fix(archive): require a real bzip2 stream header, not a stray BZh

[zoza1982]

What

The multi-stream guard stops matching bzip2's bare 3-byte magic, and zstd stops guessing from bytes at all.

Why

The guard exists so a concatenated archive is refused rather than silently mounted with its later members missing — good. But it scanned the compressed file for a second occurrence of the format's magic, and bzip2's magic is three bytes. Three bytes recur by chance roughly every 16 MB of compressed entropy:

Archive size	Chance of a false refusal
~16 MB .tar.bz2	~60%
~100 MB .tar.bz2	essentially always

So the feature was effectively broken above ~10 MB, and the error told the user their archive was multi-stream when it was not. The verifier reproduced it end-to-end: a 14.5 MB single-stream .tar.bz2 built from this repo contained a stray BZh at offset 11714699 in every variant tried, and ArchiveVfs::open refused it with compressed_tar_multi_stream.

zstd's 4-byte magic has the same defect at 1/256 the rate — negligible on small files, roughly a one-in-five false refusal on a 1 GB archive.

Closes #

How

bzip2 — a real stream header, not a magic fragment. A genuine concatenated stream begins with all ten bytes: BZh, a level digit 1–9, then either the 48-bit block magic (digits of π) or the end-of-stream magic (digits of √π). Concatenated streams are whole files, so byte alignment holds and a plain byte scan is exact for this case. Chance match ≈ 2⁻⁷².

zstd — ask, don't guess. StreamingDecoder decodes exactly one frame, so anything still readable behind it is a second frame. into_inner() and a one-byte probe answer it exactly. No heuristic left.

The generic contains_magic_after_start (and Compression::magic, now unused) go with them: gzip never needed the guard (MultiGzDecoder handles concatenation properly) and xz is unsupported.

The old doc comment called a false positive "extremely rare". For a 3-byte magic that was off by many orders of magnitude, and the new comment says so with the numbers.

Testing

cargo test -p cairn-backend-archive     # 53 passed (+2)
cargo clippy -p cairn-backend-archive --all-targets --all-features -- -D warnings

The discrimination is what matters here, so the mutation check covers both directions:

Mutant	Result
restore the bare 3-byte match	a_bare_bzh_inside_the_compressed_body_is_not_a_second_stream fails, while a_real_second_bzip2_stream_is_still_found and bzip2_multi_stream_is_refused_not_silently_truncated still pass — i.e. it reproduces exactly the old bug
zstd never reports multi-frame	zstd_multi_frame_is_refused_not_silently_truncated fails

a_real_second_bzip2_stream_is_still_found covers both the block-magic and the end-of-stream form (an empty second stream has no block magic at all), so the tightened signature is not accidentally narrow.

Screenshots / output

No UI change — affected archives now open instead of erroring.

Checklist

• PR title follows Conventional Commits
• Branched off main
• cargo fmt, cargo clippy -D warnings, cargo test, cargo doc pass locally
• Tests added (2, plus the 2 existing true-positive tests as the other half of the mutation check)
• Docs updated — the rustdoc now states the real collision rate and why bzip2 needs a scan while zstd does not
• CHANGELOG.md updated under "Unreleased"
• Security review — n/a; this loosens a guard, so the true-positive tests are the safety argument
• No secrets, credentials, or generated artifacts committed

Risk & rollback

• This makes a refusal less likely, so the failure mode to worry about is the opposite one: a genuine multi-stream bzip2 archive mounting with later members missing. The two true-positive tests are exactly that case and both still pass, and the signature is the format's own documented stream header rather than a heuristic.
• A bzip2 stream whose header is not byte-aligned would be missed — but concatenated streams are whole files, so that cannot arise from concatenation, which is the only way a multi-stream file is produced.
• Rollback: revert the single commit.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SM3dDrioC5KntqjB6WDqjW