fix: x2APIC ICR format + build system durability docs

- Fix LocalX2Apic handler: use local_apic.x2 to select correct ICR
  format (<<32 for x2APIC, <<56 for xAPIC) instead of hardcoded <<32
- Promote x2APIC/xAPIC detection from debug! to info! for bootlog
- Document build system durability in AGENTS.md: cardinal rule,
  two-layer architecture, correct workflow, anti-patterns

AGENTS.md

@@ -19,7 +19,109 @@ human-initiated operations. Durable Red Bear state belongs in `local/patches/`,
 The current baseline is **Red Bear OS 0.1.0** (Redox snapshot at build-system commit `f55acba68`).
 All recipe sources are pinned and archived in `sources/redbear-0.1.0/`.
 
-## NO SILENT UPSTREAM PULLS — OFFLINE-FIRST POLICY
+## BUILD SYSTEM DURABILITY — THE CARDINAL RULE
+
+**Never edit files under `recipes/*/source/` directly. Those trees are ephemeral — they are
+destroyed and regenerated on every `repo fetch`, `repo cook`, `make clean`, and
+`make distclean`. Any edit made there will be silently lost on the next build.**
+
+This is the #1 mistake AI agents and new contributors make. It has caused repeated work loss
+in this project. The rule is:
+
+| What you want to do | Where to do it |
+|---|---|
+| Change a kernel source file | Create or update a patch in `local/patches/kernel/` |
+| Change an init or daemon source file | Create or update a patch in `local/patches/base/` |
+| Change relibc | Create or update a patch in `local/patches/relibc/` |
+| Change a driver | Create or update a patch in `local/patches/base/` or `local/patches/<driver>/` |
+| Add a new package | Create a recipe in `local/recipes/<category>/<name>/` |
+| Change build config | Edit `config/redbear-*.toml` |
+| Add documentation | Write to `local/docs/` |
+
+### How the build system works
+
+```
+repo cook <package>
+  ├── repo fetch <package>
+  │   ├── Clone/fetch upstream source → recipes/<pkg>/source/
+  │   ├── Apply patches from recipe.toml → patches are read from local/patches/<pkg>/
+  │   └── Source tree is now fully patched and ready for build
+  ├── Cargo/cmake/configure build
+  └── Stage artifacts into sysroot
+```
+
+The `source/` directory is a disposable working copy. It is produced at the start of every
+build by cloning the upstream source + applying patches sequentially. The recipe's
+`patches = [...]` list in `recipe.toml` controls which patches are applied.
+
+### Two-layer architecture
+
+```
+Layer 1: Ephemeral (destroyed on clean/fetch/rebuild)
+  recipes/<pkg>/source/       ← working tree, cloned + patched
+  build/                      ← build outputs
+  target/                     ← cargo target dir
+
+Layer 2: Durable (survives clean/fetch/rebuild/release provisioning)
+  local/patches/<pkg>/        ← .patch files — the actual source code changes
+  local/recipes/<pkg>/        ← custom recipe directories
+  config/redbear-*.toml       ← Red Bear OS build configs
+  local/docs/                 ← planning and integration docs
+  recipes/<pkg>/recipe.toml   ← the patches list (git-tracked)
+```
+
+### The correct workflow for any source change
+
+1. **Make the change** in `recipes/<pkg>/source/` to validate it compiles
+2. **Generate a patch**: `cd recipes/<pkg>/source && git diff > ../../../local/patches/<pkg>/my-fix.patch`
+3. **Wire the patch**: add `"my-fix.patch"` to the recipe's `recipe.toml` `patches = [...]` list
+4. **Validate**: `./target/release/repo validate-patches <pkg>`
+5. **Rebuild**: `./target/release/repo cook <pkg>`
+6. **Commit**: `git add local/patches/ recipes/<pkg>/recipe.toml && git commit`
+
+### Common anti-patterns
+
+| Anti-pattern | Why it fails |
+|---|---|
+| Editing `source/` files then running `make all` | `make all` calls `repo fetch` which regenerates `source/` — edits are lost |
+| Creating a patch but not wiring it into `recipe.toml` | Patch file exists but is never applied — build uses unpatched source |
+| Editing `recipe.toml` patches list without creating the actual `.patch` file | Build fails with "missing patch" error |
+| Expecting `source/` changes to survive `make clean` | `make clean` deletes `source/` directories |
+| Running `repo cook` without `--allow-protected` for core packages | Protected recipes (kernel, relibc, base) are offline-only by default |
+
+### Patch file location convention
+
+- `local/patches/base/` — for the `base` package (init, daemon, all drivers)
+- `local/patches/kernel/` — for the kernel
+- `local/patches/relibc/` — for relibc
+- `local/patches/installer/` — for the installer
+- `local/patches/bootloader/` — for the bootloader
+- `local/patches/<package>/` — for any other patched package
+
+### Recipe patch wiring
+
+Each recipe's `recipe.toml` lists patches relative to `local/patches/<pkg>/`:
+
+```toml
+[source]
+git = "https://gitlab.redox-os.org/redox-os/base.git"
+rev = "463f76b96..."
+patches = [
+    "P0-daemon-fix-init-notify-unwrap.patch",   # applied first
+    "P9-init-scheduler-completed.patch",          # applied second
+    # ... more patches
+]
+```
+
+Patches are applied in listed order. Dependencies between patches must be respected (a patch
+that defines a type must come before a patch that uses it).
+
+### Kernel-specific notes
+
+The kernel source at `recipes/core/kernel/source/` is a separate git worktree (rev `866dfad`).
+The kernel recipe is at `recipes/core/kernel/recipe.toml` and patches are at
+`local/patches/kernel/`. The same durability rules apply — all kernel changes must be
+in `local/patches/kernel/*.patch`, never in the `source/` tree directly.
 
