Stage 2: cooperative VM context switch primitive, proven on all 3 arches (FABRIC-3.md §XXVIII)

Third stage of the preemptive context-switching plan. The real
save/restore switch mechanism now exists -- the first time anything has
ever executed on a VM's own native stack (Stage 1 allocated them,
unused).

New sk_vm_switch_to() (switch.S, one per arch) is an ordinary function
call, not an interrupt -- so unlike Stage 0's trap frame, the ABI already
covers every caller-saved register; only the callee-saved set needs
explicit save/restore (amd64: rbx/rbp/r12-r15, no FP at all since SysV
has no callee-saved XMM; aarch64: x19-x28/x29/x30 + d8-d15; riscv64:
s0-s11/ra + fs0-fs11, FS-gated like Stage 0 but read once and reused for
both halves within one call, since FS is genuine global CPU state, not
part of what's switched). A sibling sk_vm_switch_prime() in the same file
builds the synthetic first-entry frame, kept in assembly so the layout
can never drift out of sync with sk_vm_switch_to() itself.

New switch.c/switch.h: sk_vm_context_switch(from, to) handles first-entry
priming vs. resuming a parked context, and updates registry state (new
VM_STATE_SWITCHED_OUT, distinct from VM_STATE_STOPPED -- STOPPED means no
live frame, this means the opposite). sk_vm_switch_entry() is the minimal
permanent trampoline every freshly-entered VM lands in: no production
behavior defined yet, so it just yields straight back to whoever switched
to it, forever.

Closes the confirmed unguarded-KILL UAF found during planning:
capsule_vm_kill(), mama_word_kill(), and capsule_vm_kill_all_nonmama()
all now refuse (or silently leak rather than free, on the cold-restart
path where arch_cold_reset() wipes everything immediately after anyway)
tearing down a switched-out VM. Side effect found, not built on purpose:
the existing MSG-TICK idle-pump already filters on VM_STATE_LIVE, so it
automatically stopped dispatching into a switched-out VM with zero
changes needed there.

Verified via a temporary SWITCH-TEST probe (boot-triggered, since nothing
can type interactively into a foreground-only QEMU session) that
round-tripped a sentinel through 5 real Hera<->Hermes switches on all 3
architectures: 5/5 rounds, 0 failures, clean continuation to ok>. Probe
fully reverted after capture; kernel_main.c shows zero diff.

Also: Makefile.starkernel's LOADER_EXTRA_SRCS/LOADER_ASM needed the new
files added explicitly (this project's "loader" PE binary is the full
running kernel, not a thin bootstrap stage), and aarch64's switch.S
needed the same #ifndef _WIN32 guard around .hidden that isr.S already
carries (aarch64's loader assembles via clang targeting a PE/COFF
target with no .hidden equivalent) -- caught by a build failure, fixed.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016UNhH1mhi52i6Qihh7ZV5S

FABRIC-3.md

@@ -3579,3 +3579,59 @@ All 3 architectures re-verified clean boot to `ok>`, no native-stack allocation
 for any Tripod-fleet VM on any arch. Stage 2 (the actual save/restore switch primitive,
 cooperative only, no timer) is next.
 
