VU upper-op census v2: Q-forms plumbed, flags architecturally observed
=====================================================================
Rig: audit exports with flag-instance seeding and Q/P representation
mirroring (this commit); v3 program shape: op, four NOP pairs for
FMAC flag latency, then FSAND vi4,0xFFF and FMAND vi5,vi6 so status
and MAC are read the way a program reads them -- the flag hack then
has to compute them, and lazy sync questions disappear because the
observation is architectural on both engines. 600 cases per op, 12%
special-value mix, VU1, default recompiler vs ps2float interpreter.

op       value lanes /2400   status /600   mac /600
ADD             968                5            5
SUB            1025                1            1
MUL             423              266          266
MADD            436              276          284
OPMSUB          833              317          471
MULq            368              199          199
ADDq            994                1            1

Findings:
- Q-form rows are live (the v1 near-zero rows were the documented
  plumbing artifact): MULq and ADDq diverge at their family scale.
- Add-family flags are near-exact in default mode: the divergence
  that remains is in the handful of special-value cases.
- Multiply-family flags diverge on a third to half of observed
  cases. First decoded example: rec status 0x082 vs oracle 0x28A;
  the missing 0x208 is the O/U sticky pair -- the recompiled
  multiply path does not model mul overflow/underflow flag setting,
  which the old fpaudit flag-incidence rows (mul OF/UF ~12% of ops)
  predicted and this measures through real engines end to end.
- OPMSUB MAC divergence (471) exceeds its status divergence (317)
  and both exceed the plain-mul scale; the deep dive stays queued.

Campaign criteria (maintainer): accuracy against the oracle first;
performance improvement without accuracy loss; bit-exactness ideal,
equal-or-better accepted where exactness is prohibitively expensive.
Next: OPMSUB decomposition; multiply flag-model options costed
against those criteria; per-op throughput columns (default vs
accurate paths) from timed exec batches; lower table; EFU; EE COP1.
