Optimize: representative benchmarks, cold_path hints, fair parse benchmark

[gfx]

Summary

Performance/benchmark work on the fpfmt crate, keeping correctness intact throughout (full comprehensive suite passes on both feature sets; the parse-scan change was additionally cross-checked with a 5.3M-string differential test before being reverted — see below).

Measured on Intel Xeon (2.10 GHz), Ubuntu 24.04, x86_64, rustc 1.95.

Changes

1. Benchmark at opt-level 3 ([profile.bench]). The workspace [profile.release] uses opt-level = "z" to keep the wasm binary small, and bench inherited it — so benchmarks were measured size-optimized, which no consumer does (a dependency's profile never propagates; downstream builds fpfmt at their own opt-level). A dedicated [profile.bench] opt-level = 3 fixes this while release stays "z" for the wasm story.

• format 235 ns → 197 ns in the reported benchmark; ryu now measured fairly (was penalized ~5× by size-opt). Wasm binaries unchanged.

2. core::hint::cold_path on rarely-taken branches (subnormal/zero in unpack64, exact-power-of-two and subnormal in short). Zero-cost, changes no results. Runtime effect is within noise for realistic input (those branches are already never taken for normal doubles), but the hint still guides code layout at no cost to the hot path.

• Requires the toolchain pinned to exactly 1.95 (the release that stabilized cold_path). 1.96/1.97 regress the wasm-size build ~4× on small for this crate, so they are avoided. rust-version = "1.95" declares the new minimum.

3. Fair parse benchmark. The parse row fed each value's f64::to_string(), which expands 5e-324/f64::MAX to 300+ digit decimals that fpfmt's 19-significant-digit parser rejects while stdlib parses — timing rejection against parsing. It now parses each value's shortest round-trip string (valid input both accept).

• Honest parse numbers: fpfmt 67 ns / small 80 ns / stdlib 113 ns.
• Note: I prototyped a digit-count early-exit in parse_text (huge speedup on the old benchmark) but an A/B showed it slows valid parsing ~35%, so it was not kept — the per-digit branch pessimizes the common path to speed up rare adversarial rejection.

Benchmarks (8 representative f64 values)

Task	fpfmt	fpfmt small	ryu	stdlib
format	150 ns	237 ns	308 ns	927 ns
parse	67 ns	80 ns	—	113 ns

Wasm size unchanged from the 1.93 baseline (14,291 / 4,078 bytes on 1.95).

Validation

• cargo test + cargo test --features small — pass
• cargo clippy + --features small — clean
• cargo fmt --check — clean
• wasm-size builds — 14,291 / 4,078 bytes

🤖 Generated with Claude Code

---

Generated by Claude Code