parity tool: a new-kernel mismatch must fail the run; dispatch comment fallback was stale

[pipeob0]

The two follow-ups you listed in your #43 merge note. Both are yours to shape; nothing here changes a kernel.

1. The comment above the gate/up dispatch described a fallback that no longer exists. It said STRATA_NO_IQ512 / STRATA_NO_IQ256 fall back to ggml-cpu's single-token vec_dot. Since #43, STRATA_NO_IQ512 drops an AVX-512 CPU to the AVX-2 kernel, STRATA_NO_IQ256 drops the AVX-2 kernel, and ggml's vec_dot is reached only with both set (and on a CPU without AVX-512, STRATA_NO_IQ512 changes nothing at all). Comment-only commit.

2. The parity tool printed the new kernels' distance from ggml's vec_dot without affecting the exit code. A kernel that silently diverged still reported 0 failures, which is the one property of that tool people rely on. Both comparisons (the gate/up rows in check, and the IQ4_NL down) are now gated at 1e-5 and increment failures. Measured headroom: gate/up rows 3.1-3.4e-08, IQ4_NL down ~1e-7, so the gate sits ~100x above the observed spread.

Verified on this box (IQ2_XS, IQ2_XXS, IQ3_S, IQ4_NL layers):

• threshold 1e-5 → native_expert_parity: 0 failures (unchanged, as it should be);
• threshold temporarily dropped to 1e-9, i.e. below the observed spread, to prove the gate is live:

iq2_xs  avx2 gate rows MISMATCH (rel 3.12e-08 > 1e-9)
iq4_nl  down           MISMATCH (rel 9.91e-08 > 1e-9)
iq2_xxs avx2 gate rows MISMATCH (rel 3.35e-08 > 1e-9)
iq3_s   avx2 gate rows MISMATCH (rel 3.34e-08 > 1e-9)

Both sites fire, and the run exits non-zero. The threshold is back at 1e-5 in the commit.

Left alone deliberately: the q2_0 down comparison prints ~2.1e-04, which is the pre-existing Q2_0 kernel against ggml and is expected to be that coarse — gating it at 1e-5 would be a false alarm, so it stays print-only.

[Niko1221]

Thanks @pipeob0 - this closes both follow-ups from #43, and showing the gate fires at 1e-9 is exactly the proof I wanted. Merged, in v0.1.20.