the runtime entry shim into the image
+}

system/drivers/usb-xhci-bus/usb-xhci-bus.zig

@@ -1,68 +1,172 @@
 //! /system/drivers/usb-xhci-bus — the xHCI (USB 3) host-controller bus driver.
-//! The device manager spawns **one instance per controller** it discovers (a machine
-//! can carry several), passing the controller's device-tree id as argv[1]; this
-//! instance claims that device and no other, so multiple instances never fight over
-//! hardware. This increment proves the plumbing: parse the id, claim the controller,
-//! and report its MMIO window. The next increments map the registers and bring the
-//! controller up (reset, rings, port scan), then enumerate the USB devices on the
-//! bus with the usb-abi request builders and publish each with `device_register`.
+//! The device manager spawns **one instance per controller** it discovers (a
+//! machine can carry several), passing the controller's device-tree id as
+//! argv[1]; this instance claims that device and no other, so multiple
+//! instances never fight over hardware.
+//!
+//! M18.2 (this increment): after the hello, real hardware — map the xHC's
+//! register window (the first memory BAR; resource 0 is the ECAM config
+//! space), read the capability registers, and walk the root-hub ports: one
+//! `child_added` report to the manager per connected port, carrying the port
+//! number and the PORTSC speed class as identity. No transfer rings yet —
+//! descriptors and USB class matching are the USB track; the connect bit and
+//! speed come straight from PORTSC, which reflects hardware state whether or
+//! not the controller is running.
 
 const std = @import("std");
 const runtime = @import("runtime");
+const protocol = runtime.device_manager_protocol;
 const device = runtime.device;
 
-/// Format one whole log line and emit it in a single `debug_write`, so concurrent
-/// instances (one per controller) can never interleave mid-line.
+/// Format one whole log line and emit it in a single `debug_write`, so
+/// concurrent instances (one per controller) can never interleave mid-line.
 fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
     var line: [128]u8 = undefined;
     _ = runtime.system.write(std.fmt.bufPrint(&line, fmt, arguments) catch return);
 }
 
+var controller_id: u64 = protocol.no_device;
+
+/// Claim the assigned controller, find its register window, and hello the
+/// manager. Any failure returns false: the process exits cleanly, which the
+/// manager reads as "meant to stop" — a missing assignment is not a crash loop.
+fn initialise(endpoint: runtime.ipc.Handle) bool {
+    _ = endpoint;
+    if (!device.claim(controller_id)) {
+        writeLine("usb-xhci-bus: unable to claim controller device {d}\n", .{controller_id});
+        return false;
+    }
+
+    // Fetch our own descriptor back for the controller's resources.
+    const buffer = runtime.allocator().alloc(device.DeviceDescriptor, 64) catch {
+        _ = runtime.system.write("usb-xhci-bus: out of memory\n");
+        return false;
+    };
+    const total = device.enumerate(buffer);
+    const descriptor = for (buffer[0..@min(total, buffer.len)]) |d| {
+        if (d.id == controller_id) break d;
+    } else {
+        writeLine("usb-xhci-bus: device {d} not in the device tree\n", .{controller_id});
+        return false;
+    };
+
+    // The xHC's registers live behind the first memory BAR. Resource 0 is the
+    // function's ECAM configuration space (M15), so the walk starts at 1.
+    var register_index: u64 = 0;
+    const register_window = for (descriptor.resources[1..@intCast(descriptor.resource_count)], 1..) |resource, index| {
+        if (resource.kind == @intFromEnum(device.ResourceKind.memory)) {
+            register_index = index;
+            break resource;
+        }
+    } else {
+        writeLine("usb-xhci-bus: controller device {d} has no register BAR\n", .{controller_id});
+        return false;
+    };
+    writeLine("usb-xhci-bus: claimed controller device {d} (registers at 0x{x}, {d} bytes)\n", .{
+        controller_id,
+        register_window.start,
+        register_window.len,
+    });
+    register_base = device.mmioMap(controller_id, register_index) orelse {
+        _ = runtime.system.write("usb-xhci-bus: mmio_map failed\n");
+        return false;
+    };
+
+    // The handshake: role, protocol version, assignment — inside the manager's
+    // deadline (the lookup retries cover the manager still registering).
+    var manager: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (manager == null and tries < 100) : (tries += 1) {
+        manager = runtime.ipc.lookup(.device_manager);
+        if (manager == null) runtime.system.sleep(20);
+    }
+    const h = manager orelse {
+        _ = runtime.system.write("usb-xhci-bus: no device manager to hello\n");
+        return false;
+    };
+    const hello = protocol.Hello{ .role = @intFromEnum(protocol.Role.bus), .device_id = controller_id };
+    var reply: [protocol.message_maximum]u8 = undefined;
+    const n = runtime.ipc.call(h, std.mem.asBytes(&hello), &reply) catch {
+        _ = runtime.system.write("usb-xhci-bus: hello call failed\n");
+        return false;
+    };
+    if (n < protocol.reply_size or std.mem.bytesToValue(protocol.HelloReply, reply[0..protocol.reply_size]).status != 0) {
+        _ = runtime.system.write("usb-xhci-bus: hello refused\n");
+        return false;
+    }
+    _ = runtime.system.write("usb-xhci-bus: hello acknowledged\n");
+
+    scanPorts(h);
+    return true;
+}
+
+var register_base: usize = 0;
+
+/// One 32-bit volatile register read at `offset` from the mapped window.
+fn readRegister(offset: usize) u32 {
+    const register: *volatile u32 = @ptrFromInt(register_base + offset);
+    return register.*;
+}
+
+/// The root-hub port scan: read the capability registers for the port count
+/// and the operational-register offset, then one PORTSC per port. The connect
+/// bit (CCS) and the speed field reflect hardware state directly — no
+/// controller reset or run needed to *see* the devices; driving them needs the
+/// rings (the USB track).
+fn scanPorts(manager: runtime.ipc.Handle) void {
+    // Capability registers: CAPLENGTH is byte 0 of the first dword; HCSPARAMS1
+    // carries MaxPorts in bits 31:24.
+    const capability_length = readRegister(0) & 0xFF;
+    const structural = readRegister(0x04);
+    const maximum_ports: u32 = structural >> 24;
+    writeLine("usb-xhci-bus: {d} root-hub ports\n", .{maximum_ports});
+
+    // PORTSC registers: operational base + 0x400 + 0x10 per port (1-based).
+    var port: u32 = 1;
+    var connected: u32 = 0;
+    while (port <= maximum_ports) : (port += 1) {
+        const port_status = readRegister(capability_length + 0x400 + 0x10 * (port - 1));
+        if (port_status & 1 == 0) continue; // CCS: nothing connected
+        connected += 1;
+        const speed = (port_status >> 10) & 0xF; // the PORTSC port-speed class
+        writeLine("usb-xhci-bus: port {d} connected (speed class {d})\n", .{ port, speed });
+
+        const report = protocol.ChildAdded{
+            .parent = controller_id,
+            .bus_address = port,
+            .identity = speed,
+        };
+        var reply: [protocol.message_maximum]u8 = undefined;
+        _ = runtime.ipc.call(manager, std.mem.asBytes(&report), &reply) catch {
+            writeLine("usb-xhci-bus: child report for port {d} failed\n", .{port});
+            continue;
+        };
+    }
+    if (connected == 0) _ = runtime.system.write("usb-xhci-bus: no devices connected\n");
+}
+
+/// No bus protocol to serve yet — transfer requests arrive with the USB track.
+fn onMessage(message: []const u8, reply: []u8, sender: u32, capability: ?runtime.ipc.Handle) usize {
+    _ = message;
+    _ = reply;
+    _ = sender;
+    _ = capability;
+    return 0;
+}
+
 pub fn main(init: runtime.process.Init) void {
     const argument = init.arguments.get(1) orelse {
         _ = runtime.system.write("usb-xhci-bus: missing controller device id (argv[1])\n");
         return;
     };
-    const controller_id = std.fmt.parseInt(u64, argument, 10) catch {
+    controller_id = std.fmt.parseInt(u64, argument, 10) catch {
         writeLine("usb-xhci-bus: malformed controller device id '{s}'\n", .{argument});
         return;
     };
-
-    if (!device.claim(controller_id)) {
-        writeLine("usb-xhci-bus: unable to claim controller device {d}\n", .{controller_id});
-        return;
-    }
-
-    // Fetch our own descriptor back for the controller's resources.
-    const buffer = runtime.allocator().alloc(device.DeviceDescriptor, 64) catch {
-        _ = runtime.system.write("usb-xhci-bus: out of memory\n");
-        return;
-    };
-    const total = device.enumerate(buffer);
-    const descriptor = for (buffer[0..@min(total, buffer.len)]) |d| {
-        if (d.id == controller_id) break d;
-    } else {
-        writeLine("usb-xhci-bus: device {d} not in the device tree\n", .{controller_id});
-        return;
-    };
-
-    // The controller's operational registers live behind BAR0, enumerated as the
-    // device's first memory resource.
-    const register_window = for (descriptor.resources[0..@intCast(descriptor.resource_count)]) |resource| {
-        if (resource.kind == @intFromEnum(device.ResourceKind.memory)) break resource;
-    } else {
-        writeLine("usb-xhci-bus: controller device {d} has no MMIO window\n", .{controller_id});
-        return;
-    };
-    writeLine("usb-xhci-bus: claimed controller device {d} (registers at 0x{x}, {d} bytes)\n", .{
-        controller_id,
-        register_window.start,
-        register_window.len,
+    runtime.service.run(protocol.message_maximum, .{
+        .init = initialise,
+        .on_message = onMessage,
     });
-
-    // Controller bring-up (map the window, reset, rings, port scan) is the next
-    // increment; stay resident as the bus's supervisor in the meantime.
-    while (true) runtime.system.sleep(1000);
 }
 
 pub const panic = runtime.panic;

system/kernel/devices-broker.zig

@@ -104,6 +104,19 @@ pub fn ownerOf(id: u64) ?u32 {
     return claimed[@intCast(id)];
 }
 
+/// Release every claim held by `owner` — called by the process layer on every
+/// path out of a process (exit, fault, kill), so a restarted driver can claim its
+/// hardware again (docs/process-lifecycle.md iron rule 1: cleanup is the kernel's
+/// job). The devices stay in the table — they describe hardware, which did not go
+/// away — only their ownership clears.
+pub fn releaseAllOwnedBy(owner: u32) void {
+    for (claimed[0..count]) |*slot| {
+        if (slot.*) |o| {
+            if (o == owner) slot.* = null;
+        }
+    }
+}
+
 /// Resource `index` of device `id`, or null if out of range.
 pub fn resourceOf(id: u64, index: u64) ?device_abi.ResourceDescriptor {
     if (id >= count) return null;
@@ -167,6 +180,26 @@ pub fn register(parent_id: u64, owner: u32, descriptor: *const device_abi.Device
         if (!ok) return error.NotContained;
     }
 
+    // Idempotent on exact match (docs/m19-m20-plan.md decision 3): a restarted
+    // registering bus re-registers what it rediscovers, and the table has no
+    // unregister — an identical (class, identity, resources) child under the
+    // same parent returns the existing id instead of appending a duplicate.
+    for (devices[0..count]) |*existing| {
+        if (existing.parent != parent_id) continue;
+        if (existing.class != descriptor.class) continue;
+        if (existing.pci_class != descriptor.pci_class) continue;
+        if (existing.hid_len != descriptor.hid_len) continue;
+        if (!std.mem.eql(u8, existing.hid[0..@intCast(existing.hid_len)], descriptor.hid[0..@intCast(descriptor.hid_len)])) continue;
+        if (existing.resource_count != descriptor.resource_count) continue;
+        var same = true;
+        for (0..@intCast(descriptor.resource_count)) |i| {
+            const a = existing.resources[i];
+            const b = descriptor.resources[i];
+            if (a.kind != b.kind or a.start != b.start or a.len != b.len) same = false;
+        }
+        if (same) return existing.id;
+    }
+
     var d = std.mem.zeroes(device_abi.DeviceDescriptor);
     d.id = count;
     d.parent = parent_id;

system/kernel/kernel.zig

@@ -412,13 +412,26 @@ fn recoverableFault(vector: u64) bool {
 /// plus a POST code and a persistent breadcrumb. (A ring-3 fault on a *borrowed*
 /// kernel thread — process.run, the user-pf isolation probe — also lands here: there
 /// is no scheduled process to kill.)
+/// Classify a CPU exception vector as the ExitReason a supervisor reads — the
+/// fault classes of docs/process-lifecycle.md. Faults are exit reasons, never
+/// signals delivered to the faulting process: recovery is restart, not a handler.
+fn exitReasonForVector(vector: u64) abi.ExitReason {
+    return switch (vector) {
+        14 => .segmentation_fault, // page fault
+        6 => .illegal_instruction, // invalid opcode
+        0, 16, 19 => .arithmetic_fault, // divide error, x87, SIMD
+        13 => .protection_fault, // general protection
+        else => .fault,
+    };
+}
+
 fn onException(state: *const architecture.CpuState) noreturn {
     if (architecture.fromUser(state) and scheduler.currentIsUserProcess() and recoverableFault(state.vector)) {
         statusPrint("\ndanos: process {d} ({s}) killed by {s} (vector {d}) on core {d}\n", .{ scheduler.currentId(), scheduler.current().name(), architecture.exceptionName(state.vector), state.vector, scheduler.currentCpuIndex() });
         statusPrint("  error code : 0x{x}\n", .{state.error_code});
         statusPrint("  IP         : 0x{x:0>16}\n", .{architecture.instructionPointer(state)});
         if (architecture.faultAddress(state)) |address| statusPrint("  fault addr : 0x{x:0>16}\n", .{address});
-        process.killCurrentProcess(); // reclaims everything, reschedules; never returns
+        process.killCurrentProcess(exitReasonForVector(state.vector)); // reclaims everything, reschedules; never returns
     }
 
     log.checkpoint(cp_exception);

system/kernel/process.zig

@@ -137,6 +137,7 @@ pub fn init() void {
     architecture.setSystemCallHandler(system_call);
     scheduler.terminate_current_hook = terminateCurrentLocked;
     scheduler.reap_task_hook = reapTaskLocked;
+    scheduler.timer_tick_hook = timerSweepLocked;
 }
 
