feat: embed feature — minizlib-backed deflate family for embedded targets

[MagicalTux]

What

An embed Cargo feature: a mode under which algorithms with a small variant are built as that variant under the same names, so the same downstream code compiles either way. First family: deflate / zlib / gzip, as native compcol code under src/embed/ (the code was absorbed from the author's minizlib crate, which is being retired): a push decoder in the manner of zlib's puff (~1.1 KiB of state between pieces, Huffman codes decoded bit by bit from the count of codes per length, no tables to build) and a greedy fixed-Huffman LZ77 encoder, both written straight against the RawEncoder / RawDecoder traits. No new dependency.

compcol = { version = "0.6", default-features = false, features = ["embed", "gzip"] }

• No allocation. deflate/zlib/gzip no longer imply alloc; the standard build compile_error!s with guidance when neither alloc nor embed is on. Breaking for --no-default-features --features gzip builds that relied on the implied alloc.
• Codec structs own everything (decoder ≈ 34 KB, encoder ≈ 6 KB), are const-constructible with an all-zero state (a static lands in .bss), and index nothing that can panic.
• The decoder writes straight into the caller's slice (a match copy interrupted by a full output resumes as its own state); the encoder queues bits in a 64-bit buffer so any non-empty output buffer makes progress.
• Same error variants and gzip multi-member semantics as the standard build. compcol::embed documents the trade-offs (fixed ratio, no sync flush, no preset dictionary) and the size contract.

Sizes CI now enforces (thumbv7em-none-eabi, opt-level = "z", LTO)

configuration	code	max	stack	max
gzip decode	3964	4200	88	256
zlib decode	4036	4300	92	256
raw deflate decode	3672	3900	84	256
gzip encode	1844	2000	96	256
zlib encode	1804	2000	96	256
raw deflate encode	1514	1700	76	256
gzip encode + decode	5642	6000	96	256

Zero panic symbols, zero .data, no static RAM beyond the codec struct. tools/footprint/check.sh (new CI job) measures code/RAM/stack and fails on any ceiling; stack is the deepest call path from the entry point, read off the disassembly (stack.py), since -stack-size-section is not reachable from stable rustc.

Tests

• tests/embed.rs: CPython-zlib reference streams (dynamic/fixed/stored blocks, gzip with every optional header field, incompressible, empty), hand-built streams (a back-reference at exactly 32768, which zlib never emits; a "bomb" block where every two bits decode to 258 bytes), every input/output chunking, error variants, poison/reset/reuse, discard_output, concatenated gzip members, struct-size ceilings, and a 16 KiB-stack thread.
• tools/embed-crosscheck.sh (new CI job): the embed and standard CLI builds decompress each other's gzip/zlib/deflate output, both ways.
• Existing suites run in both modes (--features all and --all-features); standard-only tests (levels, FLEVEL/XFL, sync flush, dictionaries, one fixture with an RFC-invalid CINFO=15) are gated on not(embed).
• tests/cli.rs now feeds stdin from a thread: the embed ratio pushed the child's output past the pipe buffer and the old helper deadlocked.

Also: docs.rs and the release-binaries workflow now build with features = ["all"] rather than --all-features, which would otherwise ship the embed codecs.

🤖 Generated with Claude Code