tools: add sled (zero-drift) detector + capstone disassembly history in trap handler

docs/sbl-bringup.md

@@ -64,3 +64,83 @@
 Фронтир b18: sbl_main возвращается, новый код `0x801Bxxx`
 (`strb [r7,#2]`-паттерн, b17-NOP на 0x801B71E), падеж pop'ом пустого
 стека у верха IMEM — следующий фронт.
+
+## b18 parity (QEMU == uni, 2026-09-12)
+- QEMU поправлен до паритета: dstub-overlay + vec-maze для Thumb
+  (null-fetch исполняет maze и возвращается) + убран висячий префикс
+  bl (bogus odd store -> alignment fault, DFAR=0x8006B8D).
+- QEMU b18 доходит до того же фронтира, что uni b18 (38961 инстр.):
+  `0x801Bxxx`, r1=0x80522A4, r2/r3=STUBV, lr=0x801B73F, sp марширует
+  к верху IMEM. Смерти чуть различаются по таймингу (uni: read
+  0x8610000; QEMU: null-jump → maze-hang c lr-top 0x80), регион тот же.
+- Следующий фронт (оба): лабиринт `0x801B6xx` (логгер, r7=0/r4=0,
+  pop пустого стека). Кандидаты: S8-стаблет, таблица логгера.
+
+## b21-b23: логгер, S9, S10, rollback (2026-09-12)
+- Логгер `0x8034810` = голый `pop.w 0x28` без push (жрал стек) → S8:
+  вход → `bx lr` (b21). Возврат игнорируется (callers ставят r0=0).
+- Лабиринт `0x801B6A0`: счётчик `[r4]++` + `pop.w` на пустых фреймах
+  (r4/r7=STUBV из fill) → S9 (b22): `[sp+0x14]` ненулевой → исходный
+  pop; ноль → restore r0/r4/r5/sl/fp + **reset sp=0x085FFF00** +
+  прыжок в sbl_main `0x802f69d` через ip (без литералов).
+- Rollback `boot_rollback_version.c:580` (query всегда `0x302E`) → NOP NOP
+  (b22, возврат игнорируется). Query зовёт `*table[0x080528D0]` (нули) →
+  фабрикация: poke STUBV (maze вернёт 0 = версии OK) (b23).
+- Rollback выходил боком в helper-mid (`pop{r4,pc}` без push) → S10 (b23):
+  `movw/movt ip + bx ip` в sbl_main `0x802f6a3`.
+- b23: 2M инстр. без падений (sp здоров), sbl_main идёт дальше
+  (`0x80164C4`-subtree → прыжок в OCIMEM-нули `0x222F84` — фронт PBL-данных).
+  QEMU-паритет: R15=0x223000, sp=0x85FFF00, lr=0x80164CB — там же.
+
+## b24: world-switch stub, clock S11 (2026-09-12)
+- Диспетчер (`0x80164C4`-subtree) зовёт ARM jump-table `0x8007158`
+  (`ldr pc`), та — в OCIMEM-трамплин world-switch `0x221EF8`
+  (ARM: save regs, `cps MON`, `mrc`, `ldr sp/pc` из PBL-таблиц
+  `[0x221F6C]=0x223620`/`[0x221F70]=0x223930` → вызов TZ; TZ не загружен).
+  Вход получается в Thumb → b24-стаб (4B `movs r0,#0; bx lr`):
+  возврат 0 в диспетчер (TZ deferred). Перезаписывает голову ARM'а
+  (восстановимо из ELF).
+- Дальше: clock-fn `0x800E960` (udiv-подсчёты) с голым выходом
+  `pop{r4,pc}` (без push, tail-chain через NOP-идиому) → S11 (b24):
+  вход → `bx lr` (lr = брошенный-но-валидный `0x80076CD`).
+- Poll `[0x73A100]` (lsls#29+bpl = проверка bit2!): SBL только читает →
+  фабрикация bit2=1 (b24 scaffold, обе стороны).
+- `blx #0x8006964` в файловые нули паддинга (PBL scratch) → S12v2 (b25
+  scaffold, ARM! т.к. blx-imm на чётный): preset ip=S16 + ACK bit1
+  `[0x73A100]` (side effect для poll#2) + `bx lr`. Литерал `[0x8006980]`.
+- Helper-pop `0x8007B66` читает scratch (счётчики 0-4 вместо pushed lr,
+  фреймы съедены null-циклами) → S16v2 (b25 smart): B66 (`pop` → `bx ip`,
+  2B) → S16 проверяет ip (odd + top 8 = legit pushed-lr → `bx ip`;
+  иначе default: restore r7, прыжок `0x8007779`). S12v3 грузит ip из
+  `[sp+4]` (pushed lr) + ACK bit1. B66-патч в ELF (b25), S12/S16 в scaffold.
+
+## b26: S17 safe-logger (2026-09-13)
+- S8 (`bx lr`) гасил весь init-лог (UART пуст by design) вслепую.
+  S17 (50B @0x08006CF0): печать r1-строки в UARTDM-TF, только если r1
+  в rodata `[0x804B800,0x8058000)` (wild-указатели скипаются), затем
+  pop-return (без leak'а, возврат игнорируется). Логгер `0x8034810`
+  (`pop.w` 4B) → `b.w S17`. Видимость прогресса вместо счёта инстр.
+
+## b27: poll re-init skip (2026-09-13)
+- Poll-fn `0x8007748`: `cbz r6` (r6=2 константа caller'а, никогда 0) →
+  вечный re-init (`movs r4,#0`, table loop) вместо poll#2-direct.
+  Патч 2B→2B в `b`. Биты 0,1,2 `[0x73A100]` pref-set (scaffold).
+- S16v3: `add sp,#8` перед `bx ip` (снимает helper-фрейм, пропущенный
+  B66 — un-leak стека вниз).
+- S16v4: отбраковка `ip==STUBV` (fill/maze — никогда legit return) →
+  default вместо возврата в maze (разрыв maze-B66 петли).
+- S12v4 (scaffold, без ELF): smart dispatcher в самом S12 (проверка
+  pushed-lr + ACK + возврат) — S16 больше не нужен в цепочке (остаётся
+  как fallback).
+- S12v5: +restore r7 (peripheral base, S16 dormant) прямо в S12.
+
+## b28: poll-epilogue march kill (2026-09-13)
+- Poll-fn входит mid-way (без prologue-push), выходит через
+  `add sp,#0x1C + pop.w 0x14` (+0x30/проход, пушей нет) → марш вверх
+  через fill в IMEM-нули → fault на верху. `add sp` → NOP (2B, b28).
+  Pop читает helper-фреймы (pushed-lr, валидные возвраты).
+
+## b29: S18 poll-epilogue router (2026-09-13)
+- b28-NOP убрал только +0x1C; `pop.w 0x14` без пушей всё равно марширует.
+  S18 (10B @0x08006D22): слепой прыжок в sbl_main `0x802F6A9`
+  (post-dispatcher). Эпилог `pop.w` (4B) → `b.w S18`.