 /// Return -1 (as an unsigned bit pattern) in the system_call result register.
@@ -164,6 +165,7 @@ fn system_call(state: *architecture.CpuState) void {
             // A scheduled process tears down fully (terminateCurrent); a borrowed
             // test thread unwinds back to the kernel that entered it.
             if (scheduler.currentIsUserProcess()) {
+                scheduler.current().exit_reason = .exited;
                 terminateCurrent();
             } else architecture.userExit();
         },
@@ -199,6 +201,11 @@ fn system_call(state: *architecture.CpuState) void {
         .clock => systemClock(state),
         .process_enumerate => systemProcessEnumerate(state),
         .process_kill => systemProcessKill(state),
+        .process_exit_reason => systemProcessExitReason(state),
+        .process_subscribe => systemProcessSubscribe(state),
+        .signal_bind => systemSignalBind(state),
+        .process_signal => systemProcessSignal(state),
+        .timer_bind => systemTimerBind(state),
         _ => fail(state),
     }
 }
@@ -291,6 +298,8 @@ fn systemDeviceEnumerate(state: *architecture.CpuState) void {
 
 /// device_claim(id) -> 0/-1: take exclusive ownership of a device for this process.
 fn systemDeviceClaim(state: *architecture.CpuState) void {
+    const claim_flags = sync.enter();
+    defer sync.leave(claim_flags);
     if (devices_broker.claim(architecture.systemCallArg(state, 0), scheduler.current().id))
         architecture.setSystemCallResult(state, 0)
     else
@@ -305,10 +314,22 @@ fn systemMmioMap(state: *architecture.CpuState) void {
     const resource_index = architecture.systemCallArg(state, 1);
     const t = scheduler.current();
     if (t.aspace == 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 r = devices_broker.resourceOf(device_id, resource_index) orelse 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
+    // below on integer overflow).
+    const r = blk: {
+        const flags = sync.enter();
+        defer sync.leave(flags);
+        const owner = devices_broker.ownerOf(device_id) orelse return fail(state);
+        if (owner != t.id) return fail(state); // not claimed by this process
+        break :blk devices_broker.resourceOf(device_id, resource_index) orelse return fail(state);
+    };
     if (r.kind != @intFromEnum(device_abi.ResourceKind.memory)) return fail(state);
+    // A zero-length or wrapping window is not mappable — fail cleanly rather
+    // than underflow `r.len - 1`.
+    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);
@@ -444,6 +465,11 @@ fn systemDeviceRegister(state: *architecture.CpuState) void {
     var descriptor: device_abi.DeviceDescriptor = undefined;
     if (!ipc.copyFromUser(t.aspace, 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 —
+    // ring-3 registration (M19) made those genuinely concurrent.
+    const flags = sync.enter();
+    defer sync.leave(flags);
     const id = devices_broker.register(parent_id, t.id, &descriptor) catch return fail(state);
     architecture.setSystemCallResult(state, id);
 }
@@ -552,7 +578,11 @@ pub var fault_kill_count: u64 = 0;
 ///   endpoint reference destroys the Endpoint, and a still-bound GSI would have an
 ///   ISR call notifyFromIsr on freed memory the next time the device fired.
 ///   `releaseOwner` also leaves the line masked, so a dead driver's device goes
-///   quiet rather than storming.
+///   quiet rather than storming. (It drops MSI vectors by the same owner sweep.)
+/// - Device claims are released with the IRQ bindings, so a restarted driver can
+///   claim the same hardware again — the cleanup half of process-lifecycle.md's
+///   iron rule 1. Claims hold no pointers, so ordering is free; they go here so
+///   the exit notification (below, last) observes a fully-released child.
 /// - A client this task still owes a reply to (it died between receive and reply)
 ///   is failed with -EPEER rather than left blocked forever — a dead server must
 ///   not hang its callers.
@@ -565,7 +595,33 @@ pub var fault_kill_count: u64 = 0;
 ///   reference taken at spawn is dropped with it.
 /// Precondition: the big kernel lock is held.
 fn releaseTaskResourcesLocked(t: *scheduler.Task) void {
+    recordExitLocked(t);
     irq.releaseOwner(t.id);
+    devices_broker.releaseAllOwnedBy(t.id);
+    // The dying task's signal endpoint and one-shot timers go with it.
+    if (t.signal_endpoint) |raw| {
+        ipc.dropRef(@ptrCast(@alignCast(raw)));
+        t.signal_endpoint = null;
+    }
+    t.pending_signals = 0;
+    for (&one_shot_timers) |*slot| {
+        if (slot.*) |timer| {
+            if (timer.owner == t.id) {
+                ipc.dropRef(timer.endpoint);
+                slot.* = null;
+            }
+        }
+    }
+    // A dead subscriber's own subscriptions go first: it must not hear about
+    // itself, and the slots' endpoint references drop with it.
+    for (&exit_subscribers) |*slot| {
+        if (slot.*) |subscriber| {
+            if (subscriber.owner == t.id) {
+                ipc.dropRef(subscriber.endpoint);
+                slot.* = null;
+            }
+        }
+    }
     if (t.ipc_client) |client| {
         t.ipc_client = null;
         client.ipc_status = -ipc.EPEER;
@@ -575,6 +631,12 @@ fn releaseTaskResourcesLocked(t: *scheduler.Task) void {
     scheduler.removeFromWaitQueueLocked(t);
     scheduler.forgetIpcClientLocked(t);
     ipc.closeHandles(t);
+    // Publish the exit to every subscriber (docs/process-lifecycle.md): the same
+    // badge encoding as the supervisor's notification, and equally late, so a
+    // subscriber also observes a fully-released child.
+    for (&exit_subscribers) |*slot| {
+        if (slot.*) |subscriber| ipc.notifyLocked(subscriber.endpoint, abi.notify_exit_bit | t.id);
+    }
     if (t.exit_endpoint) |raw| {
         const endpoint: *ipc.Endpoint = @ptrCast(@alignCast(raw));
         t.exit_endpoint = null;
@@ -628,6 +690,7 @@ pub fn killProcess(caller_id: u32, target_id: u32) i64 {
     const target = scheduler.taskByIdLocked(target_id) orelse return -ipc.ESRCH;
     if (target.aspace == 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) {
         target.kill_pending = true;
     } else {
@@ -640,12 +703,170 @@ pub fn killProcess(caller_id: u32, target_id: u32) i64 {
 /// The fault is confined to the process — the kernel trapped it on the task's own
 /// kernel stack and is intact — so everything the process held is reclaimed and the
 /// core reschedules. The system keeps running; only the faulting process dies
-/// (docs/resilience.md: fault -> kill -> continue).
-pub fn killCurrentProcess() noreturn {
+/// (docs/resilience.md: fault -> kill -> continue). `reason` is the fault class
+/// (from the vector), recorded for the supervisor's `process_exit_reason`.
+pub fn killCurrentProcess(reason: abi.ExitReason) noreturn {
+    scheduler.current().exit_reason = reason;
     fault_kill_count += 1;
     terminateCurrent();
 }
 
+/// The bounded record of recent deaths, for `process_exit_reason`: ids are never
+/// reused, so a ring keyed by id is enough — a record evicted by wraparound reads
+/// as -ESRCH, the same as an id that never lived, which a supervisor treats as
+/// "too late to ask". Written under the big kernel lock by the reap.
+const exit_record_capacity = 64;
+const ExitRecord = struct { id: u32 = 0, supervisor: u32 = 0, reason: abi.ExitReason = .exited, valid: bool = false };
+var exit_records: [exit_record_capacity]ExitRecord = .{ExitRecord{}} ** exit_record_capacity;
+var exit_record_next: usize = 0;
+
+/// Record a dying task's (id, supervisor, reason) — called by the reap before the
+/// exit notification is posted, so a supervisor that hears the notification can
+/// always still query the reason. Precondition: the big kernel lock is held.
+fn recordExitLocked(t: *scheduler.Task) void {
+    exit_records[exit_record_next] = .{ .id = t.id, .supervisor = t.supervisor, .reason = t.exit_reason, .valid = true };
+    exit_record_next = (exit_record_next + 1) % exit_record_capacity;
+}
+
+/// How dead process `id` ended, for `caller` — the kernel half of the
+/// process_exit_reason system call. Returns the ExitReason value, -ESRCH (never
+/// lived, still alive, or evicted from the ring), or -EPERM (the caller was not
+/// its supervisor — the same authority gate as process_kill).
+pub fn exitReasonOf(caller_id: u32, target_id: u32) i64 {
+    const flags = sync.enter();
+    defer sync.leave(flags);
+    for (&exit_records) |*record| {
+        if (record.valid and record.id == target_id) {
+            if (record.supervisor != caller_id) return -ipc.EPERM;
+            return @intFromEnum(record.reason);
+        }
+    }
+    return -ipc.ESRCH;
+}
+
+/// The published exit events' subscribers (docs/process-lifecycle.md "Who learns
+/// of a death"): stateful services — the VFS's file handles, input's
+/// subscriptions — that must release what a dead client held and cannot learn it
+/// any other way (a client that simply never calls again looks like silence).
+/// Bounded like every kernel table; each entry holds its own endpoint reference.
+const exit_subscriber_capacity = 8;
+const ExitSubscriber = struct { endpoint: *ipc.Endpoint, owner: u32 };
+var exit_subscribers: [exit_subscriber_capacity]?ExitSubscriber = .{null} ** exit_subscriber_capacity;
+
+/// process_subscribe(endpoint): subscribe the caller's endpoint to published exit
+/// events. Ungated, like process_enumerate — what is running (and dying) is not a
+/// 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);
+    const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 0)) orelse return failErr(state, ipc.EBADF);
+    const flags = sync.enter();
+    defer sync.leave(flags);
+    for (&exit_subscribers) |*slot| {
+        if (slot.* == null) {
+            endpoint.refcount += 1; // the slot's own reference, dropped on unsubscribe-by-death
+            slot.* = .{ .endpoint = endpoint, .owner = t.id };
+            return architecture.setSystemCallResult(state, 0);
+        }
+    }
+    failErr(state, ipc.ENOSPC);
+}
+
+/// signal_bind(endpoint): nominate where this process's signals arrive — the
+/// IRQ-as-IPC pattern a fourth time (docs/process-lifecycle.md). Replacing a
+/// binding drops the old reference; signals that pended while unbound are
+/// 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);
+    const endpoint = ipc.resolveHandle(t, architecture.systemCallArg(state, 0)) orelse return failErr(state, ipc.EBADF);
+    const flags = sync.enter();
+    defer sync.leave(flags);
+    if (t.signal_endpoint) |raw| ipc.dropRef(@ptrCast(@alignCast(raw)));
+    endpoint.refcount += 1;
+    t.signal_endpoint = @ptrCast(endpoint);
+    if (t.pending_signals != 0) {
+        ipc.notifyLocked(endpoint, abi.notify_signal_bit | t.pending_signals);
+        t.pending_signals = 0;
+    }
+    architecture.setSystemCallResult(state, 0);
+}
+
+/// process_signal(id, signal): post a signal — a one-way, coalescing statement,
+/// never a question (docs/process-lifecycle.md). The authority gate is the
+/// supervision link, like kill; a process may also signal itself. Unbound
+/// 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);
+    const id = architecture.systemCallArg(state, 0);
+    const signal = architecture.systemCallArg(state, 1);
+    if (id > std.math.maxInt(u32)) return failErr(state, ipc.ESRCH);
+    if (signal > 31) return failErr(state, ipc.EBADF); // not a Signal bit position
+    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.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| {
+        const endpoint: *ipc.Endpoint = @ptrCast(@alignCast(raw));
+        ipc.notifyLocked(endpoint, abi.notify_signal_bit | target.pending_signals);
+        target.pending_signals = 0;
+    }
+    architecture.setSystemCallResult(state, 0);
+}
+
+/// The one-shot timers of timer_bind: the missing timed wait. A service arms a
+/// deadline and keeps serving; the expiry arrives in the same replyWait as
+/// everything else (notify_timer_bit). What stop-sequence escalation, hello
+/// deadlines, and restart backoff are built from — and later, `alarm`.
+const timer_capacity = 16;
+const OneShotTimer = struct { deadline: u64, endpoint: *ipc.Endpoint, owner: u32 };
+var one_shot_timers: [timer_capacity]?OneShotTimer = .{null} ** timer_capacity;
+
+/// Sweep expired timers — hung on scheduler.timer_tick_hook, so it runs on every
+/// tick with the big kernel lock held, like the sleeper wake it rides beside.
+fn timerSweepLocked() void {
+    const now = architecture.millis();
+    for (&one_shot_timers) |*slot| {
+        if (slot.*) |timer| {
+            if (now >= timer.deadline) {
+                ipc.notifyLocked(timer.endpoint, abi.notify_timer_bit);
+                ipc.dropRef(timer.endpoint);
+                slot.* = null;
+            }
+        }
+    }
+}
+
+/// 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);
+    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();
+    defer sync.leave(flags);
+    for (&one_shot_timers) |*slot| {
+        if (slot.* == null) {
+            endpoint.refcount += 1;
+            slot.* = .{ .deadline = architecture.millis() + ms, .endpoint = endpoint, .owner = t.id };
+            return architecture.setSystemCallResult(state, 0);
+        }
+    }
+    failErr(state, ipc.ENOSPC);
+}
+
+fn systemProcessExitReason(state: *architecture.CpuState) void {
+    const t = scheduler.current();
+    if (t.aspace == 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));
+    architecture.setSystemCallResult(state, @bitCast(r));
+}
+
 /// Resolve `(device_id, resource_index)` to a GSI this process is entitled to bind, or null.
 /// The two checks are the whole security story: the device must be *claimed* by the
 /// caller, and the resource must be one of that device's `irq` resources as recorded

system/kernel/scheduler.zig

@@ -50,6 +50,17 @@ pub const Task = struct {
     // null. Holds its own reference, dropped when the notification is posted.
     // Opaque here for the same reason as `handles` below.
     exit_endpoint: ?*anyopaque = null,
+    // How this process ended — set by the death paths (exit, fault, kill) just
+    // before the reap records it for `process_exit_reason`. Meaningless while
+    // the task lives.
+    exit_reason: abi.ExitReason = .exited,
+    // Endpoint this process's signals arrive on (signal_bind), or null — same
+    // ownership rules as exit_endpoint (holds a reference; opaque here).
+    signal_endpoint: ?*anyopaque = null,
+    // Signals posted but not yet delivered: the coalescing pending mask
+    // (docs/process-lifecycle.md). Bits are abi.Signal values. Signals pend here
+    // until an endpoint is bound; two pending terminates are one terminate.
+    pending_signals: u32 = 0,
     // Set by process_kill on a task that is running on another core; the kernel
     // finishes the kill at that task's next system call or timer tick.
     kill_pending: bool = false,
@@ -339,8 +350,9 @@ pub fn spawnUserLocked(aspace: u64, entry: u64, user_sp: u64, priority: Priority
 /// context switch and lock release.
 fn startUserTask() void {
     const t = current();
-    var buffer: [96]u8 = undefined;
-    architecture.serialWrite(std.fmt.bufPrint(&buffer, "DBG startUserTask ip=0x{x} sp=0x{x} aspace=0x{x} kstack=0x{x}\n", .{ t.user_ip, t.user_sp, t.aspace, t.kstack_top }) catch "");
+    // No serial chatter here: this runs on every spawn, unserialized against
+    // user-space writes, and its output used to shear concurrent log lines in
+    // half — the largest source of corrupted markers in the QEMU scenarios.
     architecture.jumpToUser(t.user_ip, t.user_sp); // noreturn
 }
 