+**Stage 2 CLOSED, same day.** The real save/restore switch primitive now exists and is proven
+correct on all 3 architectures -- the first time anything has ever executed on a VM's own
+native stack (Stage 1's stacks were allocated but unused until now).
+
+Deliberately NOT built the same way as Stage 0's interrupt trap frame -- `sk_vm_switch_to()`
+(new `switch.S` beside each arch's `isr.S`) is an ordinary function call, so the standard ABI
+already guarantees every caller-saved register is the caller's own problem; only the
+callee-saved set needs explicit handling (amd64: rbx/rbp/r12-r15, no FP at all since SysV has
+no callee-saved XMM; aarch64: x19-x28/x29/x30 + d8-d15; riscv64: s0-s11/ra + fs0-fs11, the FP
+half still correctly FS-gated like Stage 0's trap frame, but read once and reused for both the
+save and restore decision within one call rather than persisted -- FS is genuine single global
+CPU state, not part of what's "switched"). A sibling `sk_vm_switch_prime()` in each same file
+builds the synthetic first-entry frame (zeroed callee-saved slots + a trampoline address in the
+return-address slot) so a VM's first-ever switch-in lands somewhere real instead of into
+whatever real prior caller the layout would otherwise imply -- kept in assembly, not C, so the
+frame layout can never drift out of sync between the two functions.
+
+New C glue (`vm/switch.c`, `vm/switch.h`): `sk_vm_context_switch(from, to)` -- handles first-entry
+priming vs. resuming a previously-parked context, and updates registry state around the switch
+(new `VM_STATE_SWITCHED_OUT`, distinct from `VM_STATE_STOPPED` per the Stage 1 correction --
+STOPPED means no live frame, SWITCHED_OUT means the opposite). `sk_vm_switch_entry()` is the
+minimal permanent trampoline every freshly-entered VM lands in: no production behavior is
+defined yet (that's later-stage work), so it just yields straight back to whoever switched to
+it, forever -- must never fall off the end, since there is no legitimate return address below a
+synthetic first-entry frame.
+
+**Closed the confirmed unguarded-KILL UAF** found during planning: `capsule_vm_kill()`,
+`mama_word_kill()`, and `capsule_vm_kill_all_nonmama()` all now refuse (or, for the
+cold-restart-only path, silently leak rather than free -- harmless there specifically, since
+`arch_cold_reset()` always follows immediately and wipes everything regardless) tearing down a
+VM in `VM_STATE_SWITCHED_OUT`. A nice side effect discovered while verifying, not something
+built deliberately: the existing MSG-TICK idle-pump (`repl.c`) already filters on
+`VM_STATE_LIVE`, so it automatically stopped dispatching into Hermes the moment she was
+switched out, with zero code changes needed there.
+
+**Verification, per this project's write-probe-capture-revert discipline:** a temporary
+`SWITCH-TEST` word (registered in Hera's own vocabulary, triggered once at boot right after
+Hermes's birth, since there's no way to type interactively into a foreground-only QEMU session)
+round-tripped a distinctive sentinel through 5 real Hera<->Hermes context switches. All 3
+architectures: 5/5 rounds, 0 failures, PASS, clean continuation of boot to `ok>` afterward. The
+probe word, its boot-time trigger, and its helper were then fully reverted -- re-verified clean
+on all 3 architectures again with the probe gone, confirming no residual effects (Hermes back to
+normal `VM_STATE_LIVE` MSG-TICK dispatch once no longer switched out).
+
+Also fixed along the way: `Makefile.starkernel`'s `LOADER_EXTRA_SRCS`/`LOADER_ASM` lists needed
+`switch.c`/`switch.S` added explicitly (this project's "loader" PE binary is effectively the
+full running kernel, not a thin bootstrap stage -- confirmed by `isr.S` already being in
+`LOADER_ASM`); and aarch64's `switch.S` needed the same `#ifndef _WIN32` guard around `.hidden`
+that `isr.S` already carries, since aarch64's loader is assembled via clang targeting
+`aarch64-pc-windows-msvc` (PE/COFF, no `.hidden` equivalent) -- missed on the first pass, caught
+by the aarch64 build failing outright.
+
+Stage 3 (timer-driven preemption, Tripod fleet only, the new run-readiness signal) is next --
+the biggest remaining stage, and the first one that actually amends the §21.1
+nothing-is-concurrent-on-one-hart ruling for real.
+

Makefile.starkernel

@@ -462,7 +462,8 @@ LOADER_SRCS_BASE := \
 
 LOADER_ASM := \
     $(KERNEL_SRC)/arch/$(ARCH)/boot.S \
-    $(KERNEL_SRC)/arch/$(ARCH)/isr.S
+    $(KERNEL_SRC)/arch/$(ARCH)/isr.S \
+    $(KERNEL_SRC)/arch/$(ARCH)/switch.S
 
 ifeq ($(ARCH),amd64)
 LOADER_ASM += $(KERNEL_SRC)/arch/$(ARCH)/kernel_entry.S
@@ -539,6 +540,7 @@ VM_CORE_SRCS := $(filter-out $(VM_EXCLUDE),$(VM_ALL_SRCS))
 
 LOADER_EXTRA_SRCS := \
     $(KERNEL_SRC)/vm/arena.c \
+    $(KERNEL_SRC)/vm/switch.c \
     $(KERNEL_SRC)/vm/parity.c \
     $(KERNEL_SRC)/vm/bootstrap/sk_vm_bootstrap.c \
     $(KERNEL_SRC)/vm/host/shim.c \

capsules/BLOCK_MAP.md

@@ -1,5 +1,5 @@
 # Capsule Block Manifest — Auto-generated
-<!-- Generated by mkcapsule --manifest  2026-09-13T18:28:32Z -->
+<!-- Generated by mkcapsule --manifest  2026-09-13T19:23:06Z -->
 <!-- DO NOT EDIT — re-run mkcapsule --manifest to refresh.  -->
 <!-- Hand-written justifications and immutability notes live  -->
 <!-- in MANIFEST.md alongside this auto-generated index.      -->