 **Red Bear OS is offline-first by default. No script, build target, or tool may silently pull
 from any upstream repository without explicit user instruction.**

local/patches/kernel/P20-x2apic-icr-mode-fix.patch

@@ -0,0 +1,64 @@
+diff --git a/src/acpi/madt/arch/x86.rs b/src/acpi/madt/arch/x86.rs
+--- a/src/acpi/madt/arch/x86.rs
++++ b/src/acpi/madt/arch/x86.rs
+@@ -444,28 +444,40 @@
+                 // Send INIT IPI (Assert)
+                 {
+                     let mut icr = 0x4500u64;
+-                    icr |= u64::from(apic_id) << 32;
++                    if local_apic.x2 {
++                        icr |= u64::from(apic_id) << 32;
++                    } else {
++                        icr |= u64::from(apic_id as u8) << 56;
++                    }
+                     local_apic.set_icr(icr);
+                 }
+-
++ 
+                 // Intel SDM Vol 3A §8.4.4: wait 10ms after INIT
+                 early_udelay(10_000);
+-
++ 
+                 // Send START IPI #1
+                 {
+                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
+                     let mut icr = 0x0600u64 | ap_segment as u64;
+-                    icr |= u64::from(apic_id) << 32;
++                    if local_apic.x2 {
++                        icr |= u64::from(apic_id) << 32;
++                    } else {
++                        icr |= u64::from(apic_id as u8) << 56;
++                    }
+                     local_apic.set_icr(icr);
+                 }
+-
++ 
+                 // Intel SDM: wait 200µs between SIPIs
+                 early_udelay(200);
+-
++ 
+                 // Send START IPI #2 (recommended for compatibility)
+                 {
+                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
+                     let mut icr = 0x0600u64 | ap_segment as u64;
+-                    icr |= u64::from(apic_id) << 32;
++                    if local_apic.x2 {
++                        icr |= u64::from(apic_id) << 32;
++                    } else {
++                        icr |= u64::from(apic_id as u8) << 56;
++                    }
+                     local_apic.set_icr(icr);
+                 }
+diff --git a/src/arch/x86_shared/device/local_apic.rs b/src/arch/x86_shared/device/local_apic.rs
+--- a/src/arch/x86_shared/device/local_apic.rs
++++ b/src/arch/x86_shared/device/local_apic.rs
+@@ -60,7 +60,7 @@
+ 
+             if !self.x2 {
+-                debug!("Detected xAPIC at {:#x}", physaddr.data());
++                info!("Detected xAPIC at {:#x}", physaddr.data());
+                 self.address = map_device_memory(physaddr, 4096).data();
+             } else {
+-                debug!("Detected x2APIC");
++                info!("Detected x2APIC");
+             }

recipes/core/kernel/recipe.toml

@@ -34,6 +34,7 @@ patches = [
     "../../../local/patches/kernel/P17-3-sched-affinity.patch",
     "../../../local/patches/kernel/P17-3-syscall-dispatch.patch",
     "../../../local/patches/kernel/P19-2-irq-debug.patch",
+    "../../../local/patches/kernel/P20-x2apic-icr-mode-fix.patch",
 ]
 
