sc/kernels: apply Owen scramble family at full stoc_len too

[heroarmor]

Summary

_prepare_rng_prefix previously skipped the per-dim Owen XOR when stoc_len == 2**sc_prec (full-length Sobol), applying the scramble only to truncated prefixes. This re-applies the scramble family at full length too.

The skip was wrong for the enable-signal matmul. That path reads joint counts |{t : rng_a[d,t]<ba AND rng_b[d,t]<bb}|, which depend on the per-d trajectory — not just the marginal count count(r<v)=v (the only quantity invariant under XOR over a Sobol permutation). With make_sobol_simple_config broadcasting the same Sobol-Q/Sobol-K pair across all D dims, skipping the scramble at full length left every dim with an identical joint trajectory, so SC noise across D accumulated instead of averaging out.

Always scrambling — selecting counter / bitrev / random via SC_OWEN_MODE inside _owen_scramble — restores per-dim decorrelation at stoc_len=256 (sc_prec=8).

Relation to prior work

Supersedes the closed #11, rebased onto main so it also carries the halve_bipolar_stoc_len flag + housekeeping from #12.

Effect (expected, Llama-3.1-8B-Instruct, wikitext-2 test, ctx=1024 stride=512, per_row, sc_prec=8)

config	PPL	xFP16
FP16 baseline	6.7711	1.000
INT8 per_row deterministic	6.9328	x1.024
SC sl=256 (no scramble at full length, prior)	7.9383	x1.172
SC sl=256 (scramble family applied)	to be re-measured	expect ≈ sl=128 floor

Test plan

• Re-run sl=256 per_row PPL with SC_OWEN_MODE=counter and SC_OWEN_MODE=bitrev; confirm both drop from x1.172 toward the sl=128 / INT8 floor.
• SC_DISABLE_OWEN=1 (or SC_OWEN_MODE=off) reproduces the prior unscrambled sl=256 number — confirms the gate is the only behavioral change.
• Truncated stoc_len < 256 numbers unchanged (that path already scrambled).

🤖 Generated with Claude Code

[Allenjin123]

let's discuss tomorrow on how to do it first