@@ -643,9 +655,15 @@ fn reapKillPendingLocked() void {
 /// other critical section, but releases it *without* touching the interrupt flag
 /// — the handler's `iretq` restores the interrupted context's flags, so
 /// re-enabling here would open a nested-interrupt window before the return.
+/// Called from the tick with the big kernel lock held — process.zig hangs the
+/// one-shot timer sweep here (timer_bind), the same call-up pattern as the
+/// teardown hooks below.
+pub var timer_tick_hook: ?*const fn () void = null;
+
 pub fn tick() void {
     _ = sync.enter();
     wakeExpired();
+    if (timer_tick_hook) |hook| hook();
     reapKillPendingLocked();
     if (preemption_enabled) schedule();
     sync.leaveIsr();

system/kernel/tests.zig

@@ -132,6 +132,20 @@ pub fn run(case: []const u8, boot_information: *const BootInformation) void {
         processKillTest(boot_information);
     } else if (eql(case, "supervision")) {
         supervisionTest(boot_information);
+    } else if (eql(case, "claim-release")) {
+        claimReleaseTest(boot_information);
+    } else if (eql(case, "vfs-client-death")) {
+        vfsClientDeathTest(boot_information);
+    } else if (eql(case, "signals")) {
+        signalsTest(boot_information);
+    } else if (eql(case, "driver-restart")) {
+        driverRestartTest(boot_information);
+    } else if (eql(case, "usb-report")) {
+        usbReportTest(boot_information);
+    } else if (eql(case, "device-list")) {
+        deviceListTest(boot_information);
+    } else if (eql(case, "pci-scan")) {
+        pciScanTest(boot_information);
     } else if (eql(case, "initial-ramdisk")) {
         initialRamdiskTest(boot_information);
     } else if (eql(case, "vfs")) {
@@ -256,20 +270,56 @@ fn discoveryTest() void {
 
     // M15: every PCI function now carries its own 4 KiB ECAM configuration space as
     // resource 0 — the window a driver mmio_maps to walk its capability list (MSI etc).
+    // M19.3: the kernel seeds only the bridge; functions arrive by the ring-3
+    // scan (proven equivalent in pci-scan before the walk retired).
     var buffer: [64]device_abi.DeviceDescriptor = undefined;
     const n = @min(devices_broker.enumerate(&buffer), buffer.len);
-    var pci_functions: u32 = 0;
-    var pci_config_ok = true;
+    var bridges: u32 = 0;
+    var bridge_shape_ok = false;
+    for (buffer[0..n]) |d| {
+        if (d.class != @intFromEnum(device_abi.DeviceClass.pci_host_bridge)) continue;
+        bridges += 1;
+        var has_bus_range = false;
+        var has_io = false;
+        var memory_windows: u32 = 0;
+        for (d.resources[0..@intCast(d.resource_count)]) |resource| {
+            if (resource.kind == @intFromEnum(device_abi.ResourceKind.bus_range)) has_bus_range = true;
+            if (resource.kind == @intFromEnum(device_abi.ResourceKind.io_port)) has_io = true;
+            if (resource.kind == @intFromEnum(device_abi.ResourceKind.memory)) memory_windows += 1;
+        }
+        // ECAM plus at least one MMIO aperture, the bus range, the I/O window.
+        if (has_bus_range and has_io and memory_windows >= 2) bridge_shape_ok = true;
+    }
+    check("a PCI host bridge was seeded (MCFG)", bridges >= 1);
+    check("the bridge carries ECAM, apertures, bus range, and the I/O window", bridge_shape_ok);
+
+    // M19.0: every PCI memory resource (config slice and BARs alike) must be
+    // contained in one of its parent bridge's windows — the aperture derivation
+    // from the memory map is what makes a future user-space device_register of
+    // these functions pass containment. This is the assert that catches a
+    // too-coarse hole computation before M19.2 would.
+    var bars_contained = true;
     for (buffer[0..n]) |d| {
         if (d.class != @intFromEnum(device_abi.DeviceClass.pci_device)) continue;
-        pci_functions += 1;
-        const has_config = d.resource_count >= 1 and
-            d.resources[0].kind == @intFromEnum(device_abi.ResourceKind.memory) and
-            d.resources[0].len == abi.page_size;
-        if (!has_config) pci_config_ok = false;
+        if (d.parent >= n) {
+            bars_contained = false;
+            continue;
+        }
+        const bridge = buffer[@intCast(d.parent)];
+        for (d.resources[0..@intCast(d.resource_count)]) |r| {
+            if (r.kind != @intFromEnum(device_abi.ResourceKind.memory)) continue;
+            var inside = false;
+            for (bridge.resources[0..@intCast(bridge.resource_count)]) |w| {
+                if (w.kind != @intFromEnum(device_abi.ResourceKind.memory)) continue;
+                if (r.start >= w.start and r.start + r.len <= w.start + w.len) inside = true;
+            }
+            if (!inside) {
+                bars_contained = false;
+                log("  escaping BAR: 0x{x}+0x{x} on device {d}\n", .{ r.start, r.len, d.id });
+            }
+        }
     }
-    check("PCI functions were enumerated (MCFG/ECAM)", pci_functions >= 1);
-    check("each PCI function exposes its ECAM config space as resource 0", pci_config_ok);
+    check("every PCI BAR lies inside a bridge aperture (M19.0)", bars_contained);
 
     result();
 }
@@ -1207,15 +1257,15 @@ fn userPfTest() void {
 /// hand — address space, code page RO+X, stack page RW+NX — because the blob is a
 /// raw code fragment, not an ELF `spawnProcess` could load. Returns false if any
 /// allocation fails.
-fn spawnFaultingProcess() bool {
+fn spawnFaultingProcess() ?u32 {
     const blob = process.pfBlob();
     const flags = sync.enter();
     defer sync.leave(flags);
 
-    const aspace = architecture.createAddressSpace() orelse return false;
+    const aspace = architecture.createAddressSpace() orelse return null;
     const code_frame = pmm.alloc() orelse {
         architecture.destroyAddressSpace(aspace);
-        return false;
+        return null;
     };
     // Fill through the physmap (the user mapping is read-only); pad with int3 so a
     // stray jump traps instead of sliding.
@@ -1226,15 +1276,16 @@ fn spawnFaultingProcess() bool {
 
     const stack_frame = pmm.alloc() orelse {
         architecture.destroyAddressSpace(aspace); // frees code_frame too — it's mapped
-        return false;
+        return null;
     };
     architecture.mapUserPageInto(aspace, process.stack_base_virtual, stack_frame, true, false); // RW + NX
 
-    if (scheduler.spawnUserLocked(aspace, process.code_virtual, process.stack_base_virtual + abi.page_size, 4, "fault-probe", 0, null) == null) {
+    // 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, 4, "fault-probe", scheduler.currentId(), null) orelse {
         architecture.destroyAddressSpace(aspace);
-        return false;
-    }
-    return true;
+        return null;
+    };
+    return id;
 }
 
 /// Fault recovery (docs/resilience.md step 2): a scheduled ring-3 process that
@@ -1263,7 +1314,8 @@ fn faultRecoveryTest(boot_information: *const BootInformation) void {
     scheduler.setPriority(4);
     check("init heartbeat before the fault", process.write_count >= 1);
 
-    check("faulting process spawned", spawnFaultingProcess());
+    const probe = spawnFaultingProcess() orelse 0;
+    check("faulting process spawned", probe != 0);
 
     // The kill: the faulting process #PFs on its first instruction and the kernel
     // reaps it instead of halting.
@@ -1272,6 +1324,7 @@ fn faultRecoveryTest(boot_information: *const BootInformation) void {
     while (process.fault_kill_count < 1 and architecture.millis() < deadline) scheduler.yield();
     scheduler.setPriority(4);
     check("faulting process was killed (not the machine)", process.fault_kill_count == 1);
+    check("the probe's reason reads segmentation_fault", process.exitReasonOf(scheduler.currentId(), probe) == @intFromEnum(abi.ExitReason.segmentation_fault));
 
     // Life after the kill: init must keep beating on the same core.
     const beats_at_kill = process.write_count;
@@ -1423,6 +1476,12 @@ fn processKillTest(boot_information: *const BootInformation) void {
     check("the sleeper's exit notification arrived (length 0)", r == 0);
     check("its badge carries the exit bit and the child id", badge == abi.notify_badge_bit | abi.notify_exit_bit | sleeper);
 
+    // M17.2: the recorded reason — the notification is the fence, so it is
+    // already readable, and gated by the same supervisor check as the kill.
+    check("the sleeper's reason reads killed", process.exitReasonOf(me, sleeper) == @intFromEnum(abi.ExitReason.killed));
+    check("a non-supervisor may not read the reason (-EPERM)", process.exitReasonOf(me + 12345, sleeper) == -ipcsync.EPERM);
+    check("an unknown id has no reason (-ESRCH)", process.exitReasonOf(me, 0xFFFF_FF00) == -ipcsync.ESRCH);
+
     const beats_at_kill = process.write_count;
     scheduler.sleep(1500); // more than one heartbeat period
     check("the heartbeat stopped with the kill", process.write_count == beats_at_kill);
@@ -1443,6 +1502,21 @@ fn processKillTest(boot_information: *const BootInformation) void {
     check("the spinner's exit notification arrived (length 0)", r == 0);
     check("its badge carries the exit bit and the child id", badge == abi.notify_badge_bit | abi.notify_exit_bit | spinner);
 
+    // M17.2: a child that ends on its own must read exited, not killed —
+    // args-echo with arguments echoes once and returns from main.
+    var clean: u32 = 0;
+    i = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "args-echo")) continue;
+        clean = process.spawnProcessSupervised(item.blob, 4, &.{ "args-echo", "clean-exit" }, me, endpoint) catch 0;
+        break;
+    }
+    check("args-echo spawned as the clean-exit child", clean != 0);
+    r = ipcsync.replyWait(endpoint, 0, 0, 0, 0, abi.no_cap, &badge, &received_cap);
+    check("the clean child's exit notification arrived", badge == abi.notify_badge_bit | abi.notify_exit_bit | clean);
+    check("the clean child's reason reads exited", process.exitReasonOf(me, clean) == @intFromEnum(abi.ExitReason.exited));
+
     var table: [32]abi.ProcessDescriptor = undefined;
     const total = scheduler.enumerate(&table);
     var still_listed = false;
@@ -1454,6 +1528,390 @@ fn processKillTest(boot_information: *const BootInformation) void {
     result();
 }
 