include/starkernel/capsule_run.h

@@ -87,9 +87,23 @@ typedef struct {
 typedef enum {
     VM_STATE_EMBRYO = 0,     /* Allocated but not yet born */
     VM_STATE_LIVE,           /* Successfully born, operational */
-    VM_STATE_STOPPED,        /* Suspended — execution state saved */
+    VM_STATE_STOPPED,        /* Suspended — execution state saved. Set by
+                              * the START word after STOP cleanly unwinds
+                              * its C stack back to START's own frame: no
+                              * live native frame remains, safe to free. */
     VM_STATE_STILLBORN,      /* Birth failed */
     VM_STATE_DEAD,           /* Terminated */
+    VM_STATE_SWITCHED_OUT,   /* FABRIC-3.md §XXVIII, Stage 2 (2026-09-13):
+                              * a live saved register/stack context is
+                              * parked on this VM's own native stack
+                              * (SWITCH-TO switched control away mid-
+                              * execution). Deliberately distinct from
+                              * VM_STATE_STOPPED -- that state means "no
+                              * live native frame," this one means the
+                              * opposite. KILL must refuse/defer here: the
+                              * parked frame still points into vm->memory
+                              * and the native stack, both of which a free
+                              * would invalidate out from under it. */
 } VMState;
 
 typedef struct {

include/starkernel/vm/switch.h

@@ -0,0 +1,51 @@
+/*
+  StarForth — Steady-State Virtual Machine Runtime
+
+  Copyright (c) 2023–2025 Robert A. James
+  All rights reserved.
+
+  This file is part of the StarForth project.
+
+  Licensed under the StarForth License, Version 1.0 (the "License");
+  you may not use this file except in compliance with the License.
+
+  You may obtain a copy of the License at:
+      https://github.com/star.4th@proton.me/StarForth/LICENSE.txt
+
+  This software is provided "AS IS", WITHOUT WARRANTY OF ANY KIND,
+  express or implied, including but not limited to the warranties of
+  merchantability, fitness for a particular purpose, and noninfringement.
+
+ See the License for the specific language governing permissions and
+ limitations under the License.
+*/
+
+/**
+ * switch.h - Cooperative VM context switch (FABRIC-3.md §XXVIII, Stage 2,
+ * 2026-09-13).
+ *
+ * sk_vm_context_switch() suspends `from`'s execution exactly where it is
+ * (mid-C-call-stack, on from's own native stack -- Stage 1) and resumes
+ * `to` -- either for the first time ever (a synthesized initial frame,
+ * entering sk_vm_switch_entry()) or exactly where `to` was itself last
+ * switched out. Returns to the caller only once something later switches
+ * back to `from` -- from from's own point of view, this call simply
+ * takes a while to return, like any blocking call.
+ *
+ * Cooperative only: nothing here is interrupt-driven yet (that's Stage
+ * 3), so this is safe to call only from ordinary FORTH word dispatch,
+ * never from ISR context.
+ */
+
+#ifndef STARKERNEL_VM_SWITCH_H
+#define STARKERNEL_VM_SWITCH_H
+
+#ifdef __STARKERNEL__
+
+struct VM;
+
+int sk_vm_context_switch(struct VM *from, struct VM *to);
+
+#endif /* __STARKERNEL__ */
+
+#endif /* STARKERNEL_VM_SWITCH_H */

include/vm.h

@@ -650,9 +650,14 @@ typedef struct VM
      */
     uint64_t native_stack_paddr;       /**< Physical base (for teardown); 0 = not allocated */
     uint64_t native_stack_guard_vaddr; /**< Virtual base of the whole guarded region */
-    uint64_t native_stack_top;         /**< Initial SP value once something switches onto
+    uint64_t native_stack_top;         /**< Top of the allocated region (fixed, for the
-                                         *   this stack -- Stage 1 only allocates it; nothing
+                                         *   initial frame construction); 0 = not allocated. */
-                                         *   yet runs here (that's Stage 2). */
+    uint64_t native_stack_saved_sp;    /**< Stage 2 (FABRIC-3.md §XXVIII, 2026-09-13): the
+                                         *   parked SP of a live, switched-out register
+                                         *   context on this VM's own native stack. 0 = never
+                                         *   entered (no context parked yet); non-zero only
+                                         *   while VMRegistryEntry.state ==
+                                         *   VM_STATE_SWITCHED_OUT for this VM. */
     /** @} */
 
