Iconvex Extras

iconvex_extras is the GNU libiconv 1.19 --enable-extra-encodings companion for Iconvex. It adds the exact 86-codec complement of Core's 112 GNU-default codecs: 7 GNU extra, 9 AIX, 15 DOS, 2 OSF/1, and 53 z/OS encodings.

def deps do
  [
    {:iconvex, "~> 0.1.1"},
    {:iconvex_extras, "~> 0.1.1"}
  ]
end

Starting the application is enough:

{:ok, "Café"} = Iconvex.convert(<<0x43, 0x61, 0x66, 0x82>>, "CP437", "UTF-8")
{:ok, "IBM-1047"} = Iconvex.canonical_name("CP1047")
{:ok, all} = Application.ensure_all_started(:iconvex_extras)
198 = length(Iconvex.encodings())

Stopping Extras returns a Core-only runtime to exactly 112 codecs; restarting it restores all 198. Extras contributes 342 normalized GNU spelling keys, for 758 Core-plus-Extras spellings in total.

One atomic extension

The application publishes all 86 codec routes and the :euc_jisx0213 table provider through one Iconvex.Extension.register/2 transaction. One opaque token owns the complete commit. A conflict at even the final route rolls back every earlier route and the provider; shutdown removes only records still owned by that token. Stop/restart and a supervised Core registry-worker restart retain the same exact ownership guarantees.

Extras uses managed priority 300 so GNU compatibility aliases outrank aliases from ordinary source leaves. Core built-ins still resolve first, and canonical claims still outrank aliases regardless of package priority.

The package contains 85 packed mapping tables plus the ISO-2022-JP-3 state machine. Shared Core engines provide native linear strict, discard, callback, and substitution paths without a NIF or operating-system iconv dependency.

Extras is one leaf in the final split, whose runtime packages expose 2,105 unique canonical codecs. The unpublished research/integration workspace owns no runtime registry routes, and Extras depends only on Iconvex Core.

Exact support and evidence

SUPPORTED_ENCODINGS.md is the generated complete 86-codec list with GNU definition origin. SUPPORTED_CODEC_INVENTORY.csv is the machine-readable canonical-name, alias, module, and statefulness snapshot.

The tests port all 85 applicable GNU mapping and inverse fixtures, ISO-2022-JP-3 state and streaming behavior, IBM-1047 policy surfaces, long invalid-input recovery, exact alias coverage, and atomic lifecycle behavior. The committed sequential UCS-4BE corpus contains every code point from U+0000 through U+10FFFF: 1,112,064 Unicode scalar values and all 2,048 surrogate code points.

EXHAUSTIVE_UNICODE_DIFFERENTIAL.md records the source-bound 198-codec forward, reverse, and cross-decode comparison with GNU libiconv 1.19. The final run is byte-exact over all 1,114,112 code points, with zero mismatches. Forward and reverse timings are independently gated at 30x using a C11 in-memory GNU engine helper. Both engines calibrate one fixed iteration count using the fastest of three probes at every doubling; that fastest probe must reach 10,000,000 ns, so one delayed probe cannot end calibration early. They then report the fastest of three final batches at the fixed count. EUC-JISX0213 is 2.95x forward/15.38x reverse and CP943 is 4.03x/20.92x. Across all 198 codecs, ISO-2022-CN is worst at 24.80x under the unchanged 30x ceiling.

ICONVEX_PATH=../iconvex mix test --warnings-as-errors
ICONVEX_PATH=../iconvex mix compile --warnings-as-errors
ICONVEX_PATH=../iconvex MIX_ENV=prod mix run bench/benchmark.exs
GNU_ICONV=/path/to/gnu-libiconv-1.19/bin/iconv \
  ICONVEX_PATH=../iconvex MIX_ENV=prod \
  mix run tools/exhaustive_unicode_differential.exs

The C helper is offline benchmark/test tooling, not shipped runtime code. The Hex artifact contains only runtime source and tables, licenses, inventories, and public evidence; it excludes test/, bench/, tools/, PDFs, and large development corpora.

ICONVEX_PATH is only the sibling-checkout development override. Published builds use {:iconvex, "~> 0.1.1"} from Hex.