Starfield Encoding Specification (SFENC) v0.2

SFENC stores bytes in a lossless starfield PNG. It is not encryption. Key backups must encrypt their payload before encoding. v0.1 was unimplemented and not recoverable as specified; reject its version byte.

Limits and canonical encoder containers

Input is 0–131072 bytes. Image width equals height and is the first of 256, 512, 1024, 2048 with capacity for all stars below. Reject files above 20 MiB, other dimensions, unsupported versions, lengths, compression, nonzero reserved bytes, or checksum failures before allocating from claimed lengths. All arithmetic below is integer arithmetic (division means floor), except no floating-point operations are needed at all.

The payload is canonical gzip: header hex 1f8b08000000000000ff, DEFLATE stored blocks, then little-endian CRC32 and original length. Split input into maximal 65535-byte blocks, with one empty block for empty input. Each block starts with byte 0 (nonfinal) or 1 (final), little-endian LEN and its 16-bit one's complement, then LEN bytes. No additional blocks, members or trailing data are permitted. Compression field is 1; v0.2 does not support zstd or uncompressed payloads. Stored blocks deliberately trade file size for reproducibility and bounded decoding without browser compression dependencies.

PNG uses signature 89504e470d0a1a0a, then exactly IHDR, IDAT, IEND. IHDR is width/height big-endian, depth 8, color 6 (RGBA), compression/filter/interlace 0. Each scanline has filter byte 0 then opaque RGBA bytes. One IDAT contains zlib header 7801, the same stored-block algorithm over scanlines, then big-endian Adler32. Chunks use standard big-endian length and CRC32(type + data). IEND is empty. CRC32 is reflected IEEE polynomial edb88320, initial/final XOR ffffffff. Adler32 starts a=1,b=0, modulus 65521. These exact container rules define canonical encoder output. Decoders also accept lossless PNG re-encoding as specified below. No canvas/color conversion is used.

Header

The first 22 pixels of row 0 contain 64 header bytes followed by two zero pad bytes, three sequential bytes per pixel in RGB; alpha is 255. Write the header last. All header integers are little-endian.

PRNG and placement

Every operation on PRNG words wraps modulo 2^64. SplitMix64 starts at seed; four times add 9e3779b97f4a7c15, set z=state, set z=(z XOR (z>>30))bf58476d1ce4e5b9, then z=(z XOR (z>>27))94d049bb133111eb, and output z XOR (z>>31). Those outputs initialize s0..s3.

xoshiro256** outputs rotl(s15,7)9. Then t=s1<<17; s2^=s0; s3^=s1; s1^=s2; s0^=s3; s2^=t; s3=rotl(s3,45). rotl is a 64-bit left rotation. Draws return uint64, never floats.

Let C=floor(W/5), R=floor((H-1)/5). Cells are row-major integers 0..CR-1. Shuffle them with Fisher–Yates: for i=CR-1 down to 1, j=nextUint64() mod (i+1), swap cells[i],cells[j]. Cell k has center (2+5(k mod C), 3+5floor(k/C)). Cells cannot overlap or reach the header.

For gzip length L, N=ceil(105L/100)+50. Choose the smallest permitted dimensions with CR >= N, else reject. The first L shuffled cells carry gzip bytes; the remaining N-L carry nextUint64() mod 256, one draw per dummy. Neither background nor payload stars consume random numbers. Seed 0 is valid.

Pixels

Fill background in row-major order. Let q=seed mod 256, base=2+((13x+7y+q) mod 6), distance=abs(10y-5H-3(x-W/2)), glow=max(0,24-floor(distance/10)). RGB=(base+floor(3glow/5),base+floor(7*glow/10),base+glow), alpha=255. This integer galactic band replaces v0.1's platform-dependent transcendental math.

For byte b let v=192+(b mod 64), hue=floor(b/64). RGB by hue: 0=(v,v,v); 1=(floor(85v/100),floor(90v/100),v); 2=(v,v,floor(70v/100)); 3=(v,floor(70v/100),floor(60v/100)). Hue is uniquely identifiable; v is max(R,G,B); decode b=64hue+v-192 and require re-encoding b to reproduce RGB exactly.

Draw each star with that RGB, alpha 255. Hue 0 draws center only; hue 1 adds four axial neighbors at distance 1; hue 2 draws the 3x3 square; hue 3 draws that square plus four axial neighbors at distance 2 (13 pixels). Shapes fit within their own 5x5 cell.

Decode and conformance

Validate bounded PNG structure and every chunk CRC. Accept noninterlaced, 8-bit RGB (color type 2) or RGBA (type 6), contiguous multiple IDAT chunks, standard zlib DEFLATE compression, and row filters 0–4 (None, Sub, Up, Average, Paeth). Bound inflated data to exactly H * (1 + W * channels) before allocating/writing beyond that size. Reject unsupported critical chunks, transparency chunks, invalid chunk order, missing IEND, trailing data, and checksum/decompression failures. Other ancillary chunks may be ignored; interpret stored channel values directly, without color-profile transformations. Normalize RGB to RGBA by adding alpha 255.

Extract header; validate fields, lengths and the canonical dimensions. Recreate shuffled cells. Decode the first L centers, verify gzip CRC against header before inflating, then verify bounded canonical gzip and its CRC/length. Re-render the starfield with the extracted seed and require exact equality of the decoded RGBA pixels, including alpha, background, header padding, dummy stars and shapes. Return original payload bytes only on success. Container compression, filter selection and metadata may differ; pixel changes are not tolerated.

This decoder compatibility revision preserves version 2 and all existing canonical encoder vectors. It accepts lossless optimization such as removing the redundant opaque alpha channel, without weakening payload integrity checks. Resizing, JPEG conversion and other pixel-changing edits remain unsupported.

Given identical payload and seed, the entire PNG, including its body, is byte-identical. Encryption is randomized: this promise concerns already-encrypted payload bytes, not separate encryption operations. tests/fixtures/starfield/vectors.json contains SHA256 conformance vectors; tests/Support/starfield_reference.py is an independent Python reference codec. Verify full payload range, empty input, all byte values, capacity boundaries and malformed input using its --test entry point.