     /** @name Stadium Identity (item 4.2, FABRIC-0.md §25.5)

src/starkernel/arch/aarch64/switch.S

@@ -0,0 +1,105 @@
+/*
+ * switch.S (aarch64) - Cooperative VM context switch primitive.
+ *
+ * FABRIC-3.md §XXVIII (Stage 2, preemptive-context-switch save/restore
+ * primitive, 2026-09-13). Deliberately NOT isr.S's interrupt trap frame
+ * (Stage 0) -- that saves everything because an interrupt has no call
+ * site. sk_vm_switch_to() is an ordinary `bl`, so AAPCS64 already makes
+ * every caller-saved register (x0-x18, v0-v7, v16-v31) the caller's own
+ * problem across the call. Only the callee-saved set needs explicit
+ * handling: x19-x28, x29 (fp), x30 (lr), and d8-d15 (only the low 64
+ * bits of v8-v15 are callee-saved per AAPCS64).
+ *
+ * void sk_vm_switch_to(uint64_t *save_sp_out, uint64_t new_sp);
+ *   x0 = save_sp_out, x1 = new_sp
+ *
+ * For a VM that has never run before, new_sp must point at a synthetic
+ * 160-byte frame built by the C glue (switch.c) to this exact layout,
+ * with x30 (lr) set to the trampoline entry address and everything else
+ * zeroed, so the final `ret` below (which branches to lr) jumps there.
+ */
+
+    .section .text
+    .balign 4
+
+    .equ SW_X19, 0
+    .equ SW_X21, 16
+    .equ SW_X23, 32
+    .equ SW_X25, 48
+    .equ SW_X27, 64
+    .equ SW_X29, 80   /* x29 (fp), x30 (lr) */
+    .equ SW_D8,  96
+    .equ SW_D10, 112
+    .equ SW_D12, 128
+    .equ SW_D14, 144
+    .equ SW_SIZE, 160
+
+/* .hidden is an ELF-only visibility directive with no COFF equivalent;
+ * this file is also assembled for the aarch64 loader, which clang targets
+ * at aarch64-pc-windows-msvc (PE/COFF) -- guard on _WIN32 (defined only by
+ * that target), matching isr.S's own precedent exactly. */
+#ifndef _WIN32
+.hidden sk_vm_switch_to
+#endif
+
+    .global sk_vm_switch_to
+sk_vm_switch_to:
+    sub sp, sp, #SW_SIZE
+
+    stp x19, x20, [sp, #SW_X19]
+    stp x21, x22, [sp, #SW_X21]
+    stp x23, x24, [sp, #SW_X23]
+    stp x25, x26, [sp, #SW_X25]
+    stp x27, x28, [sp, #SW_X27]
+    stp x29, x30, [sp, #SW_X29]
+    stp d8,  d9,  [sp, #SW_D8]
+    stp d10, d11, [sp, #SW_D10]
+    stp d12, d13, [sp, #SW_D12]
+    stp d14, d15, [sp, #SW_D14]
+
+    mov x2, sp
+    str x2, [x0]         /* *save_sp_out = sp (this context's parked SP) */
+
+    mov sp, x1            /* sp = new_sp (switch stacks) */
+
+    ldp x19, x20, [sp, #SW_X19]
+    ldp x21, x22, [sp, #SW_X21]
+    ldp x23, x24, [sp, #SW_X23]
+    ldp x25, x26, [sp, #SW_X25]
+    ldp x27, x28, [sp, #SW_X27]
+    ldp x29, x30, [sp, #SW_X29]
+    ldp d8,  d9,  [sp, #SW_D8]
+    ldp d10, d11, [sp, #SW_D10]
+    ldp d12, d13, [sp, #SW_D12]
+    ldp d14, d15, [sp, #SW_D14]
+
+    add sp, sp, #SW_SIZE
+    ret
+
+/*
+ * uint64_t sk_vm_switch_prime(uint64_t stack_top, void *entry);
+ *
+ * Builds the synthetic frame above at the top of a fresh native stack, so
+ * the first ever sk_vm_switch_to() into it lands at `entry` (x30/lr) via
+ * its final `ret`. Kept in this file (not C) so the layout can never
+ * drift from sk_vm_switch_to() above -- one set of offsets, not two.
+ */
+#ifndef _WIN32
+.hidden sk_vm_switch_prime
+#endif
+
+    .global sk_vm_switch_prime
+sk_vm_switch_prime:
+    sub x2, x0, #SW_SIZE
+    stp xzr, xzr, [x2, #SW_X19]
+    stp xzr, xzr, [x2, #SW_X21]
+    stp xzr, xzr, [x2, #SW_X23]
+    stp xzr, xzr, [x2, #SW_X25]
+    stp xzr, xzr, [x2, #SW_X27]
+    stp xzr, x1,  [x2, #SW_X29]   /* x29=0, x30(lr)=entry */
+    stp xzr, xzr, [x2, #SW_D8]
+    stp xzr, xzr, [x2, #SW_D10]
+    stp xzr, xzr, [x2, #SW_D12]
+    stp xzr, xzr, [x2, #SW_D14]
+    mov x0, x2
+    ret

