Exact-multiply accuracy closure vs the interpreter oracle
=========================================================
Rig: per-op differential harness, 400 cases per op, 12% special mix,
REG_I and latches seeded. Columns: value lanes /1600 (ACC lanes for
A-forms), status /400, MAC /400.

op        default          exactmul         exactmul+addsub
MUL       284/190/190       28/190/190       28/190/190
MULx      299/147/147       22/147/147       22/147/147
MULi      328/158/158       47/158/158       47/158/158
MULq      233/121/121       15/121/121       15/121/121
MADD      709/195/204      443/195/195       16/195/195
MSUB      738/205/214      706/204/204       16/204/204
OPMSUB    523/152/158      500/149/150       15/148/148
OPMULA    235/157/167        0/153/153        0/153/153
MULA      326/217/222        0/214/214        0/214/214

Findings:
- The ACC-target multiplies close to ZERO under exactmul: the exact
  multiplier is bit-exact against the oracle where it receives raw
  operands.
- The Fd-target residuals (15-47 per 1600, ~1-3%) are entirely the
  mandatory per-title operand clamps (TOTA/DoM/Ice Age on MUL forms,
  KH on SUB forms): every residual case carries an exponent-255
  operand that the pre-clamp flattens to FLT_MAX before the exact
  multiply, while the oracle computes the raw PS2 semantics (exp-255
  is a valid huge number there). The 7f800000 cases sit exactly one
  ulp under their raw results.
- Both accurate options together take the fused family to ~1%
  (MADD/MSUB/OPMSUB 15-16 per 1600), same clamp attribution.
- Flags are untouched by exactmul, by design: it is a value path.
  The multiply flag gap remains the class-gated detector's territory
  (see the overflow-hack commit).

Clamp elision (landed in this commit): under CHECK_VU_EXACTMUL the
cFs/cFt operand clamps on multiply-stage operands are elided (the
exact path has no infinity-propagation failure, so the clamp only
destroyed information there). Measured: MUL/MULx/MULi/MULq residuals
28/22/47/15 -> ALL ZERO -- plain multiplies are bit-exact against
the oracle under exactmul. Default emission proven byte-identical
via JIT hash (option off = zero effect).

Fused-residual attribution CORRECTED: plain MADD carries no operand
clamps (clampType 0), so the earlier blanket clamp attribution
overreached for the fused family. Case anatomy under both accurate
options shows every fused residual is an overflow-saturation
intermediate: the mul stage overflows, the oracle saturates the
product (exp-255 all-ones), and the composite add stage's INTERNAL
operand clamps (the global clampE layer) flatten that saturated
product to FLT_MAX before adding -- e.g. finite*FMAX + (-FMAX):
oracle 7fffffff, rec 7f7fffff. Next candidate, queued with the same
gate requirements: elide the composite add's internal operand
clamps when both accurate options own their stages.
