device-manager: data-driven driver matching via /etc/devices.csv

Replace the three hardcoded switch tables (pciDriverForIdentity, hidDriverFor,
usbDriverForIdentity) with an authoritative, human-readable device registry the
manager reads at boot. Matching is most-specific-wins across
base/subclass/prog_if/vendor/device/subsystem/hid, so a precise vendor:device
rule and a generic class rule coexist; an unmatched device is logged, never
guessed. This resolves docs' "matching stays code until the third bus".

- ABI: child_added and DeviceDescriptor gain vendor/device/subsystem; child_added
  gains a bus discriminator (BusKind) so PCI and USB class triples match against
  the right namespace.
- pci-bus reads vendor/device (config 0x00) and subsystem (0x2C, type-0) and
  reports them.
- library/device/registry: freestanding CSV parser + matchDriver() with
  specificity scoring; 5 unit tests wired into `zig build test`.
- etc/devices.csv bundled into the initrd; the kernel serves /etc directly, so
  the manager reads it before any filesystem service is up (fat starts later).
- virtio-gpu: drop the now-redundant post-spawn 1AF4:1050 re-confirm, since the
  registry binds this driver by exact identity.
- Remove the orphaned system/drivers/display driver (unreferenced by build or
  registry).
- docs: new devices-csv.md; device-manager.md "matching stays code" resolved.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KJqSiLLchDUUCoXn5jsiwd

library/protocol/device-manager/device-manager-protocol.zig

@@ -11,6 +11,19 @@
 /// startup instead of quiet corruption later.
 pub const version: u16 = 1;
 
+/// Which bus a `child_added` came from — stated by the reporting bus driver so
+/// the manager's /etc/devices.csv matcher knows how to read the report's identity
+/// (a PCI class triple vs a USB class triple are the same 24 bits but different
+/// namespaces) and which `bus` column a rule must name to bind it. `unknown` is
+/// the zero default, so an un-upgraded reporter fails to match rather than
+/// binding to the wrong bus's rule.
+pub const BusKind = enum(u8) {
+    unknown = 0,
+    pci = 1,
+    usb = 2,
+    acpi = 3,
+};
+
 /// 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`).
@@ -66,7 +79,9 @@ pub const reply_size = @sizeOf(HelloReply);
 /// restarted instance rediscovers and re-reports.
 pub const ChildAdded = extern struct {
     operation: u8 = @intFromEnum(Operation.child_added),
-    reserved0: u8 = 0,
+    /// A `BusKind` value: which bus reported this child, so the manager reads the
+    /// identity in the right namespace and matches against the right `bus` column.
+    bus: u8 = @intFromEnum(BusKind.unknown),
     reserved1: u16 = 0,
     reserved2: u32 = 0,
     /// The reporting driver's own device (the controller) — the child's parent.
@@ -80,6 +95,18 @@ pub const ChildAdded = extern struct {
     /// 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,
+    /// The vendor id (PCI vendor / USB idVendor), or 0 when the bus has no such
+    /// concept (ACPI). Carried so the manager's /etc/devices.csv matcher can bind
+    /// on vendor — a level the bus-native `identity` (a class triple) cannot express.
+    vendor: u16 = 0,
+    /// The device id (PCI device / USB idProduct), or 0. The most specific numeric
+    /// level: this is what lets one virtio-gpu (1AF4:1050) be told from any other
+    /// virtio display function without the driver re-confirming after it is spawned.
+    device: u16 = 0,
+    /// The PCI subsystem id, packed `(subsystem_vendor << 16) | subsystem_device`
+    /// (so it reads vendor-first, matching the CSV's `ssvid:ssid`), or 0 when the
+    /// device has no subsystem id (a bridge, or a non-PCI bus).
+    subsystem: u32 = 0,
     /// The ACPI hardware id (`_HID`), EISA-decoded (e.g. "PNP0303"), for devices
     /// discovered by firmware string rather than a numeric bus identity. Empty
     /// (all zero) otherwise. Widens for FDT `compatible` strings later.