src/starkernel/arch/amd64/switch.S

@@ -0,0 +1,79 @@
+/*
+ * switch.S (amd64) - Cooperative VM context switch primitive.
+ *
+ * FABRIC-3.md §XXVIII (Stage 2, preemptive-context-switch save/restore
+ * primitive, 2026-09-13). This is deliberately NOT the same mechanism as
+ * isr.S's interrupt trap frame (Stage 0) -- that saves every register
+ * because an interrupt is not a call site the interrupted code knows
+ * about. sk_vm_switch_to() IS an ordinary function call (SWITCH-TO's C
+ * handler calls it directly), so the standard System V AMD64 ABI already
+ * guarantees every caller-saved register (rax, rcx, rdx, rsi, rdi, r8-r11,
+ * all XMM) is the caller's own problem to preserve across the call, exactly
+ * like any other call. Only the callee-saved set -- rbx, rbp, r12-r15 --
+ * needs explicit handling here, because this function does not really
+ * "return" in the normal sense: it swaps to a completely different stack
+ * and (from the newly-resumed context's point of view) appears to return
+ * from whichever earlier call to this same function originally parked it.
+ * SysV has no callee-saved XMM registers at all, so there is no FP state
+ * to carry here.
+ *
+ * void sk_vm_switch_to(uint64_t *save_sp_out, uint64_t new_sp);
+ *   rdi = save_sp_out, rsi = new_sp
+ *
+ * For a VM that has never run before, new_sp must point at a synthetic
+ * frame built to this exact layout by the C glue (switch.c): six zeroed
+ * quadwords (r15, r14, r13, r12, rbx, rbp -- values irrelevant, never a
+ * real prior context) followed by the trampoline entry address at
+ * new_sp+48, so the final `ret` below jumps there instead of into some
+ * prior caller.
+ */
+
+    .intel_syntax noprefix
+    .section .text
+
+.hidden sk_vm_switch_to
+
+    .global sk_vm_switch_to
+sk_vm_switch_to:
+    push rbp
+    push rbx
+    push r12
+    push r13
+    push r14
+    push r15
+
+    mov [rdi], rsp      /* *save_sp_out = rsp (this context's parked SP) */
+    mov rsp, rsi         /* rsp = new_sp (switch stacks) */
+
+    pop r15
+    pop r14
+    pop r13
+    pop r12
+    pop rbx
+    pop rbp
+    ret
+
+/*
+ * uint64_t sk_vm_switch_prime(uint64_t stack_top, void *entry);
+ *
+ * Builds the synthetic frame described above at the top of a fresh native
+ * stack, so the first ever sk_vm_switch_to() into it lands at `entry`
+ * instead of some real prior caller. Returns the resulting sp value to
+ * pass as sk_vm_switch_to()'s new_sp. Kept in this file (not C) so the
+ * layout can never drift from what sk_vm_switch_to() above actually
+ * expects -- one set of offsets, not two copies to keep in sync.
+ */
+.hidden sk_vm_switch_prime
+
+    .global sk_vm_switch_prime
+sk_vm_switch_prime:
+    mov rax, rdi
+    sub rax, 56
+    mov qword ptr [rax+0],  0    /* r15 */
+    mov qword ptr [rax+8],  0    /* r14 */
+    mov qword ptr [rax+16], 0    /* r13 */
+    mov qword ptr [rax+24], 0    /* r12 */
+    mov qword ptr [rax+32], 0    /* rbx */
+    mov qword ptr [rax+40], 0    /* rbp */
+    mov [rax+48], rsi             /* return address = entry */
+    ret