+/// M17.1: a dead process's device claims are released by the reap, so a restarted
+/// driver can claim its hardware again (docs/process-lifecycle.md iron rule 1).
+/// First the broker release in isolation — two owners, one released, the other's
+/// claim must survive. Then the death-path wiring with a real child: the claim is
+/// made on the child's behalf (the broker is kernel-callable), the child is
+/// killed, and once the exit notification arrives — posted last, after release —
+/// the device must be unclaimed and claimable again.
+fn claimReleaseTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: claim-release\n", .{});
+
+    var buffer: [2]device_abi.DeviceDescriptor = undefined;
+    const total = devices_broker.enumerate(&buffer);
+    check("the device tree is seeded (>= 2 devices)", total >= 2);
+    if (total < 2) {
+        result();
+        return;
+    }
+
+    // The broker release in isolation.
+    check("device 0 claimed by owner 111", devices_broker.claim(0, 111));
+    check("device 1 claimed by owner 222", devices_broker.claim(1, 222));
+    devices_broker.releaseAllOwnedBy(111);
+    check("owner 111's claim is released", devices_broker.ownerOf(0) == null);
+    check("owner 222's claim survives", (devices_broker.ownerOf(1) orelse 0) == 222);
+    devices_broker.releaseAllOwnedBy(222);
+    check("cleanup released owner 222", devices_broker.ownerOf(1) == null);
+
+    // The death-path wiring: a real process dies holding a claim.
+    check("bootloader handed over /system/services/init", boot_information.init_len != 0);
+    if (boot_information.init_len == 0) {
+        result();
+        return;
+    }
+    const image = @as([*]const u8, @ptrFromInt(boot_handoff.physicalToVirtual(boot_information.init_base)))[0..boot_information.init_len];
+    const me = scheduler.currentId();
+    const endpoint = ipcsync.createIpcEndpoint() orelse {
+        check("exit endpoint allocated", false);
+        result();
+        return;
+    };
+    const child = process.spawnProcessSupervised(image, 4, &.{"/system/services/init"}, me, endpoint) catch 0;
+    check("supervised child spawned", child != 0);
+    check("device 0 claimed on the child's behalf", devices_broker.claim(0, child));
+
+    check("the kill is accepted", process.killProcess(me, child) == 0);
+    var badge: u64 = 0;
+    var received_cap: u64 = 0;
+    _ = ipcsync.replyWait(endpoint, 0, 0, 0, 0, abi.no_cap, &badge, &received_cap);
+    check("the exit notification arrived", badge == abi.notify_badge_bit | abi.notify_exit_bit | child);
+    check("death released the child's claim", devices_broker.ownerOf(0) == null);
+    check("the device is claimable again", devices_broker.claim(0, me));
+    devices_broker.releaseAllOwnedBy(me);
+    result();
+}
+
+/// M17.3: the published exit events, proven by their first subscriber. The VFS
+/// subscribes at startup; a client opens a file and parks holding the handle;
+/// the kill posts the exit event to the VFS's endpoint; the VFS releases the
+/// dead client's handle and says so — the service-side mirror of iron rule 1
+/// (a service must never depend on clients cleaning up after themselves).
+fn vfsClientDeathTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: vfs-client-death\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;
+    };
+
+    process.write_count = 0;
+    check("vfs spawned", spawnNamed(rd, "vfs"));
+
+    const me = scheduler.currentId();
+    const endpoint = ipcsync.createIpcEndpoint() orelse {
+        check("exit endpoint allocated", false);
+        result();
+        return;
+    };
+    var client: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "vfs-test")) continue;
+        client = process.spawnProcessSupervised(item.blob, 4, &.{ "vfs-test", "park" }, me, endpoint) catch 0;
+        break;
+    }
+    check("parked client spawned (supervised)", client != 0);
+
+    // Its heartbeat is the fence: once it beats, the handle is open.
+    const parked = "vfstest: parked";
+    scheduler.setPriority(1);
+    var deadline = architecture.millis() + 10000;
+    while (architecture.millis() < deadline) {
+        if (process.write_len >= parked.len and eql(process.write_buffer[0..parked.len], parked)) break;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+    check("client parked holding an open handle", process.write_len >= parked.len and eql(process.write_buffer[0..parked.len], parked));
+
+    check("the kill is accepted", process.killProcess(me, client) == 0);
+    var badge: u64 = 0;
+    var received_cap: u64 = 0;
+    _ = ipcsync.replyWait(endpoint, 0, 0, 0, 0, abi.no_cap, &badge, &received_cap);
+    check("the exit notification arrived", badge == abi.notify_badge_bit | abi.notify_exit_bit | client);
+
+    // The VFS heard the same published event; its release line is the proof.
+    const released = "vfs: released 1 handle(s) for dead client";
+    scheduler.setPriority(1);
+    deadline = architecture.millis() + 10000;
+    while (architecture.millis() < deadline) {
+        if (process.write_len >= released.len and eql(process.write_buffer[0..released.len], released)) break;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+    check("the VFS released the dead client's handle", process.write_len >= released.len and eql(process.write_buffer[0..released.len], released));
+    result();
+}
+
+/// M17.4 from ring 3: process-test's signal-run role drives the whole lifecycle
+/// surface — the zero-length ping (answered by the harness), signals as
+/// statements (reload logged, terminate = clean exit), the one-shot timer, and
+/// both endings of the stop sequence (polite -> exited, deaf -> killed at the
+/// deadline). Its "process-test: signals ok" is the pass marker.
+fn signalsTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: signals\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;
+    };
+
+    process.setInitialRamdisk(image); // the parent system_spawns its children by name
+    process.write_count = 0;
+    var runner: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "process-test")) continue;
+        runner = process.spawnProcessSupervised(item.blob, 4, &.{ "process-test", "signal-run" }, scheduler.currentId(), null) catch 0;
+        break;
+    }
+    check("signal-run parent spawned", runner != 0);
+
+    const pass_marker = "process-test: signals ok";
+    const fail_marker = "process-test: FAIL";
+    scheduler.setPriority(1);
+    const deadline = architecture.millis() + 15000;
+    var saw_pass = false;
+    var saw_fail = false;
+    while (architecture.millis() < deadline and !saw_pass and !saw_fail) {
+        if (process.write_len >= pass_marker.len and eql(process.write_buffer[0..pass_marker.len], pass_marker)) saw_pass = true;
+        if (process.write_len >= fail_marker.len and eql(process.write_buffer[0..fail_marker.len], fail_marker)) saw_fail = true;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+    check("the signal-run parent reported ok", saw_pass and !saw_fail);
+    result();
+}
+
+/// M18.1: the device manager's restart machinery, end to end. In test-restart
+/// mode the manager also supervises crash-test: a fixture that claims device 0,
+/// hellos, and faults. The scenario asserts three markers in order — the real
+/// xHCI driver hellos clean and stays; crash-test is restarted with backoff
+/// (each respawn re-claiming the device the dead instance held, M17.1 through
+/// the manager's path); the crash loop caps and the manager gives up.
+fn driverRestartTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: driver-restart\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;
+    };
+
+    process.setInitialRamdisk(image); // the manager system_spawns drivers by name
+    process.write_count = 0;
+    var manager: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "device-manager")) continue;
+        manager = process.spawnProcessSupervised(item.blob, 4, &.{ "device-manager", "test-restart" }, scheduler.currentId(), null) catch 0;
+        break;
+    }
+    check("device-manager spawned in test-restart mode", manager != 0);
+    // The assertions live in the harness: its expect regex requires, in order,
+    // the xHCI hello ack, a crash-test restart, and the crash-loop cap — read
+    // from the whole serial capture, immune to the transient-line races a
+    // write_buffer poll would have here (many processes log concurrently).
+    result();
+}
+
+/// M18.2: bus tree reports, end to end. The manager (test-usb-restart mode)
+/// spawns the xHCI driver; the driver maps its BAR, scans the root-hub ports,
+/// and reports the two QEMU devices; the manager mirrors them, kills the
+/// reporter (the test trigger), prunes both children, restarts the driver with
+/// backoff, and the respawned instance re-claims, re-scans, and re-reports.
+/// The harness's ordered expect regex is the assertion; this test only
+/// orchestrates the spawn.
+fn usbReportTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: usb-report\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;
+    };
+
+    process.setInitialRamdisk(image);
+    var manager: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "device-manager")) continue;
+        manager = process.spawnProcessSupervised(item.blob, 4, &.{ "device-manager", "test-usb-restart" }, scheduler.currentId(), null) catch 0;
+        break;
+    }
+    check("device-manager spawned in test-usb-restart mode", manager != 0);
+    result();
+}
+
+/// M18.3: the application surface. device-list enumerates the manager's tree
+/// over IPC, subscribes with its endpoint as a capability, and prints every
+/// published event; the manager's delayed test-kill of the reporter produces a
+/// removed/added storm the subscriber must observe. The harness's ordered
+/// expect regex is the assertion.
+fn deviceListTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: device-list\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;
+    };
+
+    process.setInitialRamdisk(image);
+    var manager: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "device-manager")) continue;
+        manager = process.spawnProcessSupervised(item.blob, 4, &.{ "device-manager", "test-usb-restart" }, scheduler.currentId(), null) catch 0;
+        break;
+    }
+    check("device-manager spawned in test-usb-restart mode", manager != 0);
+    check("device-list spawned", spawnNamed(rd, "device-list"));
+    result();
+}
+
+/// M19.1: the ring-3 PCI scan agrees with the kernel's. The manager spawns
+/// pci-bus for the host bridge; the driver walks the same ECAM window through
+/// its mmio_map grant and must find exactly the functions the kernel's own
+/// enumeration recorded — the equivalence that licenses retiring the kernel
+/// walk in M19.3.
+fn pciScanTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: pci-scan\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;
+    };
+
+    // Post-flip (M19.3) ground truth: the kernel no longer enumerates PCI
+    // functions, so equivalence inverts — the broker's function count after
+    // the scan must equal what the driver itself reported finding.
+    var buffer: [64]device_abi.DeviceDescriptor = undefined;
+    const n = @min(devices_broker.enumerate(&buffer), buffer.len);
+    var boot_pci: u32 = 0;
+    for (buffer[0..n]) |d| {
+        if (d.class == @intFromEnum(device_abi.DeviceClass.pci_device)) boot_pci += 1;
+    }
+    check("the kernel seeded no PCI functions (the walk retired)", boot_pci == 0);
+
+    process.setInitialRamdisk(image);
+    process.write_count = 0;
+    var manager: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "device-manager")) continue;
+        manager = process.spawnProcessSupervised(item.blob, 4, &.{ "device-manager", "test-pci-restart" }, scheduler.currentId(), null) catch 0;
+        break;
+    }
+    check("device-manager spawned (test-pci-restart mode)", manager != 0);
+
+    // First scan: wait for the driver's count line and parse the number.
+    const count_prefix = "pci-bus: ";
+    const count_suffix = " functions found";
+    var reported: u32 = 0;
+    scheduler.setPriority(1);
+    var deadline = architecture.millis() + 15000;
+    while (architecture.millis() < deadline and reported == 0) {
+        if (process.write_len > count_prefix.len + count_suffix.len and eql(process.write_buffer[0..count_prefix.len], count_prefix)) {
+            const line = process.write_buffer[0..process.write_len];
+            const digits_end = std.mem.indexOf(u8, line, count_suffix) orelse {
+                scheduler.yield();
+                continue;
+            };
+            reported = std.fmt.parseInt(u32, line[count_prefix.len..digits_end], 10) catch 0;
+        }
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+    check("the ring-3 scan reported a function count", reported >= 1);
+
+    // Every reported function was registered: the broker holds exactly them.
+    var registered: [64]device_abi.DeviceDescriptor = undefined;
+    const r = @min(devices_broker.enumerate(&registered), registered.len);
+    var registered_pci: u32 = 0;
+    for (registered[0..r]) |d| {
+        if (d.class == @intFromEnum(device_abi.DeviceClass.pci_device)) registered_pci += 1;
+    }
+    check("the broker holds exactly the reported functions", registered_pci == reported);
+    const kernel_count = reported; // the no-duplicate check below reuses it
+
+    // The restart drill: the manager kills pci-bus after its reports; the
+    // respawn re-claims, re-scans, and re-registers.
+    const restart_marker = "device-manager: restarting pci-bus";
+    scheduler.setPriority(1);
+    deadline = architecture.millis() + 15000;
+    var restarted = false;
+    while (architecture.millis() < deadline and !restarted) {
+        if (process.write_len >= restart_marker.len and eql(process.write_buffer[0..restart_marker.len], restart_marker)) restarted = true;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+    check("the manager restarted pci-bus", restarted);
+
+    var marker_buffer: [48]u8 = undefined;
+    const marker = std.fmt.bufPrint(&marker_buffer, "pci-bus: {d} functions found", .{reported}) catch "";
+    scheduler.setPriority(1);
+    deadline = architecture.millis() + 15000;
+    var seen = false;
+    while (architecture.millis() < deadline and !seen) {
+        if (process.write_len >= marker.len and eql(process.write_buffer[0..marker.len], marker)) seen = true;
+        scheduler.yield();
+    }
+    scheduler.setPriority(4);
+    check("the respawned scan reported the same count", seen);
+
+    // No duplicates: the registrations deduped against the kernel's own nodes
+    // on the first pass, and against themselves on the second.
+    var after: [64]device_abi.DeviceDescriptor = undefined;
+    const m = @min(devices_broker.enumerate(&after), after.len);
+    var after_count: u32 = 0;
+    for (after[0..m]) |d| {
+        if (d.class == @intFromEnum(device_abi.DeviceClass.pci_device)) after_count += 1;
+    }
+    check("no duplicate PCI nodes after register + restart + re-register", after_count == kernel_count);
+    result();
+}
+
 /// The whole user-side surface at once: spawn process-test's supervisor role,
 /// which — entirely from ring 3 — creates an exit endpoint, spawns its two
 /// children supervised, sees them in process_enumerate, kills them (one blocked,
@@ -1774,6 +2232,35 @@ fn hpetGsi() ?u32 {
 /// land in the device table with the containment invariant intact.
 fn busTest(boot_information: *const BootInformation) void {
     log("DANOS-TEST-BEGIN: bus\n", .{});
+
+    // M19.0: device_register is idempotent on exact match — a restarted
+    // registering bus must not duplicate its children. Driven directly against
+    // the broker: claim an unclaimed node, register the same (class, hid,
+    // resourceless) child twice, expect one id and one table entry.
+    {
+        const me = scheduler.currentId();
+        var probe: [1]device_abi.DeviceDescriptor = undefined;
+        const total = devices_broker.enumerate(&probe);
+        check("device tree is seeded for the idempotence check", total >= 1);
+        if (devices_broker.ownerOf(0) == null) {
+            check("claimed device 0 for the idempotence check", devices_broker.claim(0, me));
+            var child = std.mem.zeroes(device_abi.DeviceDescriptor);
+            child.class = @intFromEnum(device_abi.DeviceClass.unknown);
+            child.pci_class = device_abi.no_pci_class;
+            child.hid_len = 4;
+            child.hid[0..4].* = "idem".*;
+            const first = devices_broker.register(0, me, &child) catch 0;
+            check("first register succeeded", first != 0);
+            const before = devices_broker.enumerate(&probe);
+            const second = devices_broker.register(0, me, &child) catch 0;
+            check("re-register returned the same id", second == first);
+            check("re-register grew nothing", devices_broker.enumerate(&probe) == before);
+            devices_broker.releaseAllOwnedBy(me);
+        } else {
+            check("device 0 unexpectedly claimed before the idempotence check", false);
+        }
+    }
+
     if (boot_information.initial_ramdisk_len == 0) {
         check("bootloader handed over an initial_ramdisk", false);
         result();

system/parameters.zig

@@ -16,8 +16,11 @@
 pub const maximum_cpus = 128;
 
 /// Maximum tasks (kernel threads) alive at once — the static task-table size. Each
-/// online core consumes one slot for its idle task, plus task 0 on the BSP.
-pub const maximum_tasks = 16;
+/// online core consumes one slot for its idle task, plus task 0 on the BSP. Sized
+/// for the initial-ramdisk sweep (15 bundled binaries spawned at once) plus the
+/// device manager's supervised children with room to grow — at 16 the sweep
+/// started failing spawns once the bundle passed a dozen binaries.
+pub const maximum_tasks = 32;
 
