ferrodec 1.11.0 — perf pass

~17 % aggregate speedup across the core_ops bench suite vs 1.10.1, with the headline operations seeing 23–27 % wall-time reductions. No public API change.

Performance (cumulative vs 1.10.1, Apple Silicon, rustc 1.95.0, release with thin LTO + 1 codegen-unit)

Bench	Before	After	Δ
add	7.98 µs	5.79 µs	−27.5 %
sub	39.92 µs	30.53 µs	−23.5 %
mul	42.45 µs	31.36 µs	−26.1 %
div	50.69 µs	44.78 µs	−11.7 %
fma	488 µs	415 µs	−14.9 %
sub_alignment_heavy	7.98 µs	5.80 µs	−27.3 %
mul_full_precision	6.60 µs	5.23 µs	−20.8 %
parse_str	3.68 µs	2.91 µs	−20.9 %
from_i128	2.98 µs	2.24 µs	−24.8 %

Load-bearing optimizations

• round_and_pack_finite now caches decimal_digit_count once per call instead of recomputing it 3× across the rounding / overflow-check / preferred-quantum branches. This single change is the bulk of the aggregate uplift (commit 15a7b98).
• U256::mul_pow10 looks up 10^k from a precomputed [u128; 39] table instead of running an iterative mul10 loop. Hot consumers: alignment shifts in addsub, the rounding pipeline's overflow-renormalize step, the up-renormalize in finalize_finite (commit a53ddb4).
• A third commit (84e4598) unified two duplicated digit-extraction loops in the rounding path.

Audit log

The full per-candidate audit — including three optimization candidates that were tested and reverted as no-op or noise-floor — is in docs/decisions/0008-perf-results.md. The pre-pass baseline lives in docs/decisions/0007-perf-baseline.md.

This release also introduces the ADR audit log under docs/decisions/. Eight ADRs (0001–0008) backfill the design log: BID over DPD, per-op status threading, method-only API, the Verus pilot outcome, the will-not-fix non-IEEE rounding directives, and the perf pass results pair.

Bench coverage additions

benches/core_ops.rs gains alignment-heavy / full-precision / magnitude-extreme variants. New benches/comparison.rs covers partial_cmp / total_cmp / compare_total_magnitude / 64-element sort. benches/conversions.rs adds from_i32 / from_u32 / from_u64 / to_i32 / to_u64 / to_u128. Permanent regression-watching shapes.

Verification

• 425 lib tests pass (--features=transcendentals,binary-float,serde,ops,num-traits).
• 63/63 Kani harnesses verify in ~2 minutes total.
• decTest conformance: 8 622 / 0 / 99 — unchanged from 1.10.1.

🤖 Generated with Claude Code