qemu/run-sbl.sh

@@ -5,7 +5,7 @@
 set -e
 cd "$(dirname "$0")/.."
 QEMU=~/qemu-src/qemu-8.2.2/build/qemu-system-arm
-B=${1:-b17}
+B=${1:-b29}
 IMG=fw/img/emmc-real.img
 if [ ! -f "$IMG" ]; then
   echo "no $IMG — сначала: python3 tools/ffu_extract.py fw/dl/RM1127_059X5R5_orig.ffu --emmc-user fw/img/user-real.bin"

qemu/saimaa.c

@@ -104,15 +104,15 @@ static void saimaa_tmr_init(Object *obj)
 /* Stubs region handle for set_readonly after scaffold implant. */
 static MemoryRegion *saimaa_stubs_mr;
 
-/* DDRLOW stub trap [0x08006B80,0x08006C50): serves the scaffolded bytes
- * (S/maze/S6/banner/S7 stubs) on read/fetch, ignores SBL heap-clobber
- * writes. Uni parity: deferred-restore guard in sbl_uni.py (wider:
- * [0x080068C0,0x08006C50), also covers PBL stub).
+/* DDRLOW stub trap [0x08006B80,0x08006D30): serves the scaffolded bytes
+ * (S/maze/S6/banner/S7/S9/S10/S16/S17 stubs) on read/fetch, ignores SBL
+ * heap-clobber writes. Uni parity: deferred-restore guard in sbl_uni.py
+ * (wider: [0x080068C0,0x08006D30), also covers PBL stub).
  * Snapshot is taken AFTER ELF load + all stub writes (readback).
  * PBL stub [0x080068C0,0x080068F4) and the ELF-text gap below stay plain
  * RAM: PBL runs once (if at all), gap holds real hot SBL code. */
 #define SAIMAA_DSTUB_BASE 0x08006B80
-#define SAIMAA_DSTUB_SIZE 0x000000D0
+#define SAIMAA_DSTUB_SIZE 0x000001B0
 static uint8_t saimaa_dstub_mem[SAIMAA_DSTUB_SIZE];
 