src/starkernel/arch/riscv64/switch.S

@@ -0,0 +1,162 @@
+/*
+ * switch.S (riscv64) - Cooperative VM context switch primitive.
+ *
+ * FABRIC-3.md §XXVIII (Stage 2, preemptive-context-switch save/restore
+ * primitive, 2026-09-13). Deliberately NOT isr.S's interrupt trap frame
+ * (Stage 0) -- that saves everything because an async trap has no call
+ * site. sk_vm_switch_to() is an ordinary `call`, so the LP64D psABI
+ * already makes every caller-saved register (ra excepted -- see below --
+ * t0-t6, a0-a7, ft0-ft11, fa0-fa7) the caller's own problem across the
+ * call. Only the callee-saved set needs explicit handling: s0-s11,
+ * fs0-fs11, and ra itself (ra is not "callee-saved" in the usual sense,
+ * but this function does not really return to its own caller -- it swaps
+ * stacks and later resumes as if returning from whichever earlier call
+ * parked this context, so ra must be captured explicitly here rather than
+ * left to a normal prologue/epilogue).
+ *
+ * The FP half is saved only when sstatus.FS != Off -- same reasoning and
+ * bit test as isr.S's trap frame, checked once per call and reused for
+ * both the save and restore decision (nothing changes FS in between; it
+ * is genuine single global CPU state, not part of what's "switched," so
+ * this function reads it but never writes it).
+ *
+ * void sk_vm_switch_to(uint64_t *save_sp_out, uint64_t new_sp);
+ *   a0 = save_sp_out, a1 = new_sp
+ *
+ * For a VM that has never run before, new_sp must point at a synthetic
+ * 208-byte frame built by the C glue (switch.c) with ra set to the
+ * trampoline entry address and everything else zeroed, so the final
+ * `ret` below (jalr via ra) jumps there.
+ */
+
+    .section .text
+    .balign 4
+
+    .equ SW_S0,   0
+    .equ SW_RA,   96
+    .equ SW_FS0,  104
+    .equ SW_SIZE, 208
+
+    .equ SSTATUS_FS_SHIFT, 13
+    .equ SSTATUS_FS_MASK,   3
+
+.hidden sk_vm_switch_to
+
+    .global sk_vm_switch_to
+sk_vm_switch_to:
+    addi sp, sp, -SW_SIZE
+
+    sd s0,  SW_S0+0(sp)
+    sd s1,  SW_S0+8(sp)
+    sd s2,  SW_S0+16(sp)
+    sd s3,  SW_S0+24(sp)
+    sd s4,  SW_S0+32(sp)
+    sd s5,  SW_S0+40(sp)
+    sd s6,  SW_S0+48(sp)
+    sd s7,  SW_S0+56(sp)
+    sd s8,  SW_S0+64(sp)
+    sd s9,  SW_S0+72(sp)
+    sd s10, SW_S0+80(sp)
+    sd s11, SW_S0+88(sp)
+    sd ra,  SW_RA(sp)
+
+    csrr t0, sstatus
+    srli t1, t0, SSTATUS_FS_SHIFT
+    andi t1, t1, SSTATUS_FS_MASK
+    beqz t1, 1f
+
+    fsd fs0,  SW_FS0+0(sp)
+    fsd fs1,  SW_FS0+8(sp)
+    fsd fs2,  SW_FS0+16(sp)
+    fsd fs3,  SW_FS0+24(sp)
+    fsd fs4,  SW_FS0+32(sp)
+    fsd fs5,  SW_FS0+40(sp)
+    fsd fs6,  SW_FS0+48(sp)
+    fsd fs7,  SW_FS0+56(sp)
+    fsd fs8,  SW_FS0+64(sp)
+    fsd fs9,  SW_FS0+72(sp)
+    fsd fs10, SW_FS0+80(sp)
+    fsd fs11, SW_FS0+88(sp)
+1:
+    sd sp, 0(a0)          /* *save_sp_out = sp (this context's parked SP) */
+    mv sp, a1              /* sp = new_sp (switch stacks) */
+
+    ld s0,  SW_S0+0(sp)
+    ld s1,  SW_S0+8(sp)
+    ld s2,  SW_S0+16(sp)
+    ld s3,  SW_S0+24(sp)
+    ld s4,  SW_S0+32(sp)
+    ld s5,  SW_S0+40(sp)
+    ld s6,  SW_S0+48(sp)
+    ld s7,  SW_S0+56(sp)
+    ld s8,  SW_S0+64(sp)
+    ld s9,  SW_S0+72(sp)
+    ld s10, SW_S0+80(sp)
+    ld s11, SW_S0+88(sp)
+    ld ra,  SW_RA(sp)
+
+    beqz t1, 2f
+
+    fld fs0,  SW_FS0+0(sp)
+    fld fs1,  SW_FS0+8(sp)
+    fld fs2,  SW_FS0+16(sp)
+    fld fs3,  SW_FS0+24(sp)
+    fld fs4,  SW_FS0+32(sp)
+    fld fs5,  SW_FS0+40(sp)
+    fld fs6,  SW_FS0+48(sp)
+    fld fs7,  SW_FS0+56(sp)
+    fld fs8,  SW_FS0+64(sp)
+    fld fs9,  SW_FS0+72(sp)
+    fld fs10, SW_FS0+80(sp)
+    fld fs11, SW_FS0+88(sp)
+2:
+    addi sp, sp, SW_SIZE
+    ret
+
+/*
+ * uint64_t sk_vm_switch_prime(uint64_t stack_top, void *entry);
+ *
+ * Builds the synthetic frame above at the top of a fresh native stack, so
+ * the first ever sk_vm_switch_to() into it lands at `entry` (ra) via its
+ * final `ret`. Kept in this file (not C) so the layout can never drift
+ * from sk_vm_switch_to() above -- one set of offsets, not two. FP slots
+ * are zeroed via plain integer stores (sd, not fsd) -- this is just
+ * writing zero bytes to memory, not an FP computation, so it needs no
+ * FS-enabled check at all, unlike every other f-register touch in this
+ * file.
+ */
+.hidden sk_vm_switch_prime
+
+    .global sk_vm_switch_prime
+sk_vm_switch_prime:
+    addi t0, a0, -SW_SIZE
+
+    sd zero, SW_S0+0(t0)
+    sd zero, SW_S0+8(t0)
+    sd zero, SW_S0+16(t0)
+    sd zero, SW_S0+24(t0)
+    sd zero, SW_S0+32(t0)
+    sd zero, SW_S0+40(t0)
+    sd zero, SW_S0+48(t0)
+    sd zero, SW_S0+56(t0)
+    sd zero, SW_S0+64(t0)
+    sd zero, SW_S0+72(t0)
+    sd zero, SW_S0+80(t0)
+    sd zero, SW_S0+88(t0)
+    sd a1,   SW_RA(t0)
+
+    sd zero, SW_FS0+0(t0)
+    sd zero, SW_FS0+8(t0)
+    sd zero, SW_FS0+16(t0)
+    sd zero, SW_FS0+24(t0)
+    sd zero, SW_FS0+32(t0)
+    sd zero, SW_FS0+40(t0)
+    sd zero, SW_FS0+48(t0)
+    sd zero, SW_FS0+56(t0)
+    sd zero, SW_FS0+64(t0)
+    sd zero, SW_FS0+72(t0)
+    sd zero, SW_FS0+80(t0)
+    sd zero, SW_FS0+88(t0)
+
+    mv a0, t0
+    ret

