Skip repeated Sinsemilla q_0 sort work

[ValarDragon]

Result

The focused serial Sinsemilla table-sort kernel improved portably:

• Apple M4: -24.25 us.
• Six-worker x86-64 Linux with x86_64-asm: -25.09 us.

The enclosing lookup-permutation phase and full-proof measurements were
noise-neutral, so this PR claims only the resolved serial-kernel improvement.
Paired proof checks produced exact proof bytes. Retained proving-key memory and
public APIs are unchanged.

Change

The Sinsemilla generator table repeats its compressed first-row value, q_0,
over roughly half of its usable rows. Table preparation now removes a dominant
nonzero q_0 run, constructs and sorts Pasta keys for the residual values, and
inserts the run at its canonical sorted position.

The optimization retains the existing one-quarter eligibility threshold. An
empty input, zero q_0, a sub-threshold run, or a non-Pasta field takes the
unchanged generic sort. Declined specialization leaves both input buffers
untouched.

Validation

• cargo +1.97.1 test -p zakura-halo2-proofs --lib
  • 279 passed.
• cargo +1.97.1 test -p zakura-halo2-proofs --lib \ plonk::lookup::prover::tests
  • 22 passed.
• cargo +1.97.1 test -p zakura-halo2-proofs --no-default-features --lib \ repeated_value_pasta_sort_matches_generic_sort_and_fallback
  • Passed for both Pasta fields; the build emitted only pre-existing
    no-default-feature warnings.
• cargo +1.97.1 fmt --all -- --check

The new equivalence tests cover all-repeated inputs, the exact threshold,
below-threshold and zero-value fallback, empty input, and insertion at the
start, middle, and end of the canonical field ordering. They compare both the
sorted values and key metadata with the generic route for both Pasta fields.

[v12-auditor]

Note

Complete: Audit complete. V12 did not find any issues that need review.

Open the full results here.

Analyzed one file, diff 331dbd7...fac0e93.