 /// Each task's kernel stack (also each AP's bring-up stack), in bytes.
 pub const kernel_stack_size = 16 * 1024;

system/services/acpi/acpi.zig

@@ -0,0 +1,30 @@
+//! /system/services/acpi — the ACPI discovery service: the x86 firmware
+//! interpreter, moved out of ring 0 (docs/m19-m20-plan.md, M20). **Placeholder:
+//! not implemented until M20.1** — it exists so the build's `-Ddiscovery`
+//! option has both of its values and the ramdisk's neutral `discovery` slot is
+//! wired before the implementation lands.
+//!
+//! What it becomes (the plan's decisions 5 and 7): claim the `acpi-tables`
+//! node the kernel publishes (table blobs + the broad io_port grant + the SCI),
+//! map the tables, and run the **shared AML module** in ring 3 behind a `Hal`
+//! backed by `mmio_map` and `io_read`/`io_write` — the interpreter cannot tell
+//! it moved. Then the bus-driver shape: `device_register` the namespace
+//! devices (`_HID`, `_CRS` resources, containment against the node's
+//! apertures), report each to the device manager, stay resident under its
+//! supervision. M21 grows the event side on the same claim: the SCI, PM1 fixed
+//! events, GPEs, Notify — published through the domain-named power protocol,
+//! never an "ACPI events" protocol.
+
+const runtime = @import("runtime");
+
+pub fn main() void {
+    // Not implemented: exit cleanly and silently (a bare spawn by the
+    // initial-ramdisk sweep must not derange other tests' markers). The
+    // supervisor reads a clean exit as "meant to stop" — correct for a
+    // placeholder.
+}
+
+pub const panic = runtime.panic;
+comptime {
+    _ = &runtime.start._start; // pull the runtime entry shim into the image
+}

system/services/crash-test/crash-test.zig

@@ -0,0 +1,44 @@
+//! crash-test — a test fixture, not a driver: claims the device it is assigned,
+//! hellos the device manager, announces itself, then faults on purpose. The
+//! driver-restart scenario drives the manager's whole restart machinery with
+//! it: fault → exit reason → backoff → respawn → the **same claim succeeding
+//! again** (claim release on death, M17.1, through the manager's path) → the
+//! crash-loop cap. Spawned bare (the initial-ramdisk sweep starts every bundled
+//! binary), it exits silently so it cannot derange other tests.
+
+const std = @import("std");
+const runtime = @import("runtime");
+const protocol = runtime.device_manager_protocol;
+
+pub fn main(init: runtime.process.Init) void {
+    const argument = init.arguments.get(1) orelse return; // bare: stay silent
+    const assigned = std.fmt.parseInt(u64, argument, 10) catch return;
+
+    // The respawn only reaches this line because the kernel released the
+    // previous instance's claim at death. A failed claim exits cleanly — the
+    // manager reads "meant to stop" and the scenario fails loudly by silence.
+    if (!runtime.device.claim(assigned)) {
+        _ = runtime.system.write("crash-test: claim failed\n");
+        return;
+    }
+
+    var manager: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (manager == null and tries < 100) : (tries += 1) {
+        manager = runtime.ipc.lookup(.device_manager);
+        if (manager == null) runtime.system.sleep(20);
+    }
+    const h = manager orelse return;
+    const hello = protocol.Hello{ .role = @intFromEnum(protocol.Role.device), .device_id = assigned };
+    var reply: [protocol.message_maximum]u8 = undefined;
+    _ = runtime.ipc.call(h, std.mem.asBytes(&hello), &reply) catch return;
+
+    _ = runtime.system.write("crash-test: faulting now\n");
+    const poison: *volatile u32 = @ptrFromInt(0xdead0000);
+    poison.* = 1; // the restart machinery's fuel: a real segmentation fault
+}
+
+pub const panic = runtime.panic;
+comptime {
+    _ = &runtime.start._start; // pull the runtime entry shim into the image
+}

system/services/device-list/device-list.zig

@@ -0,0 +1,88 @@
+//! device-list — the `ps` analog for the device tree (docs/device-manager.md
+//! M18.3): asks the device manager for the tree over IPC, prints it, then
+//! subscribes and prints every published add/remove event. The manager is the
+//! one answer to "what devices exist" for user space; nothing here touches a
+//! device_* system call.
+
+const std = @import("std");
+const runtime = @import("runtime");
+const protocol = runtime.device_manager_protocol;
+
+fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
+    var line: [96]u8 = undefined;
+    _ = runtime.system.write(std.fmt.bufPrint(&line, fmt, arguments) catch return);
+}
+
+pub fn main() void {
+    var manager: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (manager == null and tries < 200) : (tries += 1) {
+        manager = runtime.ipc.lookup(.device_manager);
+        if (manager == null) runtime.system.sleep(20);
+    }
+    const h = manager orelse {
+        _ = runtime.system.write("device-list: no device manager\n");
+        return;
+    };
+
+    // The snapshot — polled briefly, because at boot the bus drivers may still
+    // be scanning: an empty first answer usually just means "too early".
+    var reply: [protocol.message_maximum]u8 = undefined;
+    var count: u32 = 0;
+    var length: usize = 0;
+    tries = 0;
+    while (tries < 20) : (tries += 1) {
+        const request = protocol.Enumerate{};
+        length = runtime.ipc.call(h, std.mem.asBytes(&request), &reply) catch 0;
+        if (length >= @sizeOf(protocol.EnumerateReply)) {
+            count = std.mem.bytesToValue(protocol.EnumerateReply, reply[0..@sizeOf(protocol.EnumerateReply)]).count;
+            if (count != 0) break;
+        }
+        runtime.system.sleep(100);
+    }
+    writeLine("device-list: {d} devices\n", .{count});
+    var offset: usize = @sizeOf(protocol.EnumerateReply);
+    var index: u32 = 0;
+    while (index < count and offset + @sizeOf(protocol.ChildEntry) <= length) : (index += 1) {
+        const entry = std.mem.bytesToValue(protocol.ChildEntry, reply[offset..][0..@sizeOf(protocol.ChildEntry)]);
+        writeLine("device-list: device {d} port {d} identity {d}\n", .{ entry.parent, entry.bus_address, entry.identity });
+        offset += @sizeOf(protocol.ChildEntry);
+    }
+
+    // The subscription: our endpoint rides as the call's capability; events
+    // arrive as buffered messages carrying the same structs the bus sends.
+    const endpoint = runtime.ipc.createIpcEndpoint() orelse {
+        _ = runtime.system.write("device-list: no endpoint\n");
+        return;
+    };
+    const subscribe = protocol.Subscribe{};
+    _ = runtime.ipc.callCap(h, std.mem.asBytes(&subscribe), &reply, endpoint) catch {
+        _ = runtime.system.write("device-list: subscribe failed\n");
+        return;
+    };
+    _ = runtime.system.write("device-list: subscribed\n");
+
+    var receive: [protocol.message_maximum]u8 = undefined;
+    while (true) {
+        const got = runtime.ipc.replyWait(endpoint, &.{}, &receive, null);
+        if (!got.isMessage() or got.len < 1) continue;
+        switch (receive[0]) {
+            @intFromEnum(protocol.Operation.child_added) => {
+                if (got.len < protocol.child_added_size) continue;
+                const event = std.mem.bytesToValue(protocol.ChildAdded, receive[0..protocol.child_added_size]);
+                writeLine("device-list: added (device {d} port {d})\n", .{ event.parent, event.bus_address });
+            },
+            @intFromEnum(protocol.Operation.child_removed) => {
+                if (got.len < protocol.child_removed_size) continue;
+                const event = std.mem.bytesToValue(protocol.ChildRemoved, receive[0..protocol.child_removed_size]);
+                writeLine("device-list: removed (device {d} port {d})\n", .{ event.parent, event.bus_address });
+            },
+            else => {},
+        }
+    }
+}
+
+pub const panic = runtime.panic;
+comptime {
+    _ = &runtime.start._start; // pull the runtime entry shim into the image
+}

system/services/device-manager/device-manager-protocol.zig

@@ -0,0 +1,144 @@
+//! The device-manager protocol (docs/device-manager.md): what drivers and
+//! applications say to the device manager over its well-known endpoint. The
+//! vfs-protocol pattern — extern-struct messages, a version in the handshake,
+//! reserved fields — so both sides depend on the contract by name. Deliberately
+//! contains nothing lifecycle-shaped: stopping, liveness (the zero-length ping),
+//! and exit reasons are the universal vocabulary of
+//! docs/process-lifecycle.md, not this protocol.
+
+/// The protocol version a driver states in its hello. A manager that cannot
+/// serve a driver's version refuses the hello, and the mismatch is loud at
+/// startup instead of quiet corruption later.
+pub const version: u16 = 1;
+
+/// What kind of driver is talking (docs/driver-model.md's shapes).
+pub const Role = enum(u8) {
+    /// Owns a controller and reports the devices behind it (`child_added`).
+    bus = 1,
+    /// Serves one device, reached through a bus's transfer protocol.
+    device = 2,
+};
+
+/// The message kinds.
+pub const Operation = enum(u8) {
+    hello = 1,
+    child_added = 2,
+    child_removed = 3,
+    enumerate = 4,
+    subscribe = 5,
+};
+
+/// `Hello.device_id` for a driver that serves no enumerated device (a test
+/// fixture, a synthetic source).
+pub const no_device: u64 = ~@as(u64, 0);
+
+/// The handshake, sent once by every driver the manager spawns — the manager's
+/// one self-enforced deadline: spawned and silent past it means wrong binary,
+/// wrong version, or wedged before main, and the stop sequence follows.
+pub const Hello = extern struct {
+    operation: u8 = @intFromEnum(Operation.hello),
+    /// A Role value.
+    role: u8,
+    /// The protocol version this driver was built against (`version`).
+    version: u16 = version,
+    reserved: u32 = 0,
+    /// The device this driver was assigned (its argv[1]), or `no_device`.
+    device_id: u64,
+};
+
+pub const hello_size = @sizeOf(Hello);
+
+/// The manager's answer to a hello. Nonzero status = refused (version mismatch,
+/// unknown sender); a refused driver should exit cleanly.
+pub const HelloReply = extern struct {
+    status: i32,
+    reserved: u32 = 0,
+};
+
+pub const reply_size = @sizeOf(HelloReply);
+
+/// A bus driver reporting one device it discovered behind its controller
+/// (docs/device-manager.md "the tree"). Identity is the bus's native language —
+/// for USB a port-speed class; the (class, subclass, protocol) triple joins it
+/// once control transfers exist (the USB track). The manager mirrors the child
+/// into its tree; when the reporting driver dies, the manager prunes everything
+/// it reported (the children describe protocol state that died with it) and the
+/// restarted instance rediscovers and re-reports.
+pub const ChildAdded = extern struct {
+    operation: u8 = @intFromEnum(Operation.child_added),
+    reserved0: u8 = 0,
+    reserved1: u16 = 0,
+    reserved2: u32 = 0,
+    /// The reporting driver's own device (the controller) — the child's parent.
+    parent: u64,
+    /// Where on the bus (for USB: the root port number, 1-based).
+    bus_address: u64,
+    /// Bus-specific identity (for USB: the PORTSC port-speed class; for PCI:
+    /// the class triple).
+    identity: u64,
+    /// The kernel device id this child was `device_register`ed as — what the
+    /// manager hands a matched driver as its argv assignment — or `no_device`
+    /// for an unregistered leaf (a USB port before the descriptor track).
+    device_id: u64 = no_device,
+};
+
+pub const child_added_size = @sizeOf(ChildAdded);
+
+/// A bus driver reporting a device gone (hot-unplug). Not yet sent by any
+/// driver — the port scan has no unplug interrupt — but the manager handles it;
+/// death-pruning covers removal until hotplug lands.
+pub const ChildRemoved = extern struct {
+    operation: u8 = @intFromEnum(Operation.child_removed),
+    reserved0: u8 = 0,
+    reserved1: u16 = 0,
+    reserved2: u32 = 0,
+    parent: u64,
+    bus_address: u64,
+};
+
+pub const child_removed_size = @sizeOf(ChildRemoved);
+
+/// The manager's answer to a tree report.
+pub const ReportReply = extern struct {
+    status: i32,
+    reserved: u32 = 0,
+};
+
+/// An application asking for the tree (M18.3): the reply is an EnumerateReply
+/// header followed by `count` ChildEntry records.
+pub const Enumerate = extern struct {
+    operation: u8 = @intFromEnum(Operation.enumerate),
+    reserved0: u8 = 0,
+    reserved1: u16 = 0,
+    reserved2: u32 = 0,
+};
+
+pub const EnumerateReply = extern struct {
+    status: i32,
+    /// ChildEntry records following this header.
+    count: u32,
+};
+
+pub const ChildEntry = extern struct {
+    parent: u64,
+    bus_address: u64,
+    identity: u64,
+};
+
+/// An application subscribing to published add/remove events (the input-service
+/// pattern): the subscriber's endpoint rides as the call's **capability**, and
+/// events arrive on it as buffered messages whose payload is the same
+/// ChildAdded / ChildRemoved struct the bus drivers send — one encoding, both
+/// directions.
+pub const Subscribe = extern struct {
+    operation: u8 = @intFromEnum(Operation.subscribe),
+    reserved0: u8 = 0,
+    reserved1: u16 = 0,
+    reserved2: u32 = 0,
+};
+
+/// Upper bound on any message in this protocol — sizes the endpoint buffers.
+/// Capped by the kernel's IPC MESSAGE_MAXIMUM (256): an EnumerateReply carries
+/// up to ten ChildEntry records per call, plenty for the mirror's current
+/// bounds; paging joins the protocol if a tree ever outgrows one message.
+pub const message_maximum = 256;

system/services/device-manager/device-manager.zig

@@ -1,20 +1,24 @@
 //! /system/services/device-manager — the ring-3 process that turns the device
-//! tree into a running system. The kernel enumerates the hardware and enforces the
-//! claim capability (mechanism); this decides *which driver serves which device*
-//! and, eventually, spawns it (policy). Keeping that split in user space is the
-//! whole point of the microkernel: the manager is an ordinary, restartable process
-//! with no special privilege — it uses the same `device_*` system calls any process
-//! could ([drivers.md](../../../docs/drivers.md), [driver-model.md]).
+//! tree into a running system: **the matcher and the supervisor**
+//! (docs/device-manager.md). The kernel enumerates the hardware and enforces the
+//! claim capability (mechanism); this decides which driver serves which device,
+//! spawns it, and keeps it alive (policy). Keeping that split in user space is
+//! the whole point of the microkernel: the manager is an ordinary, restartable
+//! process with no special privilege.
 //!
-//! Increment 2 (this file): enumerate /system/devices, *match* each device to a
-//! driver, and *spawn* it with `system_spawn` — the kernel loads the named binary
-//! from the initial-ramdisk as a fresh ring-3 process. On QEMU this discovers the
-//! HPET, decides `hpet` serves it, and brings that driver all the way up. (The
-//! kernel still auto-spawns the whole initial-ramdisk at boot; increment 3 removes
-//! that redundancy so the manager is the sole owner of driver spawning.)
+//! M18.1 (this increment): the manager is a harness service on the well-known
+//! `.device_manager` endpoint. Every driver is spawned **supervised** — exit
+//! notifications land in the same loop as protocol messages. Drivers with an
+//! assignment must `hello` within a deadline or be stopped; a driver that dies
+//! is restarted with backoff, and a crash loop (three fast deaths) marks it
+//! failed instead of respawning forever. Exit reasons (M17.2) drive the
+//! decision: a clean exit meant to stop; only faults and missed deadlines
+//! restart. Tree reports (`child_added`) land in M18.2.
+
 const std = @import("std");
 const runtime = @import("runtime");
 const acpi_ids = @import("acpi-ids");
+const protocol = runtime.device_manager_protocol;
 const device = runtime.device;
 const system = runtime.system;
 
