feat(rt): add os/unzip

[abogoyavlensky]

Resolves: #687

(os/unzip zip-path dest-dir) extracts a zip archive into dest-dir and returns it. Motivation: dependency tooling built on let-go currently shells out to unzip to expand jars, and this removes that host-binary requirement.

Extraction is confined by os.Root, which opens each path component under the destination and refuses any that escapes. That covers both the lexical case (../evil.txt) and the symbolic one (link/x where dest/link already points elsewhere), and unlike a check-then-write guard it leaves no window for a process sharing the destination to swap a validated directory for a symlink after the check. Absolute entry names are refused rather than silently re-rooted.

Safety over fidelity otherwise: symlink, device, fifo and socket entries are skipped rather than recreated. Permissions follow the unzip(1) contract — as recorded, masked by the umask — with a 0644 fallback for DOS/FAT entries that record no unix mode. A failed extraction is never worse than no extraction: a directory entry colliding with an existing file is an error rather than a swallowed EEXIST, and each entry is opened before its destination is cleared.

Tests: pkg/rt/os_unzip_test.go builds archives programmatically, including the adversarial ones, plus a test/os_unzip_test.lg smoke test.

[mparrett]

Approving. Built it and ran the suite: 22 unzip subtests pass, and it builds for js/wasm, plan9/amd64, linux/amd64, and wasip1/wasm. The windows/amd64 failure is pre-existing in term.go and reproduces on the merge base.

Routing every write through an os.Root confined to the destination closes the check-then-write race that a lexical containment guard cannot, and the tests cover zip-slip, absolute entry names, and escape through a pre-existing symlink. Stripping setuid via .Perm() and separating a recorded mode from archive/zip's synthesized 0666 are both easy to miss.

No findings.

Not a change request: io.Copy is unbounded, so a crafted archive can expand without limit. That matches io/slurp and http/get, and there is no in-tree caller yet, so it isn't a defect here. It becomes worth revisiting when something starts unzipping artifacts fetched over the network, where an optional caller-supplied byte limit — the shape net/read! already uses — would fit better than a hidden default.

[mparrett]

Now these have landed: between them they cover half of what let-go needs for Grenadine's host contract.

Checked all 17 slots against lg on current main. Most already map onto os, io, http, and edn: :http-get, :read-edn, :getenv, :home-dir, :run-process, :mkdirs, :exists, :find-files, :read-bytes / :write-bytes, plus the os/stat fields behind :directory and :regular-file. With hash/sha1 and os/unzip merged, that leaves:

• :atomic-move — no rename in os or clojure.core
• :delete-tree — delete-file only, so no recursive form
• :canonical-path / :absolute-path — no path or filepath namespace
• :bytes->utf8 / :byte-count — depends on how bytes get represented, so more a design question than a missing function

Are you working on any of these? If not I'd pick up the first two, which look like small additions in the same shape as os/unzip. Asking first since #690 and #688 landing together read like a batch, and you may have these queued already.

The complete contract is src/grenadine/host/jvm.clj in clojurestar/grenadine, with glojure.clj beside it as the non-JVM implementation. let-go already passes Grenadine's portable resolver suite (52 tests, 180 assertions); the effect host is what's missing.

[abogoyavlensky]

@mparrett No, I'm not working on any of these, please, feel free to pick it up.