fletchr-encodings 0.0.1rc6

Arrow-native Level 1 bit-pattern decoders (encodings) for fletchr, with byteforge kernels behind the built-ins.

fletchr-encodings

Arrow-native Level 1 bit-pattern decoders ("encodings") for the fletchr ecosystem.

An encoding turns the raw bits of a fletchr.uintn(bits=N) column into a number — unsigned/signed integer, IEEE 754 float, MIL-STD-1750A, BCD, and so on. This package owns the Arrow boundary (bits-from-type, validity-bitmap preservation) and a plugin system so users can add their own codecs. The built-in codecs wrap byteforge's numpy/C kernels, but byteforge is an implementation detail behind the built-ins — the public surface is entirely fletchr's own.

This is the Level 1 backend referenced by fletchr-measurand's SPEC-decode.md; Level 2 (engineering-unit scaling) lives in fletchr-measurand.

Public API

from fletchr_encodings import create_encoding, decode_array, list_encodings
from fletchr_uintn import uintn_array

raw = uintn_array([0x00, 0x7F, 0x80, 0xFF], bits=8)

# Resolve by canonical name and decode (bits read from the array type):
decode_array(raw, "2c").to_pylist()        # -> [0, 127, -128, -1]

# Or pass a resolved Encoding instance:
enc = create_encoding("2c")
decode_array(raw, enc)

Functions

Function	Purpose
decode_array(arr, encoding)	Decode a 1-D UIntNArray or 2-D FixedSizeListArray<UIntN>. encoding is a canonical name or an Encoding. Validates storage compatibility, reads the bit width from the type, preserves the validity bitmap.
create_encoding(name, *, bits=None)	Resolve a canonical name to an Encoding. bits (optional) validates storage compatibility eagerly. Raises UnknownEncodingError on an unknown name.
list_encodings()	Sorted list of registered canonical names (loads plugins first).
register_plugin_group(group)	Add an entry-point group for downstream meta-packages.
load_plugins(*paths, strict=False)	Explicitly import plugin modules (notebooks/tests).

Key design points

• Bits come from the type, not a constructor arg. An Encoding is stateless with respect to width; a single instance decodes any width it supports. The width is read from the column's fletchr.uintn type at decode time.
• Validity bitmap preserved end to end. Nulls in raw bits stay null in the decoded array.
• One canonical name per encoding, no aliases. The name is both the registry key and the value written to fletchr.encoding field metadata, so the registry is 1:1 with the SPEC-decode table.
• Level 1 output dtype is natural: integral encodings stay integral (u/bcd/gray/boolean → unsigned; 2c/1c/sm/ob → signed), floats decode to their natural width (ieee16 → float16, ieee32 → float32, ieee64 → float64, other floats → float64). Level 2 (in fletchr-measurand) promotes everything to float64.

Built-in encodings

Name	Decodes to	Required storage
u	unsigned int (identity)	uintn(bits=N)
sm	signed int	uintn(bits=N), N>=2
1c	signed int	uintn(bits=N), N>=2
2c	signed int	uintn(bits=N), N>=2
ob	signed int	uintn(bits=N), N>=2
ieee16 / ieee32 / ieee64	float16 / float32 / float64	uintn(bits=16/32/64)
1750a32 / 1750a48	float64	uintn(bits=32/48)
bcd	unsigned int	uintn(bits=N), N % 4 == 0
boolean	unsigned int (0/1)	uintn(bits=1)
gray	unsigned int	uintn(bits=N)
ibm32 / ibm64	float64	uintn(bits=32/64)
dec32 / dec64 / dec64g	float64	uintn(bits=32/64)
ti32 / ti40	float64	uintn(bits=32/40)

A storage-width mismatch (e.g. ieee32 on a bits=16 column) raises EncodingError before any decode happens, per the SPEC's storage-compatibility rule.

Custom encodings (plugins)

Subclass Encoding, give it a canonical name, and implement _decode_np. Subclassing alone registers it; declaring the module in the fletchr_encodings.plugins entry-point group makes it discoverable without an explicit import. See custom-encodings.md.

from fletchr_encodings import Encoding

class CrcPacked(Encoding):
    name = "crc_packed"
    def _decode_np(self, dns, bits):
        ...   # your own numpy; keep padding bits masked

A width-constrained codec declares fixed_bits or overrides supports_bits; _decode_np must keep padding bits masked (the input is guaranteed < 2**bits).

License

MIT — see LICENSE.