 static uint64_t saimaa_dstub_read(void *opaque, hwaddr off, unsigned size)
@@ -167,13 +167,45 @@ struct SaimaaUartState {
 static int saimaa_vec_n;
 static uint64_t saimaa_vec_read(void *opaque, hwaddr off, unsigned size)
 {
-    if (saimaa_vec_n < 400) {
+    /* Maze blob mirrored at [0,18) for THUMB fetches: null-jump fetches
+     * (bx r0=0, pop 0, ldr pc,=0) execute the maze and return via lr —
+     * uni parity (hook_null -> STUB). ARM-mode fetches (real exception
+     * entry) keep BKPT vectors below. Data probes elsewhere read 0. */
+    static const uint8_t maze[18] = {
+        0x00, 0x20, 0x4f, 0xea, 0x1e, 0x6c, 0xbc, 0xf1,
+        0x08, 0x0f, 0x00, 0xd1, 0x70, 0x47, 0x72, 0xb6,
+        0xfd, 0xe7,
+    };
+    {
+        ARMCPU *cpu = ARM_CPU(current_cpu);
+        if (cpu && cpu->env.thumb && off < sizeof(maze)) {
+            uint64_t v = 0;
+            unsigned i, n = size;
+            if (off + n > sizeof(maze)) {
+                n = sizeof(maze) - off;
+            }
+            for (i = 0; i < n; i++) {
+                v |= (uint64_t)maze[off + i] << (8 * i);
+            }
+            return v;
+        }
+    }    if (saimaa_vec_n < 400) {
         ARMCPU *cpu = ARM_CPU(current_cpu);
         uint64_t pc = cpu ? cpu->env.regs[15] : 0;
         int th = cpu ? cpu->env.thumb : 0;
+        uint32_t dfar = 0, ifsr = 0, dfsr = 0;
+        if (cpu) {
+            /* CP15 DFAR/IFSR/DFSR for abort diagnosis */
+            dfar = cpu->env.cp15.dfar_s;
+            ifsr = (uint32_t)cpu->env.cp15.ifsr_s;
+            dfsr = (uint32_t)cpu->env.cp15.dfsr_s;
+            fprintf(stderr, " SCTLR=0x%lx", (unsigned long)cpu->env.cp15.sctlr_s);
+        }
         fprintf(stderr, "saimaa-vec: %s off=0x%x pc=0x%lx",
                 size == 2 ? "fetch?" : "read",
                 (unsigned)off, (unsigned long)pc - (th ? 4 : 8));
+        fprintf(stderr, " DFAR=0x%x IFSR=0x%x DFSR=0x%x", dfar, ifsr,
+                dfsr);
         if (cpu) {
             int b, r;
             fprintf(stderr, " regs=");
@@ -622,6 +654,153 @@ static void saimaa_machine_init(MachineState *machine)
                     cpu_physical_memory_write(0x08006C10 + k, &word, 4);
                 }
             }
+            /* S9 stublet at 0x08006C48 (62B, b22): labyrinth-epilogue
+             * exit router (0x801B6BE pop.w -> b.w S9). Empty frames ->
+             * restore r0/r4/r5/sl/fp, reset sp to 0x085FFF00, jump to
+             * sbl_main 0x802f69d via ip. Assembled with keystone. */
+            {
+                static const uint8_t s9[] = {
+                    0x05, 0x9b, 0x0b, 0xb1, 0xbd, 0xe8, 0xf0, 0x81,
+                    0x00, 0x20, 0x40, 0xf2, 0x00, 0x04, 0xc0, 0xf2,
+                    0x22, 0x04, 0x40, 0xf2, 0x00, 0x05, 0xc0, 0xf2,
+                    0x22, 0x05, 0x40, 0xf2, 0x00, 0x0a, 0xc0, 0xf6,
+                    0x06, 0x0a, 0x41, 0xf2, 0x00, 0x0b, 0xc0, 0xf6,
+                    0x06, 0x0b, 0x4f, 0xf6, 0x00, 0x73, 0xc0, 0xf6,
+                    0x5f, 0x03, 0x9d, 0x46, 0x4f, 0xf2, 0x9d, 0x6c,
+                    0xc0, 0xf6, 0x02, 0x0c, 0x60, 0x47,
+                };
+                size_t k;
+                for (k = 0; k < sizeof(s9); k += 4) {
+                    uint32_t word = 0;
+                    size_t n = sizeof(s9) - k < 4 ? sizeof(s9) - k : 4;
+                    __builtin_memcpy(&word, &s9[k], n);
+                    cpu_physical_memory_write(0x08006C48 + k, &word, 4);
+                }
+            }
+            /* S10 stublet at 0x08006C88 (10B, b23): rollback-exit router.
+             * movw/movt ip + bx ip to sbl_main 0x802f6a3. */
+            {
+                static const uint8_t s10[] = {
+                    0x4f, 0xf2, 0xa3, 0x6c, 0xc0, 0xf6, 0x02, 0x0c,
+                    0x60, 0x47,
+                };
+                size_t k;
+                for (k = 0; k < sizeof(s10); k += 4) {
+                    uint32_t word = 0;
+                    size_t n = sizeof(s10) - k < 4 ? sizeof(s10) - k : 4;
+                    __builtin_memcpy(&word, &s10[k], n);
+                    cpu_physical_memory_write(0x08006C88 + k, &word, 4);
+                }
+            }
+            /* S18 stublet at 0x08006D22 (10B, b29): poll-epilogue router.
+             * movw/movt ip + bx ip to sbl_main 0x802f6a9. */
+            {
+                static const uint8_t s18[] = {
+                    0x4f, 0xf2, 0xa9, 0x6c, 0xc0, 0xf6, 0x02, 0x0c,
+                    0x60, 0x47,
+                };
+                size_t k;
+                for (k = 0; k < sizeof(s18); k += 4) {
+                    uint32_t word = 0;
+                    size_t n = sizeof(s18) - k < 4 ? sizeof(s18) - k : 4;
+                    __builtin_memcpy(&word, &s18[k], n);
+                    cpu_physical_memory_write(0x08006D22 + k, &word, 4);
+                }
+            }
+            /* S17 safe-logger (50B @0x08006CF0, b26): print r1-string
+             * via UARTDM-TF iff r1 in rodata [0x804B800,0x8058000);
+             * pop-return (no leak). Replaces S8 silence with logs. */
+            {
+                static const uint8_t s17[] = {
+                    0x4b, 0xf6, 0x00, 0x03, 0xc0, 0xf6, 0x04, 0x03,
+                    0x99, 0x42, 0x11, 0xd3, 0x48, 0xf2, 0x00, 0x03,
+                    0xc0, 0xf6, 0x05, 0x03, 0x99, 0x42, 0x0b, 0xd2,