@@ -27,9 +31,8 @@ fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
 }
 
 /// The driver that serves each device — the policy table. In a fuller system
-/// this comes from the drivers describing what they bind (or a manifest under
-/// /system/drivers); for now it is a small static map, which is enough to prove the
-/// manager reads the tree and decides. `null` = no driver for this class yet.
+/// this comes from a manifest (docs/device-manager.md: the third bus type
+/// triggers it); for now a static map. `null` = no driver for this class yet.
 fn driverFor(d: device.DeviceDescriptor) ?[]const u8 {
     // detect device via DeviceClass
     if (d.class == @intFromEnum(device.DeviceClass.timer)) return "hpet";
@@ -47,66 +50,474 @@ fn driverFor(d: device.DeviceDescriptor) ?[]const u8 {
 /// pci-class.zig decodes.
 const xhci_pci_class: u64 = 0x0C_03_30;
 
-/// The bus driver that serves a PCI function, or null. Unlike the singleton drivers
-/// in `driverFor`, a machine can carry several identical controllers — so the caller
-/// spawns one driver instance *per device*, passing the device id as argv[1] for the
-/// instance to claim.
-fn pciDriverFor(d: device.DeviceDescriptor) ?[]const u8 {
-    if (d.class != @intFromEnum(device.DeviceClass.pci_device)) return null;
-    return switch (d.pci_class) {
+/// The driver that serves a *reported* PCI function (M19.3: matching moved
+/// from the boot snapshot to the bus reports), or null. A machine can carry
+/// several identical controllers — one driver instance per reported device,
+/// its registered id as argv[1].
+fn pciDriverForIdentity(identity: u64) ?[]const u8 {
+    return switch (identity) {
         xhci_pci_class => "usb-xhci-bus",
         else => null,
     };
 }
 
-/// Spawn one instance of `driver_name` to serve the specific device `id` — the id
-/// arrives as argv[1]. No isProcessRunning gate here: the name alone cannot tell two
-/// instances apart, and this manager is the sole spawner of drivers.
-fn spawnForDevice(driver_name: []const u8, id: u64) void {
-    var text: [20]u8 = undefined;
-    const id_text = std.fmt.bufPrint(&text, "{d}", .{id}) catch return;
-    if (system.spawnWithArguments(driver_name, &.{id_text}) != null) {
-        writeLine("device-manager: spawned {s} for device {d}\n", .{ driver_name, id });
-    } else {
-        writeLine("device-manager: failed to spawn {s} for device {d}\n", .{ driver_name, id });
+/// Whether some driver entry already serves registered device `device_id` —
+/// a re-report after a bus restart must not spawn a second instance.
+fn driverForDevice(device_id: u64) bool {
+    for (&drivers) |*driver| {
+        if (driver.used and driver.device_id == device_id) return true;
+    }
+    return false;
+}
+
+// --- supervision -------------------------------------------------------------
+
+/// How long a protocol driver has to hello after its spawn.
+const hello_deadline_ms: u64 = 3000;
+/// Deaths faster than this count toward the crash loop; slower ones reset it.
+const fast_death_ns: u64 = 2_000_000_000;
+/// Consecutive fast deaths before the manager gives up on a driver.
+const crash_loop_cap: u32 = 3;
+/// Restart backoff: base << (restarts - 1), so 300 ms, 600 ms, 1200 ms.
+const backoff_base_ms: u64 = 300;
+
+const DriverState = enum {
+    awaiting_hello, // spawned; the deadline is armed (protocol drivers only)
+    running,
+    restarting, // dead; respawn due at restart_due_ns
+    stopped, // exited cleanly — it meant to; not restarted
+    failed, // crash loop, or unspawnable; the manager gave up
+};
+
+const Driver = struct {
+    used: bool = false,
+    name_buffer: [24]u8 = undefined,
+    name_len: usize = 0,
+    // The assigned device id (becomes argv[1]), or protocol.no_device.
+    device_id: u64 = protocol.no_device,
+    // Whether this driver speaks the protocol (hello expected, deadline
+    // enforced). Legacy drivers (hpet, ps2-bus) are supervised and restarted
+    // but not yet required to hello.
+    speaks_protocol: bool = false,
+    process_id: u32 = 0,
+    state: DriverState = .running,
+    restarts: u32 = 0,
+    spawn_ns: u64 = 0,
+    hello_deadline_ns: u64 = 0,
+    restart_due_ns: u64 = 0,
+
+    fn name(driver: *const Driver) []const u8 {
+        return driver.name_buffer[0..driver.name_len];
+    }
+};
+
+const maximum_drivers = 16;
+var drivers: [maximum_drivers]Driver = .{Driver{}} ** maximum_drivers;
+var manager_endpoint: runtime.ipc.Handle = 0;
+var test_restart_mode = false;
+var test_usb_restart_mode = false;
+var test_usb_killed = false;
+var test_pci_restart_mode = false;
+var test_kill_pid: u32 = 0;
+var test_kill_due_ns: u64 = 0;
+
+/// The application subscribers (M18.3, the input-service pattern): endpoints
+/// handed over as capabilities, each receiving every child add/remove as a
+/// buffered message. A subscriber whose endpoint stops accepting (it died) is
+/// dropped on the failed send.
+const maximum_subscribers = 8;
+var subscribers: [maximum_subscribers]?runtime.ipc.Handle = .{null} ** maximum_subscribers;
+
+/// Publish one event (a ChildAdded or ChildRemoved struct, the same encoding
+/// the bus drivers send) to every subscriber.
+fn publishEvent(event: []const u8) void {
+    for (&subscribers) |*slot| {
+        if (slot.*) |handle| {
+            if (!runtime.ipc.send(handle, event)) slot.* = null; // dead subscriber
+        }
     }
 }
 
-pub fn main() void {
+/// The manager's mirror of what bus drivers report (docs/device-manager.md "the
+/// tree"): the children, keyed by (parent, bus address), each remembering which
+/// driver instance reported it — that is what death-pruning sweeps by.
+const Child = struct {
+    used: bool = false,
+    parent: u64 = 0,
+    bus_address: u64 = 0,
+    identity: u64 = 0,
+    // The kernel device id (registered by the reporter), or protocol.no_device.
+    device_id: u64 = 0,
+    reporter: u32 = 0, // the reporting driver instance's process id
+};
+
+const maximum_children = 32;
+var children: [maximum_children]Child = .{Child{}} ** maximum_children;
+
+/// Record (or refresh) a reported child. Refreshing matters: a restarted bus
+/// driver re-reports what it rediscovers, and the same (parent, port) must not
+/// duplicate.
+fn addChild(parent: u64, bus_address: u64, identity: u64, device_id: u64, reporter: u32) bool {
+    var free: ?*Child = null;
+    for (&children) |*child| {
+        if (child.used and child.parent == parent and child.bus_address == bus_address) {
+            child.identity = identity;
+            child.device_id = device_id;
+            child.reporter = reporter;
+            return true;
+        }
+        if (!child.used and free == null) free = child;
+    }
+    const slot = free orelse return false;
+    slot.* = .{ .used = true, .parent = parent, .bus_address = bus_address, .identity = identity, .device_id = device_id, .reporter = reporter };
+    return true;
+}
+
+/// Prune every child a dead driver instance reported: the children describe
+/// protocol state (slots, rings) that died with the process — keeping the nodes
+/// would be keeping a lie. The restarted instance rediscovers and re-reports.
+/// Watchers hear the honest story: removed now, added again on rediscovery.
+fn pruneChildrenOf(reporter: u32) void {
+    for (&children) |*child| {
+        if (child.used and child.reporter == reporter) {
+            writeLine("device-manager: child removed (device {d} port {d})\n", .{ child.parent, child.bus_address });
+            child.used = false;
+            const event = protocol.ChildRemoved{ .parent = child.parent, .bus_address = child.bus_address };
+            publishEvent(std.mem.asBytes(&event));
+        }
+    }
+}
+
+/// How many children a driver instance has reported (the test-usb-restart
+/// trigger counts these).
+fn childCountOf(reporter: u32) u32 {
+    var n: u32 = 0;
+    for (&children) |*child| {
+        if (child.used and child.reporter == reporter) n += 1;
+    }
+    return n;
+}
+
+fn driverByProcess(process_id: u32) ?*Driver {
+    for (&drivers) |*driver| {
+        if (driver.used and driver.process_id == process_id) return driver;
+    }
+    return null;
+}
+
+/// Whether a singleton driver is already in the table (two ACPI nodes can both
+/// map to ps2-bus; one instance serves both).
+fn alreadySupervised(name: []const u8) bool {
+    for (&drivers) |*driver| {
+        if (driver.used and std.mem.eql(u8, driver.name(), name)) return true;
+    }
+    return false;
+}
+
+/// Record a driver in the table and spawn its first instance.
+fn addDriver(name: []const u8, device_id: u64, speaks_protocol: bool) void {
+    for (&drivers) |*driver| {
+        if (driver.used) continue;
+        const n = @min(name.len, driver.name_buffer.len);
+        @memcpy(driver.name_buffer[0..n], name[0..n]);
+        driver.name_len = n;
+        driver.device_id = device_id;
+        driver.speaks_protocol = speaks_protocol;
+        driver.used = true;
+        spawnDriver(driver);
+        return;
+    }
+    writeLine("device-manager: driver table full; cannot supervise {s}\n", .{name});
+}
+
+/// (Re)spawn a driver instance: supervised on the manager's own endpoint, the
+/// device id as argv[1] when it has one, the hello deadline armed when it
+/// speaks the protocol.
+fn spawnDriver(driver: *Driver) void {
+    var id_text: [20]u8 = undefined;
+    var arguments: [1][]const u8 = undefined;
+    var argument_count: usize = 0;
+    if (driver.device_id != protocol.no_device) {
+        arguments[0] = std.fmt.bufPrint(&id_text, "{d}", .{driver.device_id}) catch return;
+        argument_count = 1;
+    }
+    const child = system.spawnSupervised(driver.name(), arguments[0..argument_count], manager_endpoint) orelse {
+        writeLine("device-manager: failed to spawn {s}\n", .{driver.name()});
+        driver.state = .failed;
+        return;
+    };
+    driver.process_id = child;
+    driver.spawn_ns = system.clock();
+    if (driver.speaks_protocol) {
+        driver.state = .awaiting_hello;
+        driver.hello_deadline_ns = driver.spawn_ns + hello_deadline_ms * 1_000_000;
+        _ = system.timerOnce(manager_endpoint, hello_deadline_ms + 100);
+    } else {
+        driver.state = .running;
+    }
+    if (driver.device_id != protocol.no_device) {
+        writeLine("device-manager: spawned {s} for device {d}\n", .{ driver.name(), driver.device_id });
+    } else {
+        writeLine("device-manager: spawned {s}\n", .{driver.name()});
+    }
+}
+
+/// A driver died. Prune what it reported first — then the exit reason (M17.2)
+/// is the whole restart decision: a clean exit meant to stop; anything else
+/// restarts with backoff until the crash-loop cap.
+fn onDriverExit(driver: *Driver) void {
+    pruneChildrenOf(driver.process_id);
+    const reason = runtime.process.exitReason(driver.process_id) orelse .fault;
+    if (reason == .exited) {
+        driver.state = .stopped;
+        writeLine("device-manager: {s} exited cleanly; not restarting\n", .{driver.name()});
+        return;
+    }
+    const now = system.clock();
+    const alive_ns = now - driver.spawn_ns;
+    driver.restarts = if (alive_ns < fast_death_ns) driver.restarts + 1 else 1;
+    if (driver.restarts >= crash_loop_cap) {
+        driver.state = .failed;
+        writeLine("device-manager: {s} is failing repeatedly (crash loop); giving up\n", .{driver.name()});
+        return;
+    }
+    const delay_ms = backoff_base_ms << @intCast(driver.restarts - 1);
+    driver.state = .restarting;
+    driver.restart_due_ns = now + delay_ms * 1_000_000;
+    writeLine("device-manager: restarting {s} in {d} ms (died: {s})\n", .{ driver.name(), delay_ms, @tagName(reason) });
+    _ = system.timerOnce(manager_endpoint, delay_ms + 50);
+}
+
+/// A timer landed: sweep every deadline. Overdue hellos are killed (the exit
+/// notification then routes through the normal restart policy); due restarts
+/// respawn. Timers carry no id on purpose — the table is the state, and one
+/// sweep serves every armed deadline.
+fn sweepDeadlines() void {
+    const now = system.clock();
+    if (test_kill_pid != 0 and now >= test_kill_due_ns) {
+        writeLine("device-manager: test mode: killing the reporter\n", .{});
+        _ = system.kill(test_kill_pid);
+        test_kill_pid = 0;
+    }
+    for (&drivers) |*driver| {
+        if (!driver.used) continue;
+        switch (driver.state) {
+            .awaiting_hello => if (now >= driver.hello_deadline_ns) {
+                writeLine("device-manager: {s} missed its hello deadline\n", .{driver.name()});
+                _ = system.kill(driver.process_id);
+                // The exit notification finishes the job via onDriverExit.
+            },
+            .restarting => if (now >= driver.restart_due_ns) spawnDriver(driver),
+            else => {},
+        }
+    }
+}
+
+// --- the harness callbacks -----------------------------------------------------
+
+fn initialise(endpoint: runtime.ipc.Handle) bool {
+    manager_endpoint = endpoint;
+
     // Enumerate into a heap buffer (too big for the one-page user stack).
     const buffer = runtime.allocator().alloc(device.DeviceDescriptor, 64) catch {
         _ = runtime.system.write("device-manager: out of memory\n");
-        return;
+        return false;
     };
     const total = device.enumerate(buffer);
     const count = @min(total, buffer.len);
 
