In-game option validation sweep
===============================

Why this exists: the exact-divide field failure proved that a gate over
code that never runs is not a gate. The menu-boot fixture parks on a
static screen and never executes a hot divide; every float-path option
had until then been pixel-gated only there. This sweep re-validates
each option on a gameplay fixture where its path actually fires.

Fixture: Tekken Tag Tournament (USA), maintainer-provided in-game
savestate (RetroArch RASTATE container around RZIP zlib chunks; the
headless harness unwraps both and retro_unserializes the MEM payload
at frame 3). Roughly 154,000 VU1 divide ops per frame; 344 mVU0 and
894 mVU1 add/sub sites compiled, against 62 and 82 on the menu
fixture. SW renderer, FNV-1a per-frame framebuffer row hashes.

Gate per option: rc == 0 across the run; bit-identical hash streams
across paired runs (determinism); engagement proven by divergence
from the default stream or an emission/call counter, never assumed.

Results (aligned-defaults build, 2f4b17f02):

  option                        rc  deterministic  engagement
  ---------------------------------------------------------------
  default (aligned)             0   yes (pair)     n/a; 897 frames
  VU Accurate Add/Sub on        0   yes (pair)     897/897 differ
  VU Exact Mul on               0   yes (pair)     897/897 differ
  VU0 Soft Float on             0   single run     447/447 differ
  VU1 Soft Float on             0   single run     447/447 differ
  EE FPU Soft Float on          0   yes (pair)     284/447 differ
  VU Exact Divide               0   yes (pair)     0 diff vs off; the
                                                   engagement proof is
                                                   the call counter:
                                                   138M calls in 897
                                                   frames, values
                                                   bit-identical to
                                                   the plain divider
                                                   on this content

Elision counters in-game: mVU0 1/344, mVU1 8/894 -- the Fs==Ft
doubling idiom exists in real content at about one percent.

Methodology note: attributing a default-change delta by keying options
back to their old values does not work. A delivered variable that
differs from a static fires the settings-update path, which a static
default never does, so a keyed run is not binary-comparable to a
static-default run even at equal option values. Per-static builds are
the only clean isolation.

Reproduction recipe: PS2_LOADSTATE=<raw MEM payload> on the headless
harness, PCSX2_FBHASH=1, option keys PS2_VU_ADDSUB / PS2_VU_EXACTMUL /
PS2_VU0_SOFT / PS2_VU1_SOFT / PS2_FPU_SOFT / PS2_EE_ACC /
PS2_VU_EXACTDIV, each 0/1-valued.