+                    0x10, 0xb5, 0x4f, 0xf2, 0x00, 0x03, 0xc0, 0xf2,
+                    0x8a, 0x73, 0x0c, 0x46, 0x14, 0xf8, 0x01, 0x0b,
+                    0x08, 0xb1, 0x18, 0x67, 0xfa, 0xe7, 0x10, 0xbd,
+                    0x70, 0x47,
+                };
+                size_t k;
+                for (k = 0; k < sizeof(s17); k += 4) {
+                    uint32_t word = 0;
+                    size_t n = sizeof(s17) - k < 4 ? sizeof(s17) - k : 4;
+                    __builtin_memcpy(&word, &s17[k], n);
+                    cpu_physical_memory_write(0x08006CF0 + k, &word, 4);
+                }
+            }
+            /* PBL version-table fabrication (b23): [0x080528D0] = STUBV
+             * so the rollback query's blx calls maze (0 = versions OK). */
+            {
+                uint32_t v = 0x08006B8D;
+                cpu_physical_memory_write(0x080528D0, &v, 4);
+            }
+            /* HW-ready bits (b25/b26): [0x073A100] bits 0,1,2 (SBL
+             * read-only polls: bit2 poll#1, bit1 poll#2, bit0 gate). */
+            {
+                uint32_t v = 0x00000007;
+                cpu_physical_memory_write(0x073A100, &v, 4);
+            }
+            /* PBL world-switch Thumb-stub (b24, 4B @0x00221EF8):
+             * movs r0,#0; bx lr (return 0 to dispatcher, TZ deferred).
+             * Overwrites ARM trampoline head (recoverable from ELF). */
+            {
+                uint32_t v = 0x47702000;
+                cpu_physical_memory_write(0x00221EF8, &v, 4);
+            }
+             /* S12-stub (b27, 92B @0x08006964, ARM): blx-to-padding
+             * becomes smart dispatcher: ldr ip,[sp,#4] (=pushed lr);
+             * ip==STUBV/even/non-code => ip=poll resumption 0x8007779;
+             * else keep pushed-lr; restore r7 (peripheral base);
+             * ACK peripheral bit1 ([literal 0x073A100]); bx lr.
+             * PBL would place the real ARM helper here. */
+            {
+                static const uint8_t s12[] = {
+                    0x04, 0xc0, 0x9d, 0xe5, 0x8d, 0x3b, 0x06, 0xe3,
+                    0x00, 0x38, 0x40, 0xe3, 0x03, 0x00, 0x5c, 0xe1,
+                    0x07, 0x00, 0x00, 0x0a, 0x01, 0x00, 0x1c, 0xe3,
+                    0x05, 0x00, 0x00, 0x0a, 0x2c, 0x3c, 0xa0, 0xe1,
+                    0x08, 0x00, 0x53, 0xe3, 0x04, 0x00, 0x00, 0x0a,
+                    0x2c, 0x3e, 0xa0, 0xe1, 0x08, 0x00, 0x53, 0xe3,
+                    0x01, 0x00, 0x00, 0x0a, 0x79, 0xc7, 0x07, 0xe3,
+                    0x00, 0xc8, 0x40, 0xe3, 0x00, 0x70, 0x0a, 0xe3,
+                    0x73, 0x70, 0x40, 0xe3, 0x0c, 0x30, 0x9f, 0xe5,
+                    0x00, 0x20, 0x93, 0xe5, 0x02, 0x20, 0x82, 0xe3,
+                    0x00, 0x20, 0x83, 0xe5, 0x1e, 0xff, 0x2f, 0xe1,
+                    0x00, 0xa1, 0x73, 0x00,
+                };
+                size_t k;
+                for (k = 0; k < sizeof(s12); k += 4) {
+                    uint32_t word = 0;
+                    size_t n = sizeof(s12) - k < 4 ? sizeof(s12) - k : 4;
+                    __builtin_memcpy(&word, &s12[k], n);
+                    cpu_physical_memory_write(0x08006964 + k, &word, 4);
+                }
+            }
+            /* S16-stublet (b27, 64B @0x08006CC0): smart return for the
+             * B66 bx-ip chain. ip==STUBV (fill/maze, never legit) =>
+             * default; ip odd + top 8 => add sp,#8 (pop the helper
+             * frame B66 skipped!) + bx ip (legit pushed-lr, un-leaks);
+             * else default (restore r7, jump poll resumption 0x8007779
+             * via ip). */
+            {
+                static const uint8_t s16[] = {
+                    0xdd, 0xf8, 0x04, 0xc0, 0x46, 0xf6, 0x8d, 0x33,
+                    0xc0, 0xf6, 0x00, 0x03, 0x9c, 0x45, 0x0e, 0xd0,
+                    0x1c, 0xf0, 0x01, 0x0f, 0x0b, 0xd0, 0x5f, 0xea,
+                    0x1c, 0x63, 0x08, 0x2b, 0x01, 0xd1, 0x02, 0xb0,
+                    0x60, 0x47, 0x5f, 0xea, 0x1c, 0x73, 0x08, 0x2b,
+                    0x01, 0xd1, 0x02, 0xb0, 0x60, 0x47, 0x4a, 0xf2,
+                    0x00, 0x07, 0xc0, 0xf2, 0x73, 0x07, 0x47, 0xf2,
+                    0x79, 0x7c, 0xc0, 0xf6, 0x00, 0x0c, 0x60, 0x47,
+                };
+                size_t k;
+                for (k = 0; k < sizeof(s16); k += 4) {
+                    uint32_t word = 0;
+                    size_t n = sizeof(s16) - k < 4 ? sizeof(s16) - k : 4;
+                    __builtin_memcpy(&word, &s16[k], n);
+                    cpu_physical_memory_write(0x08006CC0 + k, &word, 4);
+                }
+            }
             {
                 uint32_t fill = 0x08006B8D;
                 uint32_t a;
@@ -629,6 +808,15 @@ static void saimaa_machine_init(MachineState *machine)
                     cpu_physical_memory_write(a, &fill, 4);
                 }
             }
+            /* IMEM poison [0x8600000,0x8600F00): drift-sink trap (UDF in
+             * both modes). IMEM stack [0x8600F00,0x8610000) intact. */
+            {
+                uint32_t p = 0xFFFFFFFF;
+                uint32_t a;
+                for (a = 0x08600000; a < 0x08600F00; a += 4) {
+                    cpu_physical_memory_write(a, &p, 4);
+                }
+            }
             /* pop-site slots -> stublets (sp values from cpu traces) */
             v = 0x802f65d; /* S1 pop10 @fef0: resume sbl_main */
             cpu_physical_memory_write(0x085FFF14, &v, 4);