     var matched: usize = 0;
     for (buffer[0..count]) |descriptor| {
-        if (pciDriverFor(descriptor)) |driver_name| {
+        if (descriptor.class == @intFromEnum(device.DeviceClass.pci_host_bridge)) {
+            // The PCI bus driver: enumeration in ring 3 (M19), one instance
+            // per bridge, the bridge id as its assignment.
             matched += 1;
-            spawnForDevice(driver_name, descriptor.id);
+            addDriver("pci-bus", descriptor.id, true);
             continue;
         }
+        // PCI functions no longer appear in the boot snapshot (M19.3): the
+        // pci-bus driver reports them, and onChildAdded matches from reports.
         const driver_name = driverFor(descriptor) orelse continue;
         matched += 1;
-        if (!system.isProcessRunning(driver_name)) {
-            if (runtime.system.spawn(driver_name) != null) {
-                writeLine("device-manager: spawned {s}\n", .{driver_name});
-            } else {
-                writeLine("device-manager: failed to spawn {s}\n", .{driver_name});
-            }
-        } else {
-            writeLine("device-manager: already spawned {s}\n", .{driver_name});
+        // Skip a singleton that is already alive (the initial-ramdisk sweep test
+        // starts every bundled binary bare, this manager included) — spawning a
+        // second instance would only lose the claim race and churn the log.
+        if (!alreadySupervised(driver_name) and !system.isProcessRunning(driver_name)) {
+            addDriver(driver_name, protocol.no_device, false);
         }
     }
 
+    if (test_restart_mode) {
+        // The driver-restart scenario's fixture: claims device 0 (the tree
+        // root, otherwise unclaimed), hellos, then faults — driving backoff,
+        // re-claim-after-death, and the crash-loop cap deterministically.
+        addDriver("crash-test", 0, true);
+    }
+
     if (matched == 0) {
         _ = runtime.system.write("device-manager: no matchable devices\n");
+    } else {
+        _ = runtime.system.write("device-manager: ok\n");
+    }
+    return true;
+}
+
+fn onMessage(message: []const u8, reply: []u8, sender: u32, capability: ?runtime.ipc.Handle) usize {
+    if (message.len < 1) return 0;
+    switch (message[0]) {
+        @intFromEnum(protocol.Operation.child_added) => return onChildAdded(message, reply, sender),
+        @intFromEnum(protocol.Operation.child_removed) => return onChildRemoved(message, reply, sender),
+        @intFromEnum(protocol.Operation.enumerate) => return onEnumerate(reply),
+        @intFromEnum(protocol.Operation.subscribe) => return onSubscribe(reply, capability),
+        @intFromEnum(protocol.Operation.hello) => {},
+        else => return 0,
+    }
+    if (message.len < protocol.hello_size) return 0;
+    const hello = std.mem.bytesToValue(protocol.Hello, message[0..protocol.hello_size]);
+
+    var status: i32 = 0;
+    if (hello.version != protocol.version) {
+        status = -1;
+        writeLine("device-manager: refused hello (version {d}) from process {d}\n", .{ hello.version, sender });
+    } else if (driverByProcess(sender)) |driver| {
+        driver.state = .running;
+        writeLine("device-manager: hello from {s} (device {d})\n", .{ driver.name(), hello.device_id });
+    } else {
+        status = -1;
+        writeLine("device-manager: hello from unknown process {d}\n", .{sender});
+    }
+    const hello_reply = protocol.HelloReply{ .status = status };
+    @memcpy(reply[0..protocol.reply_size], std.mem.asBytes(&hello_reply));
+    return protocol.reply_size;
+}
+
+/// A bus driver reported a discovered device: mirror it, and in
+/// test-usb-restart mode kill the reporter once after its second child — the
+/// deterministic trigger for prune -> backoff -> respawn -> re-report.
+fn onChildAdded(message: []const u8, reply: []u8, sender: u32) usize {
+    if (message.len < protocol.child_added_size) return 0;
+    const report = std.mem.bytesToValue(protocol.ChildAdded, message[0..protocol.child_added_size]);
+    var status: i32 = 0;
+    if (driverByProcess(sender)) |driver| {
+        if (!addChild(report.parent, report.bus_address, report.identity, report.device_id, sender)) status = -1;
+        writeLine("device-manager: child added (device {d} port {d}, identity {d}) by {s}\n", .{ report.parent, report.bus_address, report.identity, driver.name() });
+        if (status == 0) publishEvent(message[0..protocol.child_added_size]);
+        // Matching from reports (M19.3): a registered child whose identity
+        // names a driver gets one, once — re-reports after a bus restart
+        // dedupe on the registered id, exactly like the registrations do.
+        if (status == 0 and report.device_id != protocol.no_device) {
+            if (pciDriverForIdentity(report.identity)) |child_driver| {
+                if (!driverForDevice(report.device_id)) addDriver(child_driver, report.device_id, true);
+            }
+        }
+    } else {
+        status = -1;
+    }
+    const report_reply = protocol.ReportReply{ .status = status };
+    @memcpy(reply[0..@sizeOf(protocol.ReportReply)], std.mem.asBytes(&report_reply));
+    if (test_pci_restart_mode and !test_usb_killed) {
+        if (driverByProcess(sender)) |driver| {
+            if (std.mem.eql(u8, driver.name(), "pci-bus") and childCountOf(sender) >= 3) {
+                // The pci restart drill: kill the enumerator after it has
+                // reported; the respawn must re-register without duplicates
+                // (M19.0 idempotence, proven end to end by pci-scan).
+                test_usb_killed = true;
+                test_kill_pid = sender;
+                test_kill_due_ns = system.clock() + 1_000_000_000;
+                _ = system.timerOnce(manager_endpoint, 1100);
+            }
+        }
+    }
+    if (test_usb_restart_mode and !test_usb_killed and childCountOf(sender) >= 2) {
+        // Only the xHCI reporter is the drill's victim — pci-bus also reports
+        // now, and whichever finishes second must not trigger the kill.
+        if (driverByProcess(sender)) |driver| {
+            if (std.mem.eql(u8, driver.name(), "usb-xhci-bus")) {
+                // Delayed, not immediate: the device-list scenario's subscriber
+                // needs a window to enumerate and subscribe before the events.
+                test_usb_killed = true;
+                test_kill_pid = sender;
+                test_kill_due_ns = system.clock() + 2_000_000_000;
+                _ = system.timerOnce(manager_endpoint, 2100);
+            }
+        }
+    }
+    return @sizeOf(protocol.ReportReply);
+}
+
+/// A bus driver reported a device gone (hot-unplug; no sender exists yet, but
+/// the handler is protocol-complete — death-pruning covers removal until then).
+fn onChildRemoved(message: []const u8, reply: []u8, sender: u32) usize {
+    if (message.len < protocol.child_removed_size) return 0;
+    const report = std.mem.bytesToValue(protocol.ChildRemoved, message[0..protocol.child_removed_size]);
+    var status: i32 = -1;
+    for (&children) |*child| {
+        if (child.used and child.parent == report.parent and child.bus_address == report.bus_address and child.reporter == sender) {
+            writeLine("device-manager: child removed (device {d} port {d})\n", .{ child.parent, child.bus_address });
+            child.used = false;
+            status = 0;
+        }
+    }
+    const report_reply = protocol.ReportReply{ .status = status };
+    @memcpy(reply[0..@sizeOf(protocol.ReportReply)], std.mem.asBytes(&report_reply));
+    return @sizeOf(protocol.ReportReply);
+}
+
+/// An application asked for the tree: the mirror, as a header plus entries.
+fn onEnumerate(reply: []u8) usize {
+    var count: u32 = 0;
+    var offset: usize = @sizeOf(protocol.EnumerateReply);
+    for (&children) |*child| {
+        if (!child.used) continue;
+        if (offset + @sizeOf(protocol.ChildEntry) > reply.len) break;
+        const entry = protocol.ChildEntry{ .parent = child.parent, .bus_address = child.bus_address, .identity = child.identity };
+        @memcpy(reply[offset..][0..@sizeOf(protocol.ChildEntry)], std.mem.asBytes(&entry));
+        offset += @sizeOf(protocol.ChildEntry);
+        count += 1;
+    }
+    const header = protocol.EnumerateReply{ .status = 0, .count = count };
+    @memcpy(reply[0..@sizeOf(protocol.EnumerateReply)], std.mem.asBytes(&header));
+    return offset;
+}
+
+/// An application subscribed: its endpoint arrived as the call's capability.
+fn onSubscribe(reply: []u8, capability: ?runtime.ipc.Handle) usize {
+    var status: i32 = -1;
+    if (capability) |handle| {
+        for (&subscribers) |*slot| {
+            if (slot.* == null) {
+                slot.* = handle;
+                status = 0;
+                break;
+            }
+        }
+    }
+    const report_reply = protocol.ReportReply{ .status = status };
+    @memcpy(reply[0..@sizeOf(protocol.ReportReply)], std.mem.asBytes(&report_reply));
+    return @sizeOf(protocol.ReportReply);
+}
+
+fn onNotification(badge: u64) void {
+    if (badge & runtime.ipc.notify_exit_bit != 0) {
+        const dead: u32 = @intCast(badge & ~(runtime.ipc.notify_badge_bit | runtime.ipc.notify_exit_bit));
+        if (driverByProcess(dead)) |driver| onDriverExit(driver);
         return;
     }
-    _ = runtime.system.write("device-manager: ok\n");
-    while (true) runtime.system.sleep(1000);
+    if (badge & runtime.ipc.notify_timer_bit != 0) sweepDeadlines();
+}
+
+pub fn main(init: runtime.process.Init) void {
+    if (init.arguments.get(1)) |mode| {
+        test_restart_mode = std.mem.eql(u8, mode, "test-restart");
+        test_usb_restart_mode = std.mem.eql(u8, mode, "test-usb-restart");
+        test_pci_restart_mode = std.mem.eql(u8, mode, "test-pci-restart");
+    }
+    runtime.service.run(protocol.message_maximum, .{
+        .service = .device_manager,
+        .init = initialise,
+        .on_message = onMessage,
+        .on_notification = onNotification,
+    });
 }
 
 pub const panic = runtime.panic;

system/services/fdt/fdt.zig

@@ -0,0 +1,34 @@
+//! /system/services/fdt — the devicetree discovery service: the ARM twin of the
+//! acpi service (docs/m19-m20-plan.md decision 7). **Placeholder: not
+//! implemented.** It exists so the build's `-Ddiscovery` option has both of its
+//! values from day one; the implementation lands with the Raspberry Pi
+//! bring-up (docs/arm.md).
+//!
+//! What it becomes: the per-firmware discoverer for boots that hand over a
+//! flattened device tree instead of ACPI tables. It claims the
+//! `devicetree-blob` node the kernel publishes (the FDT the loader received),
+//! walks the tree — pure data, no bytecode, so unlike the acpi service it
+//! needs no port grant and no interpreter — and, like any bus-shaped driver:
+//! `device_register`s what it finds (containment against the blob node's
+//! recorded apertures), reports each child to the device manager
+//! (`child_added`, identity = the node's `compatible` string), and stays
+//! resident under the manager's supervision (hello, restart, the usual
+//! contract).
+//!
+//! Known prerequisite recorded in the plan: `DeviceDescriptor`'s 8-byte `hid`
+//! cannot hold an FDT `compatible` string ("brcm,bcm2835-aux-uart") — identity
+//! widens before this file grows a body.
+
+const runtime = @import("runtime");
+
+pub fn main() void {
+    // Not implemented: exit cleanly and silently (a bare spawn by the
+    // initial-ramdisk sweep must not derange other tests' markers). The
+    // supervisor reads a clean exit as "meant to stop" — correct for a
+    // placeholder.
+}
+
+pub const panic = runtime.panic;
+comptime {
+    _ = &runtime.start._start; // pull the runtime entry shim into the image
+}

system/services/process-test/process-test.zig

@@ -48,11 +48,92 @@ fn awaitChildExit(endpoint: runtime.ipc.Handle) u32 {
     return received.childProcessId();
 }
 
