Wirehair Python binding

wirehair-fec is the standard Python distribution for Wirehair. New code should use the correctly spelled import:

import wirehair

message = b"application payload"
with wirehair.Encoder.create(message, block_bytes=1200) as encoder:
    first_packet = encoder.encode(0)

The historical import whirehair spelling remains supported. Both names export the same classes, constants, and functions.

Native-library policy

The wheel is pure Python (py3-none-any) and deliberately does not bundle wirehair.dll, libwirehair.so, or libwirehair.dylib. Install a matching Wirehair shared library from the same release through CMake, a system package, or your application's deployment system. The binding checks major API version 2 and requires every symbol used by that release; an older 2.x library missing newer entry points is rejected. Static archives cannot be loaded by this binding.

Discovery is deterministic and follows this order:

1. A non-empty WIREHAIR_LIBRARY is authoritative and names the exact shared library file to load. Failure does not fall back to another installation.
2. A non-empty WIREHAIR_PREFIX is authoritative and searches only its bin, lib, and lib64 trees using platform library names.
3. CMake-prefix and active Python-prefix locations are searched.
4. The operating-system loader is queried (find_library and normal loader names).

Linux uses libwirehair.so/libwirehair.so.2, macOS uses libwirehair.dylib/libwirehair.2.dylib, and Windows uses wirehair.dll/libwirehair.dll. wirehair.initialize() validates the native API version before a codec is created.

Installation modes

Build and install the Python distribution with any PEP 517 frontend:

python -m pip install .
python -m pip install -e .  # editable source checkout

A shared or dual CMake installation also installs both import names under its python directory for prefix-oriented deployments. Static-only CMake installs omit all Python artifacts because they do not contain a loadable library.

The reusable encode_into, recover_into, and recover_block_into methods write to caller-owned bytearray or writable contiguous memoryview storage. Decode/recovery success is not authentication: verify a digest or MAC obtained from trusted application metadata before using recovered bytes.

Encoder.detach_input() severs the native encoder from its source after the recovery columns created during construction are ready. For borrowed input it also releases the binding's retained buffer view, so the caller may resize, modify, or destroy the source immediately after success. For owned input it releases the native private message copy. The method is idempotent and packets stay byte-identical, but systematic packet generation becomes slower because the encoder must reconstruct those packets from recovery columns. Do not call it concurrently with encode, reuse, conversion, or close.