tools/sbl_patch.py

@@ -528,6 +528,9 @@ def main():
     # else restore r4/r5/sl/fp and return to sbl_main (0x802f681).
     # NOTE: overwrites BAE-path head (ldr r2,[pc,#0x28], only used when
     # [r7]==2 — never in our flow, r7=0).
+    # Also NOPs the FULL bl idiom at 0x8006682 (b17 NOPed only its 2nd
+    # half, leaving a dangling prefix that fused with the NOP into a
+    # bogus odd-address store -> alignment fault in QEMU).
     d = bytearray(base)
 
     for va, (a, b) in BLX2BL.items():
@@ -550,8 +553,6 @@ def main():
 
         (0x801B76C, '280b0508', '00000708'),
 
-        (0x8006684, '01f807bd', '46c046c0'),
-
         (0x8017AFC, 'bde8fc87', 'eef74fb8'),
 
         (0x802F690, '9847fee7', 'c046c046'),
@@ -564,6 +565,15 @@ def main():
 
         set4(d, va, new)
 
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs. (b17 NOPed only
+    # its 2nd half, leaving a dangling prefix that fused with the NOP
+    # into a bogus odd-address store -> alignment fault in QEMU.)
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
     o = off(0x8013AA8)
 
     assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
@@ -651,6 +661,748 @@ def main():
 
     save(d, 'img0_b20.elf')
 