+/// The harness-run child of the signals test: echoes requests, logs the two
+/// signals it handles. Terminate makes run() return, and returning from main is
+/// the clean exit the parent reads as ExitReason.exited.
+fn echo(message: []const u8, reply: []u8, sender: u32, capability: ?runtime.ipc.Handle) usize {
+    _ = sender;
+    _ = capability;
+    const n = @min(message.len, reply.len);
+    @memcpy(reply[0..n], message[0..n]);
+    return n;
+}
+
+fn onReload() void {
+    _ = runtime.system.write("process-test: reloaded\n");
+}
+
+fn onTerminate() void {
+    _ = runtime.system.write("process-test: terminating\n");
+}
+
+/// The parent of the signals test: drives ping, echo, reload, the one-shot
+/// timer, and both endings of the stop sequence (polite -> exited; deaf ->
+/// killed at the deadline). Prints "process-test: signals ok" as the marker.
+fn signalRun() void {
+    const endpoint = runtime.ipc.createIpcEndpoint() orelse fail("create exit endpoint");
+    const child = runtime.system.spawnSupervised("process-test", &.{"service"}, endpoint) orelse fail("spawn service child");
+
+    // Reach the child's endpoint through the registry (retry: it may not be up).
+    var service_handle: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (service_handle == null and tries < 200) : (tries += 1) {
+        service_handle = runtime.ipc.lookup(.input);
+        if (service_handle == null) runtime.system.sleep(20);
+    }
+    const h = service_handle orelse fail("service child never registered");
+
+    // The universal ping: a zero-length call answered zero-length by the harness.
+    var reply: [16]u8 = undefined;
+    const pong = runtime.ipc.call(h, &.{}, &reply) catch fail("ping call failed");
+    if (pong != 0) fail("ping reply not empty");
+
+    // An ordinary request still reaches on_message.
+    const n = runtime.ipc.call(h, "echo!", &reply) catch fail("echo call failed");
+    if (n != 5 or !std.mem.eql(u8, reply[0..5], "echo!")) fail("echo mismatch");
+
+    // reload: a statement — the child logs it; the kernel test reads the serial.
+    if (!runtime.process.sendSignal(child, .reload)) fail("send reload");
+    runtime.system.sleep(200);
+
+    // The one-shot timer: armed on our endpoint, lands as isTimer.
+    if (!runtime.system.timerOnce(endpoint, 100)) fail("arm timer");
+    var scratch: [8]u8 = undefined;
+    const landing = runtime.ipc.replyWait(endpoint, scratch[0..0], &scratch, null);
+    if (!landing.isTimer()) fail("expected the timer landing");
+
+    // The stop sequence, polite path: terminate, clean exit inside the deadline.
+    runtime.process.stop(child, 2000, endpoint);
+    if ((runtime.process.exitReason(child) orelse .killed) != .exited) fail("service child reason not exited");
+
+    // The deaf child: binds nothing, hears nothing — the deadline kills it.
+    const deaf = runtime.system.spawnSupervised("process-test", &.{"sleeper"}, endpoint) orelse fail("spawn deaf child");
+    runtime.system.sleep(50); // let it reach its sleep
+    runtime.process.stop(deaf, 300, endpoint);
+    if ((runtime.process.exitReason(deaf) orelse .exited) != .killed) fail("deaf child reason not killed");
+
+    _ = runtime.system.write("process-test: signals ok\n");
+}
+
 pub fn main(init: runtime.process.Init) void {
     const role = init.arguments.get(1) orelse return; // spawned bare (ramdisk sweep): stay silent
     if (std.mem.eql(u8, role, "sleeper")) {
         while (true) runtime.system.sleep(500);
     }
+    if (std.mem.eql(u8, role, "service")) {
+        // Borrowed well-known id: the input service is not part of this scenario.
+        runtime.service.run(64, .{
+            .service = .input,
+            .on_message = echo,
+            .on_reload = onReload,
+            .on_terminate = onTerminate,
+        });
+        return; // terminate arrived; returning is the clean exit
+    }
+    if (std.mem.eql(u8, role, "signal-run")) {
+        signalRun();
+        return;
+    }
     if (std.mem.eql(u8, role, "spinner")) {
         var beat: u64 = 0;
         const touch: *volatile u64 = &beat;
@@ -89,6 +170,12 @@ pub fn main(init: runtime.process.Init) void {
     if (listed(sleeper, "process-test")) fail("sleeper still listed after kill");
     if (listed(spinner, "process-test")) fail("spinner still listed after kill");
 
+    // M17.2: both children were killed by us, and the reason says so — the whole
+    // restart-policy input, read through the runtime like a real supervisor would.
+    if ((runtime.process.exitReason(sleeper) orelse .exited) != .killed) fail("sleeper reason not killed");
+    if ((runtime.process.exitReason(spinner) orelse .exited) != .killed) fail("spinner reason not killed");
+    if (runtime.process.exitReason(0xFFFF_FFF0) != null) fail("unknown id had a reason");
+
     _ = runtime.system.write("process-test: ok\n");
 }
 

system/services/vfs/vfs-test.zig

@@ -6,10 +6,30 @@
 const std = @import("std");
 const runtime = @import("runtime");
 
-pub fn main() void {
+pub fn main(init: runtime.process.Init) void {
     const u = @import("posix").unistd;
     const payload = "hello-vfs";
 
+    // The "park" role (the vfs-client-death test): open a file, then hold the
+    // handle forever without closing — the kill and the VFS's release-on-death
+    // are the point.
+    if (init.arguments.count > 1) {
+        var fd: i32 = -1;
+        var tries: u32 = 0;
+        while (fd < 0 and tries < 200) : (tries += 1) {
+            fd = u.open("parked", u.O_CREAT);
+            if (fd < 0) runtime.system.sleep(20);
+        }
+        if (fd < 0) {
+            _ = runtime.system.write("vfstest: park open failed\n");
+            return;
+        }
+        while (true) {
+            _ = runtime.system.write("vfstest: parked\n");
+            runtime.system.sleep(500);
+        }
+    }
+
     // The VFS server may not have registered yet — retry open until it's up.
     var fd: i32 = -1;
     var tries: u32 = 0;

system/services/vfs/vfs.zig

@@ -23,6 +23,10 @@ const Node = struct {
 const OpenFile = struct {
     used: bool = false,
     node: usize = 0,
+    // The client (task id — an IPC badge is one) that opened this handle. What
+    // release-on-death sweeps by: a service must never depend on its clients
+    // cleaning up after themselves (docs/process-lifecycle.md).
+    owner: u32 = 0,
 };
 
 var nodes = [_]Node{.{}} ** 8;
@@ -65,8 +69,30 @@ fn fail(out: []u8) usize {
     return writeReply(out, .{ .status = -1 }, &.{});
 }
 
-/// Handle one request; write the reply into `out`, return its length.
-fn handle(message: []const u8, out: []u8) usize {
+/// Format one whole log line and emit it in a single `debug_write`, so lines from
+/// concurrent processes can never land in the middle of it.
+fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
+    var line: [96]u8 = undefined;
+    _ = runtime.system.write(std.fmt.bufPrint(&line, fmt, arguments) catch return);
+}
+
+/// Release every open handle `client` held — called on that client's published
+/// exit event. The nodes (the files) stay: ramfs contents outlive their writers,
+/// only the dead client's handles go.
+fn releaseClientHandles(client: u32) void {
+    var released: u32 = 0;
+    for (&opens) |*o| {
+        if (o.used and o.owner == client) {
+            o.used = false;
+            released += 1;
+        }
+    }
+    if (released != 0) writeLine("vfs: released {d} handle(s) for dead client {d}\n", .{ released, client });
+}
+
+/// Handle one request from `sender`; write the reply into `out`, return its length.
+fn handle(message: []const u8, out: []u8, sender: u32, capability: ?runtime.ipc.Handle) usize {
+    _ = capability;
     if (message.len < protocol.request_size) return fail(out);
     const request = std.mem.bytesToValue(protocol.Request, message[0..protocol.request_size]);
     const payload = message[protocol.request_size..];
@@ -77,7 +103,7 @@ fn handle(message: []const u8, out: []u8) usize {
             const ni = findNode(name) orelse createNode(name) orelse return fail(out);
             for (&opens, 0..) |*o, i| {
                 if (!o.used) {
-                    o.* = .{ .used = true, .node = ni };
+                    o.* = .{ .used = true, .node = ni, .owner = sender };
                     return writeReply(out, .{ .status = 0, .node = i }, &.{});
                 }
             }
@@ -113,27 +139,36 @@ fn handle(message: []const u8, out: []u8) usize {
     }
 }
 
-pub fn main() void {
-    const endpoint = runtime.ipc.createIpcEndpoint() orelse {
-        _ = runtime.system.write("vfs: no endpoint\n");
-        return;
-    };
-    if (!runtime.ipc.register(.vfs, endpoint)) {
-        _ = runtime.system.write("vfs: register failed\n");
-        return;
+/// Startup, under the harness: subscribe to the published exit events — when a
+/// client dies holding open handles, the exit notification is how the VFS learns
+/// to release them (docs/process-lifecycle.md).
+fn initialise(endpoint: runtime.ipc.Handle) bool {
+    if (!runtime.process.subscribeExits(endpoint)) {
+        _ = runtime.system.write("vfs: exit subscription failed\n");
     }
     _ = runtime.system.write("vfs: ready\n");
+    return true;
+}
 
-    var reply_buffer: [protocol.message_maximum]u8 = undefined;
-    var reply_len: usize = 0;
-    var receive: [protocol.message_maximum]u8 = undefined;
-    while (true) {
-        const got = runtime.ipc.replyWait(endpoint, reply_buffer[0..reply_len], &receive, null);
-        // Ignore notifications (none expected here); handle a request.
-        reply_len = handle(receive[0..got.len], &reply_buffer);
+/// A non-signal notification: the only kind the VFS subscribes to is exit events.
+fn onNotification(badge: u64) void {
+    if (badge & runtime.ipc.notify_exit_bit != 0) {
+        releaseClientHandles(@intCast(badge & ~(runtime.ipc.notify_badge_bit | runtime.ipc.notify_exit_bit)));
     }
 }
 
+pub fn main() void {
+    // The harness owns the loop: requests dispatch to handle(), exit events to
+    // onNotification(), ping and terminate are answered for free — this service
+    // gained the whole lifecycle contract by deleting its hand-rolled loop.
+    runtime.service.run(protocol.message_maximum, .{
+        .service = .vfs,
+        .init = initialise,
+        .on_message = handle,
+        .on_notification = onNotification,
+    });
+}
+
 pub const panic = runtime.panic;
 comptime {
     _ = &runtime.start._start;

test/qemu_test.py

@@ -168,7 +168,7 @@ CASES = [
     # Stress the big kernel lock across cores; heavier, so a longer timeout.
     {"name": "smp-stress",
      "smp":    4,
-     "timeout": 90,
+     "timeout": 150,
      "expect": r"DANOS-TEST-RESULT: PASS",
      "fail":   r"DANOS-TEST-RESULT: FAIL"},
     # Retry: a forced first-wake failure must still bring every core online.
@@ -240,6 +240,75 @@ CASES = [
      "smp":    4,
      "expect": r"DANOS-TEST-RESULT: PASS",
      "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M17.1: a dead process's device claims are released by the reap — kill a child
+    # holding a claim, the device must be claimable again (process-lifecycle.md).
+    {"name": "claim-release",
+     "smp":    4,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M17.3: published exit events — the VFS subscribes, a client dies holding an
+    # open handle, and the VFS releases it (process-lifecycle.md "Who learns of a death").
+    {"name": "vfs-client-death",
+     "smp":    4,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M17.4: signals over IPC — ping, reload, terminate (clean exit), the one-shot
+    # timer, and the stop sequence's two endings, all driven from ring 3.
+    {"name": "signals",
+     "smp":    4,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M18.2: bus tree reports — the xHCI driver scans its root-hub ports and
+    # reports both QEMU devices; the manager mirrors, prunes on the reporter's
+    # death, and the respawned driver re-reports (docs/device-manager.md).
+    {"name": "usb-report",
+     "smp":    4,
+     "timeout": 150,
+     "qemu_extra": ["-device", "qemu-xhci,id=xhci",
+                    "-device", "usb-kbd,bus=xhci.0",
+                    "-device", "usb-mouse,bus=xhci.0"],
+     "expect": r"device-manager: child added[\s\S]*"
+               r"device-manager: child added[\s\S]*"
+               r"device-manager: test mode: killing the reporter[\s\S]*"
+               r"device-manager: child removed[\s\S]*"
+               r"device-manager: restarting usb-xhci-bus[\s\S]*"
+               r"device-manager: child added",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M19.1: the ring-3 PCI scan (pci-bus walks the ECAM through its mmio_map
+    # grant) finds exactly the functions the kernel's own walk recorded.
+    {"name": "pci-scan",
+     "smp":    4,
+     "timeout": 60,
+     "expect": r"DANOS-TEST-RESULT: PASS",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M18.3: the application surface — device-list enumerates the tree over IPC,
+    # subscribes (endpoint as capability), and observes the removed/added events
+    # the reporter's test-kill produces (docs/device-manager.md).
+    {"name": "device-list",
+     "smp":    4,
+     "timeout": 150,
+     "qemu_extra": ["-device", "qemu-xhci,id=xhci",
+                    "-device", "usb-kbd,bus=xhci.0",
+                    "-device", "usb-mouse,bus=xhci.0"],
+     "expect": r"device-list: \d+ devices[\s\S]*"
+               r"device-list: subscribed[\s\S]*"
+               r"device-manager: test mode: killing the reporter[\s\S]*"
+               r"device-list: removed \(device[\s\S]*"
+               r"device-list: added \(device",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
+    # M18.1: the device manager's hello + restart policy — xHCI hellos clean and
+    # stays; crash-test faults, is restarted with backoff (re-claiming its device
+    # each time), and hits the crash-loop cap (docs/device-manager.md).
+    {"name": "driver-restart",
+     "smp":    4,
+     "timeout": 150,
+     "qemu_extra": ["-device", "qemu-xhci,id=xhci",
+                    "-device", "usb-kbd,bus=xhci.0",
+                    "-device", "usb-mouse,bus=xhci.0"],
+     "expect": r"usb-xhci-bus: hello acknowledged[\s\S]*"
+               r"device-manager: restarting crash-test[\s\S]*"
+               r"device-manager: crash-test is failing repeatedly",
+     "fail":   r"DANOS-TEST-RESULT: FAIL"},
     # The initial_ramdisk: the loader ferries a bundle of user binaries; the kernel parses
     # it and spawns each as a ring-3 process (here the VFS-server stub heartbeats).
     {"name": "initial-ramdisk",
@@ -410,6 +479,12 @@ def main():
     for case in selected:
         print(f"  {case['name']:<12} ... ", end="", flush=True)
         ok, detail = run_case(arch, case)
+        if not ok:
+            # Keep the evidence: serial.log is otherwise overwritten by the
+            # next case, and an intermittent failure's log is unrecoverable.
+            source = os.path.join(WORK, "serial.log")
+            if os.path.exists(source):
+                shutil.copy(source, os.path.join(WORK, f"{case['name']}-failed-serial.log"))
         print(("PASS" if ok else "FAIL") + f"  ({detail})")
         if not ok:
             failures += 1