src/starkernel/capsule/capsule_birth.c

@@ -392,6 +392,19 @@ int capsule_vm_kill(const char *name) {
         return 0;
     }
 
+    /* FABRIC-3.md §XXVIII, Stage 2 (2026-09-13): a switched-out VM has a
+     * live saved context parked on its own native stack -- freeing vm_ptr
+     * here would leave that saved SP pointing into freed memory, a real
+     * use-after-free the moment anything ever tried to resume it. Refuse
+     * rather than silently corrupt; the caller can retry once the VM has
+     * been switched back in and reaches a normal LIVE state. */
+    if (entry->state == VM_STATE_SWITCHED_OUT) {
+        console_puts("KILL: ");
+        console_puts(name);
+        console_println(" refused -- switched-out (context parked)");
+        return -1;
+    }
+
     vm_id = entry->vm_id;
     vm    = (VM *)entry->vm_ptr;
 
@@ -440,7 +453,17 @@ void capsule_vm_kill_all_nonmama(void) {
             (void)stadium_evict(node->entry.stadium_patron_cell);
             node->entry.stadium_patron_cell = STADIUM_CELL_NONE;
         }
-        if (vm) {
+        /* FABRIC-3.md §XXVIII, Stage 2 (2026-09-13): a switched-out VM has
+         * a live saved context parked on its own native stack -- skip the
+         * free (leaked, not corrupted) rather than invalidate memory a
+         * parked SP still points into. Harmless here specifically: every
+         * caller of this function (BYE/cold-restart) calls
+         * arch_cold_reset() immediately afterward, which wipes all memory
+         * regardless of what this function left allocated. */
+        if (vm && node->entry.state == VM_STATE_SWITCHED_OUT) {
+            log_message(LOG_WARN, "capsule_vm_kill_all_nonmama: leaving %s allocated (switched-out)",
+                        node->entry.name);
+        } else if (vm) {
             vm->halted = 1;
             vm_cleanup(vm);
             sf_free(vm);

src/starkernel/capsule/mama_forth_words.c

@@ -586,6 +586,15 @@ void mama_word_kill(VM *vm)
      * before capsule_vm_kill()'s own Hera guard below ever runs. */
     {
         VMRegistryEntry entry;
+        /* FABRIC-3.md §XXVIII, Stage 2 (2026-09-13): mirror the refusal
+         * capsule_vm_kill() itself now enforces -- checked here too so the
+         * heat-fanout/active-VM cleanup below never runs against a VM
+         * that's about to be refused anyway. */
+        if (capsule_vm_find_by_name_nocase(name_buf, &entry) == 0 &&
+            entry.state == VM_STATE_SWITCHED_OUT) {
+            console_println("KILL: refused -- switched-out (context parked)");
+            return;
+        }
         if (capsule_vm_find_by_name_nocase(name_buf, &entry) == 0 &&
             entry.state == VM_STATE_LIVE) {
             if (!vm_uuid_is_hera(entry.vm_id)) {

