diff --git a/Makefile b/Makefile
index 5afef93..a906bad 100644
--- a/Makefile
+++ b/Makefile
@@ -95,7 +95,9 @@ endif
 ifeq ($(TARGET),F303)
   USE_FPU = hard
   USE_PROCESS_STACKSIZE = 0x480
-  USE_EXCEPTIONS_STACKSIZE = 0x200
+  # The fault reporter renders through the LCD/printf stack. Reserve room for
+  # its assembly veneer, an extended exception frame, and a nested ISR frame.
+  USE_EXCEPTIONS_STACKSIZE = 0x400
 endif
 
 #
diff --git a/docs/CHIBIOS_21_11_5_PORT.md b/docs/CHIBIOS_21_11_5_PORT.md
index c1bb16a..8caf507 100644
--- a/docs/CHIBIOS_21_11_5_PORT.md
+++ b/docs/CHIBIOS_21_11_5_PORT.md
@@ -34,6 +34,8 @@ this port is marked ready.
   explicitly. ChibiOS 21.11.x no longer adds it automatically; omitting it
   promotes unsuffixed sweep constants to software double precision and causes
   a measurable self-test sweep regression.
+- Split the F303 HardFault entry into an assembly-only MSP/PSP veneer and an
+  ordinary non-returning C reporter, while reserving a 1 KiB exception stack.
 
 ## Build
 
diff --git a/main.c b/main.c
index 745074a..119afa6 100644
--- a/main.c
+++ b/main.c
@@ -3451,8 +3451,35 @@ int main(void)
   }
 }
 
-/* The prototype shows it is a naked function - in effect this is just an
-assembly function. */
+#ifdef TINYSA4
+void hard_fault_handler_c(uint32_t *sp, const uint32_t *callee)
+    __attribute__((noreturn, noinline));
+void HardFault_Handler(void) __attribute__((naked));
+
+void HardFault_Handler(void)
+{
+  __asm volatile(
+      "tst lr, #4\n"
+      "ite eq\n"
+      "mrseq r0, msp\n"
+      "mrsne r0, psp\n"
+      "sub sp, sp, #32\n"
+      "str r4, [sp, #0]\n"
+      "str r5, [sp, #4]\n"
+      "str r6, [sp, #8]\n"
+      "str r7, [sp, #12]\n"
+      "mov r2, r8\n"
+      "str r2, [sp, #16]\n"
+      "mov r2, r9\n"
+      "str r2, [sp, #20]\n"
+      "mov r2, r10\n"
+      "str r2, [sp, #24]\n"
+      "mov r2, r11\n"
+      "str r2, [sp, #28]\n"
+      "mov r1, sp\n"
+      "b hard_fault_handler_c\n");
+}
+#else
 void HardFault_Handler(void);
 
 void hard_fault_handler_c(uint32_t *sp) __attribute__((naked));
@@ -3465,22 +3492,27 @@ void HardFault_Handler(void)
   __asm volatile("mrs %0, psp \n\t" : "=r"(sp));
   hard_fault_handler_c(sp);
 }
+#endif
 
+#ifdef TINYSA4
+void hard_fault_handler_c(uint32_t *sp, const uint32_t *callee)
+#else
 void hard_fault_handler_c(uint32_t *sp)
+#endif
 {
 #ifdef TINYSA4
   uint32_t r0  = sp[0];
   uint32_t r1  = sp[1];
   uint32_t r2  = sp[2];
   uint32_t r3  = sp[3];
-  register uint32_t  r4 __asm("r4");
-  register uint32_t  r5 __asm("r5");
-  register uint32_t  r6 __asm("r6");
-  register uint32_t  r7 __asm("r7");
-  register uint32_t  r8 __asm("r8");
-  register uint32_t  r9 __asm("r9");
-  register uint32_t r10 __asm("r10");
-  register uint32_t r11 __asm("r11");
+  uint32_t r4  = callee[0];
+  uint32_t r5  = callee[1];
+  uint32_t r6  = callee[2];
+  uint32_t r7  = callee[3];
+  uint32_t r8  = callee[4];
+  uint32_t r9  = callee[5];
+  uint32_t r10 = callee[6];
+  uint32_t r11 = callee[7];
   uint32_t r12 = sp[4];
   uint32_t lr  = sp[5];
   uint32_t pc  = sp[6];
@@ -3538,4 +3570,3 @@ void hard_fault_handler_c(uint32_t *sp)
 }
 
 
-