 [build]

recipes/core/kernel/source/src/acpi/madt/arch/x86.rs

@@ -383,6 +383,8 @@ pub(super) fn init(madt: Madt) {
                 }
 
                 RmmA::invalidate_all();
+            } else {
+                debug!("KERNEL AP: LAPIC CPU {} disabled in MADT, skipping", u32::from(ap_local_apic.id));
             }
         } else if let MadtEntry::LocalX2Apic(ap_x2apic) = madt_entry {
             let apic_id = ap_x2apic.x2apic_id;
@@ -444,7 +446,11 @@ pub(super) fn init(madt: Madt) {
                 // Send INIT IPI (Assert)
                 {
                     let mut icr = 0x4500u64;
+                    if local_apic.x2 {
                         icr |= u64::from(apic_id) << 32;
+                    } else {
+                        icr |= u64::from(apic_id as u8) << 56;
+                    }
                     local_apic.set_icr(icr);
                 }
 
@@ -455,7 +461,11 @@ pub(super) fn init(madt: Madt) {
                 {
                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
                     let mut icr = 0x0600u64 | ap_segment as u64;
+                    if local_apic.x2 {
                         icr |= u64::from(apic_id) << 32;
+                    } else {
+                        icr |= u64::from(apic_id as u8) << 56;
+                    }
                     local_apic.set_icr(icr);
                 }
 
@@ -466,7 +476,11 @@ pub(super) fn init(madt: Madt) {
                 {
                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
                     let mut icr = 0x0600u64 | ap_segment as u64;
+                    if local_apic.x2 {
                         icr |= u64::from(apic_id) << 32;
+                    } else {
+                        icr |= u64::from(apic_id as u8) << 56;
+                    }
                     local_apic.set_icr(icr);
                 }
 
@@ -520,6 +534,8 @@ pub(super) fn init(madt: Madt) {
                 }
 
                 RmmA::invalidate_all();
+            } else {
+                debug!("KERNEL AP: x2APIC CPU {} disabled in MADT (flags={:#x}), skipping", apic_id, flags);
             }
         } else if let MadtEntry::LocalApicNmi(nmi) = madt_entry {
             let target_apic = nmi.processor;