src/starkernel/vm/switch.c

@@ -0,0 +1,98 @@
+/*
+  StarForth — Steady-State Virtual Machine Runtime
+
+  Copyright (c) 2023–2025 Robert A. James
+  All rights reserved.
+
+  This file is part of the StarForth project.
+
+  Licensed under the StarForth License, Version 1.0 (the "License");
+  you may not use this file except in compliance with the License.
+
+  You may obtain a copy of the License at:
+      https://github.com/star.4th@proton.me/StarForth/LICENSE.txt
+
+  This software is provided "AS IS", WITHOUT WARRANTY OF ANY KIND,
+  express or implied, including but not limited to the warranties of
+  merchantability, fitness for a particular purpose, and noninfringement.
+
+ See the License for the specific language governing permissions and
+ limitations under the License.
+*/
+
+/**
+ * switch.c - Cooperative VM context switch glue (FABRIC-3.md §XXVIII,
+ * Stage 2, 2026-09-13). See switch.h for the contract.
+ */
+
+#ifdef __STARKERNEL__
+
+#include "starkernel/vm/switch.h"
+#include "vm.h"
+#include "starkernel/capsule_birth.h" /* capsule_vm_set_state */
+#include <stdint.h>
+
+/* Arch-specific (switch.S in each arch dir). sk_vm_switch_prime() builds
+ * the exact synthetic first-entry frame sk_vm_switch_to() expects -- kept
+ * in assembly, not duplicated here, so the layout can never drift out of
+ * sync between the two. */
+extern void     sk_vm_switch_to(uint64_t *save_sp_out, uint64_t new_sp);
+extern uint64_t sk_vm_switch_prime(uint64_t stack_top, void *entry);
+
+/* Single-writer-mainline globals, safe under this stage's cooperative-
+ * only, one-VM-runs-at-a-time-on-one-hart discipline (same class of
+ * bookkeeping as vm_core.c's g_log_attrib_vm) -- set immediately before a
+ * first-ever switch into a VM, read exactly once, synchronously, at the
+ * very top of sk_vm_switch_entry() before any other switch could occur.
+ * Preemption (Stage 3) will need to revisit this; flagged there, not here. */
+static VM *g_switch_entry_vm;
+static VM *g_switch_back_to;
+
+/**
+ * sk_vm_switch_entry - trampoline for a VM's first-ever entry.
+ *
+ * A freshly-entered VM has no production behavior defined yet -- what a
+ * switched-to VM should actually DO once running is later-stage work
+ * (message dispatch, an idle loop, whatever Stage 3+ needs). For now this
+ * is a minimal, permanent placeholder: immediately yield back to whoever
+ * switched to it, forever. Must never fall off the end -- there is no
+ * legitimate return address below a synthetic first-entry frame, unlike
+ * an ordinary function.
+ */
+void sk_vm_switch_entry(void) {
+    VM *self = g_switch_entry_vm;
+    VM *back = g_switch_back_to;
+    for (;;) {
+        sk_vm_context_switch(self, back);
+    }
+}
+
+int sk_vm_context_switch(VM *from, VM *to) {
+    uint64_t new_sp;
+
+    if (!from || !to || to->native_stack_top == 0) {
+        return -1;
+    }
+
+    if (to->native_stack_saved_sp == 0) {
+        /* First-ever entry: synthesize the initial frame. */
+        g_switch_entry_vm = to;
+        g_switch_back_to  = from;
+        new_sp = sk_vm_switch_prime(to->native_stack_top, (void *)&sk_vm_switch_entry);
+    } else {
+        new_sp = to->native_stack_saved_sp;
+        to->native_stack_saved_sp = 0; /* about to be running, not parked */
+    }
+
+    capsule_vm_set_state(to->stadium_vm_id, VM_STATE_LIVE);
+    capsule_vm_set_state(from->stadium_vm_id, VM_STATE_SWITCHED_OUT);
+
+    sk_vm_switch_to(&from->native_stack_saved_sp, new_sp);
+
+    /* Resumes here only once something later switches back to `from` --
+     * from `from`'s own point of view, this call simply took a while. */
+    capsule_vm_set_state(from->stadium_vm_id, VM_STATE_LIVE);
+    return 0;
+}
+
+#endif /* __STARKERNEL__ */