Adding python bindings

[AAVision]

Added

• Python bindings via PyO3 — new python/ crate publishing the decayfmt-py package on PyPI (import name: decayfmt).
  • corrupt_bytes(data, x, kind) — decayed copy of immutable bytes
  • corrupt_in_place(buffer, x, kind) — zero-copy decay of a bytearray
  • encode_file(source, output) / encode_bytes(source, output) — encode clean files
  • decay_file(path) — open without display: permanently corrupts on disk, returns (kind, dims, bytes)
  • parse_filename(name), read_header(data), write_header(kind, width, height), __version__
  • Errors map to built-ins: OSError (I/O), PermissionError (read-only target), ValueError (everything else) — messages match the CLI exactly
  • abi3 wheels for Python 3.11+; requires-python >=3.11
• GitHub Actions workflow (python.yml) — builds and smoke-tests wheels on 5 platforms (manylinux x86_64 + aarch64, macOS arm64 + x86_64, Windows), publishes to PyPI via trusted publishing on py-v* tags
• Docs — python/README.md (PyPI long description), python/examples/example.py (every function + every exception), python/RELEASE.md (release checklist), new ## Python and ## Performance sections in root README

Changed

• Corruption loop rewritten for throughput — bulk u16 RNG pools replace per-byte f64 draws, exact-replacement tables replace rejection sampling, payloads ≥ 1 MiB corrupt in parallel (rayon) across cores. Measured on this machine (macOS arm64, 64 MiB payload): text x=1 142 → 1366 MiB/s (9.6x), text x=10 85.6 → 754 MiB/s (8.8x), image x=1 187 → 1421 MiB/s (7.6x), image x=10 116 → 950 MiB/s (8.2x)
  • Statistics unchanged: per-byte probability stays 1 - exp(-x/10) (quantization ≤ 1/65536), image alpha never touched, trailing non-pixel bytes untouched, randomness stays OS-entropy-seeded and never replayable
• Encode path — dropped second full-size buffer in write_decayfmt (header + payload streamed); text encode measured 7.9 → 3.9 ms (2.0x)
• Smaller dependency tree — image default features trimmed (avif encode-only dead code, internal rayon); rav1e removed from lockfile. Decode format parity kept (exr retained)
• Cargo workspace — members = ["python"] with default-members = ["."]; existing CLI CI/release workflows untouched

Benchmark	Before	After	Improvement
Text corruption ×1	142 MiB/s	1,366 MiB/s	9.6× faster
Text corruption ×10	86 MiB/s	754 MiB/s	8.8× faster
Image corruption ×1	187 MiB/s	1,421 MiB/s	7.6× faster
Image corruption ×10	116 MiB/s	950 MiB/s	8.2× faster
Encode 16 MiB text	7.9 ms	3.9 ms	2× faster

Internal (no CLI behavior change)

• encode::encode_bytes, decode_image, text_payload, write_decayfmt, open::decay_in_place now pub for embedding
• New bench harness (benches/, no new dependencies, DECAY_BENCH_MB override) and large-payload statistical tests (8 MiB parallel path, seeded determinism, alpha/tail invariance)

Notes

• CLI functionality, commands, error messages, and file format are unchanged — perf work and Python binding sit on top of the existing library
• End-to-end encode is bounded by image decode + disk I/O; the large speedups above are the corruption step
• PyPI package name is decayfmt-py (import stays decayfmt); releases tagged py-v* independently of CLI v* tags

[aravpanwar]

Awesome work. Went through the corruption rewrite and it's great: nothing's seeded in the real path, and the alpha and pixel-boundary handling held up across the parallel split.

CI's red on this thing: PyO3/maturin-action@v2 doesn't exist, so the python jobs die at "Set up job" before building. Switch both uses (build + publish) to @v1 and it should go green, then I'll merge.

[AAVision]

@aravpanwar fixed, check the tests and manage how you publish it to PyPI