iconvex_platforms

# Platforms benchmarks

The representative benchmarks compare native codec paths with independent,
source-parsed Elixir references. Each operation must remain below 30× its
reference and each 20k→40k reduction count must scale within its declared
linear band.

Run the quick gates:

```console
ICONVEX_PATH=../iconvex mix run --no-compile bench/cp1090_benchmark.exs --quick
ICONVEX_PATH=../iconvex mix run --no-compile bench/evertype_source_qualified_bench.exs --quick
ICONVEX_PATH=../iconvex mix run --no-compile bench/vietunicode_vni_bench.exs --quick
```

The 2026-07-30 verification run covered:

- CP1090: 95/256 source mappings, six direct/public operations, worst observed
  native/reference ratio 5.161×, and four passing reduction-scaling gates.
- Evertype: 1,694/1,694 source rows across seven codecs and 21 operation gates;
  worst observed ratio 2.168×.
- VietUnicode VNI: 1,041/1,041 source rows across four token profiles and 12
  operation gates; worst observed ratio 1.15×.

Wall-clock rates vary by host, so CI asserts reproducible ratios, source
digests, parity, coverage counts, and reduction scaling rather than a fixed
throughput number.