+    # b21 = b18 + logger S8 (0x8034810 pop.w -> bx lr):
+    # the logger is a naked pop.w {r2-r8,sb,sl,pc} (0x28) with no push;
+    # called via bl from scaffold-frame workers, it eats 0x28 of live
+    # stack per call (slots->fill->zeros->null-loop->IMEM overflow).
+    # S8 turns it into a no-op return (bx lr, 2B; tail bytes unreachable):
+    # return value ignored (callers set r0=0 after), workers unwind via
+    # their own epilogue pops. Logging already empty (UART stubs).
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Logger S8: pop.w entry (4B bde8fc87) -> bx lr (first 2B).
+    o = off(0x8034810)
+
+    assert d[o:o + 4].hex() == 'bde8fc87', hx(d, 0x8034810)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    save(d, 'img0_b21.elf')
+
+    # b22 = b21 + rollback assert NOP (0x801671C blx + b.n self):
+    # boot_rollback_version.c:580 (code 0x302E) fires (versions zeroed);
+    # caller (sbl_main) ignores r0 (movs r0,#0 after), so fall-through
+    # to the success pop is safe. Same idiom as banner site.
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Logger S8 (see b21).
+    o = off(0x8034810)
+
+    assert d[o:o + 4].hex() == 'bde8fc87', hx(d, 0x8034810)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    save(d, 'img0_b22.elf')
+
+    # b23 = b22 + S10 (rollback sideways-exit fix):
+    # 0x8016720 pop{r4,pc} has no matching push; it exits sideways to
+    # helper-mid (0x8005BBA) instead of sbl_main (0x802f6a3), closing a
+    # ping-pong cycle (S9->rollback->query->helper->assert->labyrinth).
+    # Route to S10 (0x08006C88, scaffold): movw/movt ip + bx ip to
+    # 0x802f6a3. Eats the following NOP (movs r0,r0).
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Logger S8 (see b21).
+    o = off(0x8034810)
+
+    assert d[o:o + 4].hex() == 'bde8fc87', hx(d, 0x8034810)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    save(d, 'img0_b23.elf')
+
+    # b24 = b23 + S11 (clock-fn naked-exit fix):
+    # 0x800E988 pop{r4,pc} has no matching push (naked tail-chain worker:
+    # bl 0x8006684-idiom tail-jumps to the clock fn, abandoning lr).
+    # Each dispatcher pass eats 0x8+ ([sp] -> wild 0x902FC506).
+    # S11 turns the exit into bx lr (lr = abandoned-but-valid resumption
+    # 0x80076CD/0x80076D3 in the caller). 2B patch, like S8.
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Logger S8 (see b21).
+    o = off(0x8034810)
+
+    assert d[o:o + 4].hex() == 'bde8fc87', hx(d, 0x8034810)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    # Clock S11: pop{r4,pc} (2B 10bd) -> bx lr (2B 7047).
+    o = off(0x800E988)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x800E988, 2)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    save(d, 'img0_b24.elf')
+
+    # b25 = b24 + helper-pop S16 chain (0x8007B66 pop{r4,pc} -> bx ip):
+    # the pop reads scratch ([sp+4] = counters 0..4, frames marched by
+    # null-cycles) instead of pushed lr. bx ip uses ip preset by S12
+    # (ARM stub at 0x8006964) to S16 (0x08006CC0, scaffold): restore r7
+    # (peripheral base) and jump to poll-loop resumption 0x8007779.
+    # 2B patch, no eating (bx ip = 0x47c0 fits the pop slot).
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Logger S8 (see b21).
+    o = off(0x8034810)
+
+    assert d[o:o + 4].hex() == 'bde8fc87', hx(d, 0x8034810)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    # Clock S11 (see b24).
+    o = off(0x800E988)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x800E988, 2)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    # Helper-pop S16 chain: pop{r4,pc} (2B 10bd) -> bx ip (2B 4760;
+    # ip preset by S12 to S16). 0x47c0 would be bx lr (wrong reg!).
+    o = off(0x8007B66)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x8007B66, 2)
+
+    d[o:o + 2] = bytes.fromhex('6047')
+
+    save(d, 'img0_b25.elf')
+
+    # b26 = b25 + S17 safe-logger (0x8034810 pop.w -> b.w S17):
+    # S8 (bx lr) silenced all init logging (UART empty by design).
+    # S17 (50B @0x08006CF0, scaffold): print r1-string via UARTDM-TF
+    # only if r1 in rodata [0x804B800,0x8058000) (wild pointers skip),
+    # then pop-return (no leak, return value ignored by callers).
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+        (0x8034810, 'bde8fc87', 'd2f76eba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Helper-pop S16 chain (see b25).
+    o = off(0x8007B66)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x8007B66, 2)
+
+    d[o:o + 2] = bytes.fromhex('6047')
+
+    # Clock S11 (see b24).
+    o = off(0x800E988)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x800E988, 2)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    save(d, 'img0_b26.elf')
+
+    # b27 = b26 + poll re-init skip (0x8007748 cbz r6 -> b):
+    # r6 (table mode, caller-passed constant 2) never 0, so the poll-fn
+    # re-inits (movs r4,#0) and re-runs its table loop forever instead of
+    # proceeding to poll#2-direct + epilogue. Force direct (2B->2B).
+    # Bits 0,1,2 of [0x73A100] pre-set (scaffold), so polls exit.
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+        (0x8034810, 'bde8fc87', 'd2f76eba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Poll re-init skip (b27): cbz r6 (2B 96b1) -> b (2B 14e0).
+    o = off(0x8007748)
+
+    assert d[o:o + 2].hex() == '96b1', hx(d, 0x8007748, 2)
+
+    d[o:o + 2] = bytes.fromhex('14e0')
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Helper-pop S16 chain (see b25).
+    o = off(0x8007B66)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x8007B66, 2)
+
+    d[o:o + 2] = bytes.fromhex('6047')
+
+    # Clock S11 (see b24).
+    o = off(0x800E988)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x800E988, 2)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    save(d, 'img0_b27.elf')
+
+    # b28 = b27 + poll-epilogue march kill (0x8007794 add sp,#0x1c NOP):
+    # the poll-fn is entered mid-way (no prologue push!) but exits via
+    # add sp,#0x1c + pop.w 0x14 (+0x30/pass, no pushes!) -> marches sp
+    # UP through fill into IMEM zeros -> top fault. NOP the add (2B);
+    # the pop then reads helper frames (pushed-lrs, valid returns).
+    # Slow helper-leak-down (0x8/call, B66 skips pop) remains, harmless.
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+        (0x8034810, 'bde8fc87', 'd2f76eba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Helper-pop S16 chain (see b25).
+    o = off(0x8007B66)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x8007B66, 2)
+
+    d[o:o + 2] = bytes.fromhex('6047')
+
+    # Clock S11 (see b24).
+    o = off(0x800E988)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x800E988, 2)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    # Poll-epilogue march kill: add sp,#0x1c (2B 07b0) -> NOP (2B c046).
+    o = off(0x8007794)
+
+    assert d[o:o + 2].hex() == '07b0', hx(d, 0x8007794, 2)
+
+    d[o:o + 2] = bytes.fromhex('c046')
+
+    save(d, 'img0_b28.elf')
+
+    # b29 = b28 + S18 (poll-epilogue pop.w -> b.w S18):
+    # the b28 NOP killed +0x1c but pop.w 0x14 still marches (no pushes,
+    # mid-entries!). S18 (10B @0x08006D22, scaffold): movw/movt ip +
+    # bx ip to sbl_main 0x802f6a9 (post-dispatcher, dispatcher work
+    # done-ish via helpers). Blind jump like S10.
+    d = bytearray(base)
+
+    for va, (a, b) in BLX2BL.items():
+
+        set4(d, va, b)
+
+    for va, exp, new in [
+
+        (0x802F664, 'e4f77efc', 'c046c046'),
+
+        (0x802F672, 'e6f7bbfa', 'c046c046'),
+
+        (0x802F678, 'e9f7fef9', 'c046c046'),
+
+        (0x801B71E, '9847fee7', 'c046c046'),
+
+        (0x8008396, 'fef780f9', 'c046c046'),
+
+        (0x8016184, 'fdf790fc', 'c046c046'),
+
+        (0x801B76C, '280b0508', '00000708'),
+
+        (0x8017AFC, 'bde8fc87', 'eef74fb8'),
+
+        (0x802F690, '9847fee7', 'c046c046'),
+
+        (0x801671C, '9847fee7', 'c046c046'),
+
+        (0x8013BAC, 'f8bd0a4a', 'f3f730b8'),
+
+        (0x801B6BE, 'bde8f081', 'ebf7c3ba'),
+
+        (0x8016720, '10bd0000', 'f0f7b2ba'),
+
+        (0x8034810, 'bde8fc87', 'd2f76eba'),
+
+        (0x8007798, 'bde8f08f', 'fff7c3ba'),
+
+    ]:
+
+        assert hx(d, va) == exp, (hex(va), hx(d, va))
+
+        set4(d, va, new)
+
+    # Full bl idiom at 0x8006682 (bl + pop, 6B) -> NOPs (see b18).
+    o = off(0x8006682)
+
+    assert d[o:o + 6].hex() == '00f001f807bd', hx(d, 0x8006682, 6)
+
+    d[o:o + 6] = bytes.fromhex('c046c046c046')
+
+    o = off(0x8013AA8)
+
+    assert d[o:o + 4].hex() == '07f026be', hx(d, 0x8013AA8)
+
+    d[o:o + 4] = bytes.fromhex('c046c046')
+
+    o = off(0x8013BA6)
+
+    assert d[o:o + 2].hex() == 'b888', (hex(0x8013BA6), hx(d, 0x8013BA6, 2))
+
+    d[o:o + 2] = bytes.fromhex('ff20')
+
+    # Helper-pop S16 chain (see b25).
+    o = off(0x8007B66)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x8007B66, 2)
+
+    d[o:o + 2] = bytes.fromhex('6047')
+
+    # Clock S11 (see b24).
+    o = off(0x800E988)
+
+    assert d[o:o + 2].hex() == '10bd', hx(d, 0x800E988, 2)
+
+    d[o:o + 2] = bytes.fromhex('7047')
+
+    # Poll re-init skip (see b27).
+    o = off(0x8007748)
+
+    assert d[o:o + 2].hex() == '96b1', hx(d, 0x8007748, 2)
+
+    d[o:o + 2] = bytes.fromhex('14e0')
+
+    # Poll-epilogue add NOP (see b28).
+    o = off(0x8007794)
+
+    assert d[o:o + 2].hex() == '07b0', hx(d, 0x8007794, 2)
+
+    d[o:o + 2] = bytes.fromhex('c046')
+
+    save(d, 'img0_b29.elf')
+
 
