Skip repeated Sinsemilla input sort work

[ValarDragon]

Summary

• derive the first theta-compressed Sinsemilla table value directly from the
  proving key before lookup tasks are scheduled
• remove that repeated value from each Sinsemilla input before generating and
  sorting Pasta keys, then reinsert its canonical sorted run
• preserve independent input/table preparation, the generic fallback, proof
  bytes, sort scratch capacity, and retained proving-key memory

Motivation

#467 skips sorting the repeated first-row value in the shared fixed
Sinsemilla table. The two Sinsemilla inputs per Action have the same stronger
shape, but they were still sorting all 2,042 usable rows. In the benchmark
fixtures, 1,056--1,108 of those rows equal the table's first compressed value.

Waiting for shared table preparation would serialize independently scheduled
tasks. Instead, this change reconstructs the same value from the three
unrotated row-zero fixed values with two Horner steps, then reuses #467's
validated repeated-value sorter. Non-Pasta fields, zero or non-dominant
repetitions, and structurally unrelated tables retain the generic sort.

Benchmarks

Paired proofs used Rust 1.97.1. Every control/candidate pair produced identical
proof bytes. The focused metric is the sum of the two Sinsemilla input sorts
per Action; it excludes the already-optimized u10 lookup.

Host	Workers	Actions	Pairs	Control	Candidate	Delta (95% CI)
Apple M4	10	1	64	122.2 us	81.6 us	-40.6 us [-48.8, -32.6]
Apple M4	10	4	32	497.4 us	305.2 us	-192.3 us [-211.5, -173.4]
x86-64 Linux, x86_64-asm	6	1	64	175.0 us	119.0 us	-56.0 us [-58.3, -53.9]
x86-64 Linux, x86_64-asm	6	4	32	639.7 us	431.8 us	-207.9 us [-218.7, -197.8]

Uninstrumented full-proof measurements were noise-limited: M4/10 measured
43.231 -> 43.246 ms at one Action and 110.261 -> 110.055 ms at four Actions,
with both paired confidence intervals crossing zero. This PR therefore claims
the resolved focused saving, not a resolved full-proof delta.

Correctness and API

• the direct q_0 derivation uses the same fixed columns, row, expression
  order, and theta-Horner order as table compression
• the existing repeated-value sorter validates the nonzero dominant run and
  otherwise leaves both buffers untouched for the generic fallback
• exact sorted-output equivalence remains covered for both Pasta fields, and a
  new test covers direct row-zero derivation and rotated-query rejection
• 23 lookup-prover tests and the Orchard Post-NU6.3 proof/version test pass
• no public or pub(crate) API, transcript, RNG consumption, cache, allocation
  capacity, or retained-memory shape changes; only a private helper signature
  gains the optional value

[v12-auditor]

Note

Complete: Audit complete. No review-worthy issues remain after automated review. One finding was auto-invalidated.

Open the full results here.

Analyzed one file, diff 3729c32...ce5cdf7.