recipes/core/kernel/source/src/acpi/madt/arch/x86.rs.orig

@@ -0,0 +1,162 @@
+use core::{
+    hint,
+    sync::atomic::{AtomicU8, Ordering},
+};
+
+use crate::{
+    arch::{
+        device::local_apic::the_local_apic,
+        start::{kstart_ap, KernelArgsAp},
+    },
+    cpu_set::LogicalCpuId,
+    memory::{
+        allocate_p2frame, Frame, KernelMapper, Page, PageFlags, PhysicalAddress, RmmA, RmmArch,
+        VirtualAddress, PAGE_SIZE,
+    },
+    startup::AP_READY,
+};
+
+use super::{Madt, MadtEntry};
+
+const TRAMPOLINE: usize = 0x8000;
+static TRAMPOLINE_DATA: &[u8] = include_bytes!(concat!(env!("OUT_DIR"), "/trampoline"));
+
+pub(super) fn init(madt: Madt) {
+    let local_apic = unsafe { the_local_apic() };
+    let me = local_apic.id();
+
+    if local_apic.x2 {
+        debug!("    X2APIC {}", me.get());
+    } else {
+        debug!("    XAPIC {}: {:>08X}", me.get(), local_apic.address);
+    }
+
+    if cfg!(not(feature = "multi_core")) {
+        return;
+    }
+
+    // Map trampoline
+    let trampoline_frame = Frame::containing(PhysicalAddress::new(TRAMPOLINE));
+    let trampoline_page = Page::containing_address(VirtualAddress::new(TRAMPOLINE));
+    let (result, page_table_physaddr) = unsafe {
+        //TODO: do not have writable and executable!
+        let mut mapper = KernelMapper::lock_rw();
+
+        let result = mapper
+            .map_phys(
+                trampoline_page.start_address(),
+                trampoline_frame.base(),
+                PageFlags::new().execute(true).write(true),
+            )
+            .expect("failed to map trampoline");
+
+        (result, mapper.table().phys().data())
+    };
+    result.flush();
+
+    // Write trampoline, make sure TRAMPOLINE page is free for use
+    for (i, val) in TRAMPOLINE_DATA.iter().enumerate() {
+        unsafe {
+            (*((TRAMPOLINE as *mut u8).add(i) as *const AtomicU8)).store(*val, Ordering::SeqCst);
+        }
+    }
+
+    unsafe {
+        let preliminary_cpu_count = madt.iter().filter(|e| matches!(e, MadtEntry::LocalApic(entry) if u32::from(entry.id) == me.get() || entry.flags & 1 == 1)).count();
+        crate::profiling::allocate(preliminary_cpu_count as u32);
+    }
+
+    for madt_entry in madt.iter() {
+        debug!("      {:x?}", madt_entry);
+        if let MadtEntry::LocalApic(ap_local_apic) = madt_entry {
+            if u32::from(ap_local_apic.id) == me.get() {
+                debug!("        This is my local APIC");
+            } else if ap_local_apic.flags & 1 == 1 {
+                let cpu_id = LogicalCpuId::next();
+
+                // Allocate a stack
+                let stack_start = RmmA::phys_to_virt(
+                    allocate_p2frame(4)
+                        .expect("no more frames in acpi stack_start")
+                        .base(),
+                )
+                .data();
+                let stack_end = stack_start + (PAGE_SIZE << 4);
+
+                let pcr_ptr = crate::arch::gdt::allocate_and_init_pcr(cpu_id, stack_end);
+
+                let idt_ptr = crate::arch::idt::allocate_and_init_idt(cpu_id);
+
+                let args = KernelArgsAp {
+                    stack_end: stack_end as *mut u8,
+                    cpu_id,
+                    pcr_ptr,
+                    idt_ptr,
+                };
+
+                let ap_ready = (TRAMPOLINE + 8) as *mut u64;
+                let ap_args_ptr = unsafe { ap_ready.add(1) };
+                let ap_page_table = unsafe { ap_ready.add(2) };
+                let ap_code = unsafe { ap_ready.add(3) };
+
+                // Set the ap_ready to 0, volatile
+                unsafe {
+                    ap_ready.write(0);
+                    ap_args_ptr.write(&args as *const _ as u64);
+                    ap_page_table.write(page_table_physaddr as u64);
+                    #[expect(clippy::fn_to_numeric_cast)]
+                    ap_code.write(kstart_ap as u64);
+
+                    // Ensure all trampoline writes are visible to the AP before
+                    // it starts executing.  asm!("") is only a compiler barrier;
+                    // fence(SeqCst) is a full hardware memory barrier.
+                    core::sync::atomic::fence(Ordering::SeqCst);
+                };
+                AP_READY.store(false, Ordering::SeqCst);
+
+                // Send INIT IPI
+                {
+                    let mut icr = 0x4500;
+                    if local_apic.x2 {
+                        icr |= u64::from(ap_local_apic.id) << 32;
+                    } else {
+                        icr |= u64::from(ap_local_apic.id) << 56;
+                    }
+                    local_apic.set_icr(icr);
+                }
+
+                // Send START IPI
+                {
+                    let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
+                    let mut icr = 0x4600 | ap_segment as u64;
+
+                    if local_apic.x2 {
+                        icr |= u64::from(ap_local_apic.id) << 32;
+                    } else {
+                        icr |= u64::from(ap_local_apic.id) << 56;
+                    }
+
+                    local_apic.set_icr(icr);
+                }
+
+                // Wait for trampoline ready
+                while unsafe { (*ap_ready.cast::<AtomicU8>()).load(Ordering::SeqCst) } == 0 {
+                    hint::spin_loop();
+                }
+                while !AP_READY.load(Ordering::SeqCst) {
+                    hint::spin_loop();
+                }
+
+                RmmA::invalidate_all();
+            }
+        }
+    }
+
+    // Unmap trampoline
+    let (_frame, _, flush) = unsafe {
+        KernelMapper::lock_rw()
+            .unmap_phys(trampoline_page.start_address())
+            .expect("failed to unmap trampoline page")
+    };
+    flush.flush();
+}