 
 

tools/sbl_uni.py

@@ -101,9 +101,28 @@ def main():
     last = [0]
     from collections import deque
     hist = deque(maxlen=20000)
+    # sled detector: sequential runs (pc == prev+2/4) longer than THRESH
+    # are zero-drifts; log run start (the jump/fall that entered it).
+    _sled_start = [0]
+    _sled_n = [0]
+    _sled_log = []
+    _SLED_TH = 100
     calls = [] if "--calltrace" in sys.argv else None
     rets = [] if "--retrace" in sys.argv else None
     qlog = [] if "--querylog" in sys.argv else None
+    # frozen-maze breaker: maze entries with identical lr and no bl
+    # between them mean a zero-state-change spin (maze->mid-fn->maze).
+    # After K, redirect once to poll resumption (fresh bls unfreeze lr).
+    _fmlr = [None]
+    _fmct = [0]
+    _FMK = 50
+    _FMPC = 0x8007779
+    _traps = []
+    _trap_n = [0]
+    if "--trap" in sys.argv:
+        _traps = [int(x, 0) for x in sys.argv[sys.argv.index("--trap") + 1].split(",")]
+    if "--trapn" in sys.argv:
+        _trap_n = [int(sys.argv[sys.argv.index("--trapn") + 1])]
     _spmax = [0x085FFF00]
     _splog = [] if "--spytrack" in sys.argv else None
     elog = [] if "--entrylog" in sys.argv else None
@@ -121,6 +140,35 @@ def main():
     def hook_code(mu, addr, size, data):
         count[0] += 1
         last[0] = addr
+        _pp = _prev[0]
+        _prev[0] = addr
+        if addr == _pp + 2 or addr == _pp + 4:
+            if _sled_n[0] == 0:
+                _sled_start[0] = _pp
+            _sled_n[0] += 1
+        else:
+            if _sled_n[0] >= _SLED_TH and len(_sled_log) < 20:
+                _sled_log.append((count[0], _sled_start[0], _pp, _sled_n[0]))
+            _sled_n[0] = 0
+        if 0x08006B8C <= addr < 0x08006B9E:
+            try:
+                _lr = mu.reg_read(_LR)
+            except Exception:
+                _lr = None
+            if _lr == _fmlr[0]:
+                _fmct[0] += 1
+                if _fmct[0] >= _FMK:
+                    _fmct[0] = 0
+                    _fmlr[0] = None
+                    mu.reg_write(_PC, _FMPC)
+                    return
+            else:
+                _fmlr[0] = _lr
+                _fmct[0] = 0
+        elif _fmlr[0] is not None:
+            # any non-maze fetch: reset only on bl (lr change means
+            # fresh calls; other insns keep the freeze assessment)
+            pass
         if _splog is not None:
             try:
                 sp = mu.reg_read(UC_ARM_REG_SP)
@@ -145,6 +193,38 @@ def main():
         except Exception:
             thumb = 1
         hist.append((addr, thumb))
+        if _traps and addr in _traps:
+            try:
+                vals = [mu.reg_read(_REGS[i]) for i in range(8)]
+            except Exception:
+                vals = []
+            print(f"TRAP {addr:#x} insn#{count[0]} r0-r7={[hex(x) for x in vals]} sp={mu.reg_read(UC_ARM_REG_SP):#x} lr={mu.reg_read(_LR):#x}")
+            try:
+                print(f"  [73A100]={bytes(mu.mem_read(0x073A100, 4)).hex()}")
+            except Exception as e:
+                print(f"  [73A100] unreadable: {e}")
+            try:
+                from capstone import Cs, CS_ARCH_ARM, CS_MODE_THUMB, CS_MODE_ARM
+                md_t = Cs(CS_ARCH_ARM, CS_MODE_THUMB)
+                md_a = Cs(CS_ARCH_ARM, CS_MODE_ARM)
+                for _pc, _th in list(hist)[-60:]:
+                    try:
+                        _b = _pc & ~1
+                        _c = mu.mem_read(_b, 8 if _th else 12)
+                        for _ins in (md_t if _th else md_a).disasm(_c, _b):
+                            print(f"  hist {_b:#x}{'t' if _th else 'a'}: {_ins.mnemonic} {_ins.op_str}")
+                            break
+                    except Exception:
+                        print(f"  hist {_pc:#x}: <unreadable>")
+            except ImportError:
+                pass
+            _trap_n[0] -= 1
+            if _trap_n[0] <= 0 and "--trapn" in sys.argv:
+                mu.emu_stop()
+                return
+            if "--trapn" not in sys.argv:
+                mu.emu_stop()
+                return
         if qlog is not None and addr in (0x8017EA4, 0x8017C60):
             try:
                 vals = [mu.reg_read(_REGS[i]) for i in range(8)]
@@ -158,9 +238,15 @@ def main():
                 inside = _elo <= addr < _ehi
                 if inside and not _inrng[_i]:
                     if len(elog) < 20000:
-                        elog.append((count[0], _prev[0], addr))
+                        try:
+                            _th = (mu.reg_read(_CPSR) >> 5) & 1
+                        except Exception:
+                            _th = -1
+                        elog.append((count[0], _prev[0], addr, _th))
+                if not inside and _inrng[_i]:
+                    if len(elog) < 20000:
+                        elog.append((count[0], _prev[0], -addr, 2))
                 _inrng[_i] = inside
-            _prev[0] = addr
         if calls is not None and size in (2, 4):
             try:
                 if size == 2:
@@ -182,6 +268,15 @@ def main():
                             bxt = -1
                         if len(calls) < 8000:
                             calls.append((addr, 100 + bxrm, bxt))
+                    # pop {..,pc} (0xBD00+): log stacked target
+                    if calls is not None and (h & 0xFF00) == 0xBD00:
+                        try:
+                            _sp = mu.reg_read(UC_ARM_REG_SP)
+                            _tgt = struct.unpack("<I", mu.mem_read(_sp, 4))[0]
+                        except Exception:
+                            _tgt = -1
+                        if len(calls) < 12000:
+                            calls.append((addr, 300, _tgt))
                     if rets is not None and 0x8017E00 <= addr < 0x8018000:
                         # pop {..,pc} (0xBD00+) or bx lr (0x4770)
                         if h == 0x4770 or (h & 0xFF00) == 0xBD00:
@@ -236,6 +331,7 @@ def main():
     # immune to [0]-writes that clobber a static maze copy
     from unicorn.arm_const import UC_ARM_REG_PC as _PC
     _nulllog = []
+    _halt_on_null = "--haltonnull" in sys.argv
 
     def hook_null(mu, addr, size, data):
         if len(_nulllog) < 10:
@@ -247,6 +343,8 @@ def main():
                 r, lr, sp = [], -1, -1
             # walk back: find the branch source from hist
             _nulllog.append((count[0], r, lr, sp))
+        if _halt_on_null and len(_nulllog) == 1:
+            return False
         mu.reg_write(_PC, STUB + 1)
 
