benchmark/: three-way RE-tool comparison + first real C-lift

Three small functions extracted from the v1.19 conservative blob with
ground-truth C and per-tool (Ghidra / retdec / decomp.me) docs:
  01_memset        — byte memset, 28 B
  02_memcpy32      — word-aligned memcpy, 36 B
  03_magic_memset  — magic check + tail-call to memset, 40 B
  04_train_phy_block — first real poll-site function (104 B, 26 insts),
                       contains poll sites 12-15

Results in RESULTS.md:
  - Ghidra: A on all four. Auto-decompile is close to final.
  - retdec: A on #3, F on #1 and #2 (no register-arg inference on raw),
    C on #4 (mistakes & 0xF0000000 for < 0x10000000).

GRIND_LOG.md (in 04_train_phy_block/) records the matching-decomp
iteration: 116-byte candidate.c at -Os vs vendor 104 bytes = 89.7%
size match on first real iteration. Remaining gap is GCC's choice of
`cmp w, w_const; b.ls` over vendor's `tst w, #imm; b.eq` for the
mask tests.

gdb_debug/ holds a native-aarch64 GDB single-stepper for the three
benchmark functions — boltzmann smoke test passed (memset:
buf[10] 0x00→0xab).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

benchmark/01_memset/decompme.md

@@ -0,0 +1,58 @@
+# decomp.me recipe — 01_memset
+
+## Create a scratch
+
+Open https://decomp.me/ (or your self-hosted instance at
+http://192.168.88.64 if available). Click **New scratch**.
+
+- **Platform / Compiler:** `gcc 12.x aarch64-linux-gnu` (or whatever
+  aarch64-gcc is offered — exact version doesn't matter much for this
+  size).
+- **Compiler flags:** `-O2 -ffreestanding -nostdlib`
+- **Diff label:** `memset_byte`
+
+## Target asm
+
+Paste the following into the **"Target asm"** box:
+
+```asm
+memset_byte:
+    mov     x3, #0x0
+.Lloop:
+    cmp     x2, x3
+    b.ne    .Lbody
+    ret
+.Lbody:
+    strb    w1, [x0, x3]
+    add     x3, x3, #0x1
+    b       .Lloop
+```
+
+## Context (headers/decls)
+
+```c
+#include <stddef.h>
+#include <stdint.h>
+```
+
+## Source
+
+Paste the ground-truth C from `reference.c` (or write your own first
+and iterate).
+
+## Expected workflow
+
+- First compile: scorer usually reports a high similarity (>= 80%) if
+  the compiler picks the same `while (i != n)` pattern.
+- Fine-tune: try `i++` vs `i+=1`, try `while` vs `for`, try `uint8_t *`
+  cast placement. Each yields a distinct register-allocation order the
+  scorer rewards or punishes.
+- Perfect match possible if you hit the exact code shape GCC chose.
+
+## Benchmark notes
+
+- decomp.me's strength is the **compile-and-diff** feedback loop — every
+  edit immediately shows the byte-delta against the target.
+- Weakness for this target: the real blob was likely built with a
+  different compiler (ARMCC / Keil / vendor LLVM?). GCC may never match
+  exactly even with perfect C. Similarity >= 90% is the realistic ceiling.