recipes/core/kernel/source/src/acpi/madt/arch/x86.rs.rej

@@ -0,0 +1,194 @@
+--- src/acpi/madt/arch/x86.rs
++++ src/acpi/madt/arch/x86.rs
+@@ -20,6 +22,7 @@
+ 
+ use super::{Madt, MadtEntry};
+ 
++use alloc::collections::BTreeSet;
+ use alloc::vec::Vec;
+ 
+ /// Maximum number of APIC→CPU mappings we track for NUMA topology.
+@@ -47,6 +50,67 @@
+ const TRAMPOLINE: usize = 0x8000;
+ static TRAMPOLINE_DATA: &[u8] = include_bytes!(concat!(env!("OUT_DIR"), "/trampoline"));
+ 
++/// Estimate TSC frequency in MHz from CPUID.
++///
++/// Tries CPUID leaf 0x16 (Processor Frequency Information) first,
++/// then CPUID leaf 0x15 (TSC/Core Crystal Clock Ratio).
++/// Returns None if frequency cannot be determined.
++fn tsc_freq_mhz_cpuid() -> Option<u64> {
++    let max_leaf = unsafe { core::arch::x86_64::__cpuid(0).eax as u32 };
++
++    // CPUID leaf 0x16: EAX = Core Base Frequency in MHz (Intel)
++    if max_leaf >= 0x16 {
++        let mhz = unsafe { core::arch::x86_64::__cpuid(0x16) }.eax as u64;
++        if mhz > 0 {
++            return Some(mhz);
++        }
++    }
++
++    // CPUID leaf 0x15: EAX = denominator, EBX = numerator, ECX = crystal Hz
++    if max_leaf >= 0x15 {
++        let res = unsafe { core::arch::x86_64::__cpuid(0x15) };
++        let denom = res.eax as u64;
++        let numer = res.ebx as u64;
++        let crystal_hz = res.ecx as u64;
++        if denom > 0 && numer > 0 && crystal_hz > 0 {
++            // TSC freq = crystal_hz * numer / denom
++            let tsc_hz = crystal_hz * numer / denom;
++            return Some(tsc_hz / 1_000_000); // Hz → MHz
++        }
++    }
++
++    None
++}
++
++/// Early-boot microsecond delay using the Time Stamp Counter.
++///
++/// Uses CPUID-based TSC frequency estimation when available.
++/// Falls back to a conservative spin loop calibrated for the
++/// minimum expected CPU speed (1 GHz).
++///
++/// # Safety
++/// Must only be called after the BSP TSC is running (always true
++/// after CPU reset on x86).
++fn early_udelay(us: u64) {
++    if let Some(mhz) = tsc_freq_mhz_cpuid() {
++        // TSC-based delay: precise on invariant TSC (all modern x86).
++        // MHz = cycles per µs.
++        let target = unsafe { rdtsc() } + us * mhz;
++        while unsafe { rdtsc() } < target {
++            hint::spin_loop();
++        }
++    } else {
++        // Fallback: conservative spin loop.
++        // spin_loop() (PAUSE) is ~40 cycles on modern Intel, ~1 on AMD.
++        // At 1 GHz minimum: 1000 cycles/µs ÷ 40 cycles/iter = 25 iters/µs.
++        // Use 50 iters/µs for safety margin on slower/variable CPUs.
++        let iters = us.saturating_mul(50);
++        for _ in 0..iters {
++            hint::spin_loop();
++        }
++    }
++}
++
+ fn current_x2apic_processor_uid(madt: &Madt, apic_id: u32) -> Option<u32> {
+     madt.iter().find_map(|entry| match entry {
+         MadtEntry::LocalX2Apic(x2apic) if x2apic.x2apic_id == apic_id => Some(x2apic.processor_uid),
+@@ -235,20 +329,53 @@
+                     local_apic.set_icr(icr);
+                 }
+ 
+-                // Send START IPI
++                // Intel SDM Vol 3A §8.4.4: wait 10ms after INIT deassert
++                // before sending first SIPI. Modern CPUs may need less,
++                // but 10ms is the safe specification-compliant value.
++                early_udelay(10_000);
++
++                // Send START IPI #1
+                 {
+                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
+-                    let mut icr = 0x4600 | ap_segment as u64;
+-
++                    // ICR: Delivery Mode=StartUp(110), Vector=ap_segment
++                    // Note: bit 14 (Level) must be 0 for SIPI per Intel SDM.
++                    let mut icr = 0x0600 | ap_segment as u64;
+                     if local_apic.x2 {
+                         icr |= u64::from(ap_local_apic.id) << 32;
