v2.2.0 — precomputed compressed-float entry points, the normative range clamp, and string benches

Three things for users of this crate, all merged on green this week.

Precomputed compressed-float entry points — serialize_compressed_float_precomputed takes the four derived constants (step count, wire width, delta, min) computed once at build or generation time, skipping the per-call derivation. Additive: the existing API is untouched, and the two paths are wire-identical by construction, held by the differential suite.

The normative integer clamp (STANDARD.md, 2026-08-23). For declarations with step counts in [2^23, 2^24), float32 rounding could push the quantized code past max_integer_value — a top-of-range write was rejected by its own reader, or leaked a 25th bit into the stream. The writer clamps after the floor; no byte changes outside that band. Witnesses pinned in the tests and corpus.

string and wstring bench rows — first in this crate, with baselines recorded.

🤖 Generated with Claude Code