     mu.hook_add(UC_HOOK_CODE, hook_null, begin=0x0, end=0xFFF)
@@ -313,8 +411,66 @@ def main():
         "ff2c13d940f20004c0f2220440f20005c0f2220540f2000a"
         "c0f6060a41f2000bc0f6060bdff800f081f60208dff800f0"
         "a53b0108"))
+    # S9 stublet (62B @0x08006C48, b22): labyrinth-epilogue exit router.
+    # ldr r3,[sp,#0x14]: nonzero -> original pop.w (legit return);
+    # zero (empty scaffold frames) -> restore r0/r4/r5/sl/fp, reset sp
+    # to 0x085FFF00 (fresh STUBV-fill stack), jump to sbl_main 0x802f69d
+    # via ip (no literal loads).
+    mu.mem_write(0x08006C48, bytes.fromhex(
+        "059b0bb1bde8f081002040f20004c0f2220440f20005c0f2"
+        "220540f2000ac0f6060a41f2000bc0f6060b4ff60073c0f6"
+        "5f039d464ff29d6cc0f6020c6047"))
+    # S10 stublet (10B @0x08006C88, b23): rollback-exit router.
+    # movw/movt ip + bx ip to sbl_main 0x802f6a3.
+    mu.mem_write(0x08006C88, bytes.fromhex("4ff2a36cc0f6020c6047"))
+    # S18 stublet (10B @0x08006D22, b29): poll-epilogue router.
+    # movw/movt ip + bx ip to sbl_main 0x802f6a9 (post-dispatcher).
+    mu.mem_write(0x08006D22, bytes.fromhex("4ff2a96cc0f6020c6047"))
+    # S17 safe-logger (50B @0x08006CF0, b26): print r1-string via
+    # UARTDM-TF only if r1 in rodata [0x804B800,0x8058000); pop-return.
+    mu.mem_write(0x08006CF0, bytes.fromhex(
+        "4bf60003c0f60403994211d348f20003c0f605039942"
+        "0bd210b54ff20003c0f28a730c4614f8010b08b11867fae710bd7047"))
+    # PBL version-table fabrication (b23): [0x080528D0] = STUBV so the
+    # rollback query's blx calls maze (returns 0 = versions OK).
+    W32(0x080528D0, STUBV)
+    # HW-ready bit fabrication (b25): [0x073A100] bits 0,1,2 = peripheral
+    # status done-bits (poll#1 checks bit2 (lsls#29), poll#2 bit1 (lsls#30),
+    # post-poll gate bit0 (lsls#31)). SBL only reads it, never writes.
+    W32(0x073A100, 0x00000007)
+    # PBL world-switch Thumb-stub (b24): the ARM trampoline at 0x221EF8
+    # is entered in Thumb (mode confusion at dispatch); real ARM would
+    # save regs/CPS to MON and call TZ (not loaded yet). Stub returns 0
+    # to the dispatcher via stale lr (TZ bringup deferred).
+    # NOTE: overwrites 4B of real ARM trampoline (recoverable from ELF).
+    mu.mem_write(0x00221EF8, bytes([0x00, 0x20, 0x70, 0x47]))
+    # S12-stub (b27, 92B @0x08006964, ARM): blx-to-padding becomes
+    # smart dispatcher: ldr ip,[sp,#4] (=pushed lr); ip==STUBV/even/
+    # non-code => ip=poll resumption 0x8007779; else keep pushed-lr;
+    # restore r7 (peripheral base, S16 dormant); ACK peripheral bit1
+    # ([literal 0x073A100]); bx lr. PBL would place the real helper.
+    mu.mem_write(0x08006964, bytes.fromhex(
+        "04c09de58d3b06e3003840e303005ce10700000a0100"
+        "1ce30500000a2c3ca0e1080053e30400000a2c3ea0e1"
+        "080053e30100000a79c707e300c840e300700ae37370"
+        "40e30c309fe5002093e5022082e3002083e51eff2fe1"
+        "00a17300"))
+    # S16-stublet (b27, 64B @0x08006CC0): smart return for B66 bx ip.
+    # ip==STUBV (fill/maze, never a legit return) => default;
+    # ip odd + top 8 (code) => add sp,#8 (pop the helper frame B66
+    # skipped!) + bx ip (legit pushed-lr return, un-leaks the stack);
+    # else default (restore r7, jump poll resumption 0x8007779).
+    mu.mem_write(0x08006CC0, bytes.fromhex(
+        "ddf804c046f68d33c0f600039c450ed01cf0010f0bd0"
+        "5fea1c63082b01d102b060475fea1c73082b01d102b0"
+        "60474af20007c0f2730747f2797cc0f6000c6047"))
     # stack zone fill (STUBV odd)
     mu.mem_write(0x085F0000, struct.pack("<I", STUBV) * (0x10000 // 4))
+    # IMEM poison [0x8600000,0x8600F00): drift-sink trap. Fill/IMEM-zero
+    # sleds (fall-through ldr/movs/andeq) drift silently to IMEM-top
+    # fault; poison faults LOUDLY at drift-entry (UDF in both modes).
+    # IMEM stack [0x8600F00,0x8610000) left intact.
+    mu.mem_write(0x08600000, b"\xff\xff\xff\xff" * (0xF00 // 4))
     # pop-site slots (sp values from traces)
     for a, v in [(0x085FFF14, 0x802f65d), (0x085FFF1C, 0x802f665),
                  (0x085FFF44, 0x8013ba5), (0x085FFF5C, 0x802f681),
@@ -349,7 +505,7 @@ def main():
     # --- end parity ---
     # stub range is readonly on HW/QEMU: ignore SBL heap-clobber writes,
     # keep maze/stublets intact (restore-on-write with reentrancy guard)
-    STUB_LO, STUB_HI = 0x080068C0, 0x08006C50  # +S7 + banner tail
+    STUB_LO, STUB_HI = 0x080068C0, 0x08006D30  # +S7/S9/S10/S16/S17
     stub_snap = bytes(mu.mem_read(STUB_LO, STUB_HI - STUB_LO))
     _inww = [False]
     _swn = [0]
@@ -417,7 +573,7 @@ def main():
                 wlolog.append((pc, addr, size, value))
 
         mu.hook_add(UC_HOOK_MEM_WRITE, hook_wlo,
-                    begin=0x08000000, end=0x085FFFFF)
+                    begin=0x00200000, end=0x085FFFFF)
         import atexit as _ax2
 
         def _dwlo():
@@ -459,10 +615,24 @@ def main():
                 print(f"  {pc:#x}: <unreadable>")
     except ImportError:
         print("hist:", [hex(pc) for pc, _ in list(hist)[-12:]])
+    if _sled_log or _sled_n[0] >= _SLED_TH:
+        if _sled_n[0] >= _SLED_TH:
+            _sled_log.append((count[0], _sled_start[0], last[0], _sled_n[0]))
+        print(f"sleds: {len(_sled_log)}")
+        for n, s, e, c in _sled_log:
+            print(f"  insn#{n} [{s:#x}..{e:#x}] len={c}")
     if elog is not None:
         print(f"entries: {len(elog)}")
-        for n, prev, cur in elog[-30:]:
-            print(f"  insn#{n} {prev:#x} -> {cur:#x}")
+        for n, prev, cur, th in elog[:10]:
+            if cur < 0:
+                print(f"  FIRST EXIT insn#{n} {prev:#x} -> {(-cur):#x}")
+            else:
+                print(f"  FIRST insn#{n} {prev:#x} -> {cur:#x}{'t' if th == 1 else ('a' if th == 0 else '?')}")
+        for n, prev, cur, th in elog[-10:]:
+            if cur < 0:
+                print(f"  LAST EXIT insn#{n} {prev:#x} -> {(-cur):#x}")
+            else:
+                print(f"  LAST insn#{n} {prev:#x} -> {cur:#x}{'t' if th == 1 else ('a' if th == 0 else '?')}")
     if _splog is not None:
         print(f"sp-highs: {len(_splog)}")
         for n, pc, sp in _splog:
@@ -485,6 +655,12 @@ def main():
         print(f"bl-immed: {len(bls)}")
         for pc, tgt in bls[-30:]:
             print(f"  {pc:#x}: bl #{tgt:#x}")
+        pops = [(pc, tgt) for pc, rm, tgt in calls if rm == 300]
+        print(f"pop-pc: {len(pops)}")
+        for pc, tgt in pops[:15]:
+            print(f"  FIRST {pc:#x}: pop-> {tgt:#x}" if tgt >= 0 else f"  FIRST {pc:#x}: pop-> ?")
+        for pc, tgt in pops[-15:]:
+            print(f"  LAST {pc:#x}: pop-> {tgt:#x}" if tgt >= 0 else f"  LAST {pc:#x}: pop-> ?")
         for pc, rm, tgt in calls[-12:]:
             if rm == 200:
                 continue