+                     } else {
+                         icr |= u64::from(ap_local_apic.id) << 56;
+                     }
++                    local_apic.set_icr(icr);
++                }
++
++                // Intel SDM: wait 200µs between SIPIs
++                early_udelay(200);
+ 
++                // Send START IPI #2 (recommended for compatibility)
++                {
++                    let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
++                    let mut icr = 0x0600 | ap_segment as u64;
++                    if local_apic.x2 {
++                        icr |= u64::from(ap_local_apic.id) << 32;
++                    } else {
++                        icr |= u64::from(ap_local_apic.id) << 56;
++                    }
+                     local_apic.set_icr(icr);
+                 }
+ 
++                // Wait briefly for SIPI to be accepted
++                early_udelay(200);
++
++                // Check ESR for delivery errors after SIPI sequence.
++                // Bit 5 = Send Accept Error, Bit 6 = Send Illegal Vector.
++                let esr_val = unsafe { local_apic.esr() };
++                if esr_val != 0 {
++                    println!(
++                        "KERNEL AP: CPU {} SIPI delivery error (ESR={:#x}), continuing",
++                        ap_local_apic.id, esr_val
++                    );
++                }
++
+                 // Wait for trampoline ready with timeout
+                 let mut trampoline_ready = false;
+                 for _ in 0..AP_SPIN_LIMIT {
+@@ -343,34 +470,50 @@
+                 }
+                 AP_READY.store(false, Ordering::SeqCst);
+ 
++                // Clear APIC Error Status Register before starting AP.
++                unsafe { local_apic.esr(); }
++
++                // Send INIT IPI (Assert)
+                 {
+                     let mut icr = 0x4500u64;
+                     icr |= u64::from(apic_id) << 32;
+                     local_apic.set_icr(icr);
+                 }
+ 
+-                for _ in 0..100_000 {
+-                    hint::spin_loop();
+-                }
++                // Intel SDM Vol 3A §8.4.4: wait 10ms after INIT
++                early_udelay(10_000);
+ 
++                // Send START IPI #1
+                 {
+                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
+-                    let mut icr = 0x4600u64 | ap_segment as u64;
++                    let mut icr = 0x0600u64 | ap_segment as u64;
+                     icr |= u64::from(apic_id) << 32;
+                     local_apic.set_icr(icr);
+                 }
+ 
+-                for _ in 0..2_000_000 {
+-                    hint::spin_loop();
+-                }
++                // Intel SDM: wait 200µs between SIPIs
++                early_udelay(200);
+ 
++                // Send START IPI #2 (recommended for compatibility)
+                 {
+                     let ap_segment = (TRAMPOLINE >> 12) & 0xFF;
+-                    let mut icr = 0x4600u64 | ap_segment as u64;
++                    let mut icr = 0x0600u64 | ap_segment as u64;
+                     icr |= u64::from(apic_id) << 32;
+                     local_apic.set_icr(icr);
+                 }
+ 
++                // Wait briefly for SIPI acceptance
++                early_udelay(200);
++
++                // Check ESR for delivery errors.
++                let esr_val = unsafe { local_apic.esr() };
++                if esr_val != 0 {
++                    println!(
++                        "KERNEL AP: CPU {} SIPI delivery error (ESR={:#x}), continuing",
++                        apic_id, esr_val
++                    );
++                }
++
+                 let mut trampoline_ready = false;
+                 for _ in 0..AP_SPIN_LIMIT {
+                     if unsafe { (*ap_ready.cast::<AtomicU8>()).load(Ordering::SeqCst) } != 0 {

recipes/core/kernel/source/src/arch/x86_shared/device/local_apic.rs

@@ -59,10 +59,10 @@ impl LocalApic {
                 .is_some_and(|feature_info| feature_info.has_x2apic());
 
             if !self.x2 {
-                debug!("Detected xAPIC at {:#x}", physaddr.data());
+                info!("Detected xAPIC at {:#x}", physaddr.data());
                 self.address = map_device_memory(physaddr, 4096).data();
             } else {
-                debug!("Detected x2APIC");
+                info!("Detected x2APIC");
             }
 
             self.init_ap();

recipes/core/kernel/source/src/arch/x86_shared/idt.rs.rej

@@ -0,0 +1,11 @@
+--- src/arch/x86_shared/idt.rs
++++ src/arch/x86_shared/idt.rs
+@@ -110,6 +110,8 @@
+ }
+ 
+ pub fn available_irqs_iter(cpu_id: LogicalCpuId) -> impl Iterator<Item = u8> + 'static {
++    let count = (32..=254).filter(|&index| !is_reserved(cpu_id, index)).count();
++    info!("available_irqs_iter: cpu_id={} count={}", cpu_id.get(), count);
+     (32..=254).filter(move |&index| !is_reserved(cpu_id, index))
+ }
+ 

recipes/core/kernel/source/src/context/switch.rs.rej

@@ -0,0 +1,87 @@
+--- src/context/switch.rs
++++ src/context/switch.rs
+@@ -361,6 +361,7 @@
+ }
+ 
+ /// This is the scheduler function which currently utilises Deficit Weighted Round Robin Scheduler
++/// with NUMA-aware context selection preference.
+ fn select_next_context(
+     token: &mut CleanLockToken,
+     percpu: &PercpuBlock,
+@@ -386,6 +387,10 @@
+     let total_contexts: usize = contexts_list.iter().map(|q| q.len()).sum();
+     let mut skipped_contexts = 0;
+ 
++    // NUMA-aware selection: remember cross-node fallback candidate.
++    let my_numa_node = percpu.numa_node.get();
++    let mut cross_node_fallback: Option<(usize, ArcContextLockWriteGuard)> = None;
++
+     'priority: loop {
+         i = (i + 1) % 40;
+         total_iters += 1;
+@@ -450,11 +455,44 @@
+             // Is this context runnable on this CPU?
+             let sw = unsafe { update_runnable(&mut next_context_guard, cpu_id, switch_time) };
+             if let UpdateResult::CanSwitch = sw {
+-                // Cache the new context's priority for MCS lock priority donation.
+-                percpu.current_prio.set(next_context_guard.prio);
+-                next_context_guard_opt = Some(next_context_guard);
+-                balance[i] -= SCHED_PRIO_TO_WEIGHT[20];
+-                break 'priority;
++                // NUMA-aware selection: check if this context's last CPU was on the same node.
++                let same_node = if my_numa_node != u8::MAX {
++                    next_context_guard.cpu_id
++                        .map(|cid| {
++                            crate::percpu::get_for_cpu(cid)
++                                .map(|p| p.numa_node.get() == my_numa_node)
++                                .unwrap_or(false)
++                        })
++                        .unwrap_or(true) // New context (no last CPU) — treat as same node
++                } else {
++                    true // No NUMA info — treat all as same node
++                };
++
++                if same_node {
++                    // Cache-warm: select immediately
++                    percpu.current_prio.set(next_context_guard.prio);
++                    next_context_guard_opt = Some(next_context_guard);
++                    balance[i] -= SCHED_PRIO_TO_WEIGHT[20];
++                    break 'priority;
++                } else {
++                    // Cross-node candidate: save as fallback, keep scanning for same-node
++                    if cross_node_fallback.is_none() {
++                        // Cache the priority and balance for later
++                        cross_node_fallback =
++                            Some((next_context_guard.prio, next_context_guard));
++                        balance[i] -= SCHED_PRIO_TO_WEIGHT[20];
++                        // Don't break — keep looking for a same-node context
++                        continue;
++                    } else {
++                        // Already have a cross-node fallback; push this one back
++                        contexts.push_back(next_context_ref);
++                        skipped_contexts += 1;
++                        if skipped_contexts >= total_contexts {
++                            break 'priority;
++                        }
++                        continue;
++                    }
++                }
+             } else {
+                 if matches!(sw, UpdateResult::Blocked) {
+                     idle_contexts(token.token()).push_back(next_context_ref);
+@@ -469,6 +507,15 @@
+             }
+         }
+     }
++
++    // If we found a cross-node fallback but no same-node context, use it
++    if next_context_guard_opt.is_none() {
++        if let Some((prio, guard)) = cross_node_fallback {
++            percpu.current_prio.set(prio);
++            next_context_guard_opt = Some(guard);
++        }
++    }
++
+     percpu.balance.set(balance);
+     percpu.last_queue.set(i);
+