Move the bounded archive extractors out of vendor::registry_fetch into utils::archive #1012

[agent] Filed by the scheduled architecture audit routine (ecosystems and formats). Register: register comment.

Kind: refactor. Source: Part 5.6 ("registry_fetch.rs is misnamed and misplaced"). Register E29.

Problem

vendor/registry_fetch.rs is 3,048 lines at db83f01: 1,560 production lines and 38 tests from L1561. The review counted 1,535. Its own module doc (L1) describes it as "Bounded archive readers, integrity verification and registry metadata transport". Nothing in it is vendored-mode logic. It holds four unrelated families:

So agent mode (patch/), hosted restore (patch/redirect/upstream/), hosted/ and api/ all import crate::vendor::registry_fetch. That is a patch/hosted/api → vendor dependency for code that has nothing to do with vendoring.

The comment at L24–L26 justifies pub(crate) caps by a common::read_zip_members twin, which no longer exists. The caps have no importer outside the file except MAX_DOWNLOAD_BYTES.

Impact

• This is a layering inversion: the agent-mode and hosted-restore paths depend on the vendored module tree, which blocks moving vendor/ behind a backend trait (E21, Tracking: dispatch vendored backends through one per-ecosystem table instead of string matches in core and the CLI #959).
• The file name hides where the archive-bomb caps live. patch/package.rs and vendor/common.rs keep their own caps (by design; different inputs), and registry_fetch is not where a reader looks for them.
• Risk: low. This is a pure move.

Proposed change (this issue: the first two families only)

• Create crates/socket-patch-core/src/utils/archive.rs (or utils/archive/{mod,go_module}.rs). Move the archive extraction and Go module zip families into it verbatim, with the tests that cover them.
• Keep the vendor::registry_fetch paths for these items through a pub(crate) use crate::utils::archive::* for one release of the code, or update the ~16 import sites directly. Either is fine; update the imports if the diff stays reviewable.
• Delete the stale read_zip_members comment.

Out of scope, as follow-ups recorded on register row E29:

• integrity, which waits for LockIntegrity to move into formats (Move LockfileEntry, LockIntegrity, SourceKind and http_url from vendor::lock_inventory into formats::entry #834);
• registry bases and transport, where download/build_registry_client are being changed by PR Bound registry downloads by ApiTimeouts instead of a 60 s total deadline (#872) #876 for Registry downloads give up after 60 s even while the body is still arriving #872.

Size and scope

• Files: vendor/registry_fetch.rs, the new utils/archive.rs, utils/mod.rs, and the importers (vendor/{cargo,composer_lock,gem,golang,maven_repo,npm_dir,nuget_feed,redownload,service_fetch}.rs, patch/jvm_jar.rs, patch/redirect/upstream/client.rs, api/vendor_prefetch.rs).
• About 1,000 moved production lines plus their tests. The import edits are about 20 lines. There is no behavior change.

Acceptance criteria

• git diff --color-moved shows the two families as moved blocks only.
• No file outside vendor/ imports archive or Go-module-zip items from crate::vendor::registry_fetch.
• Every archive refusal message and cap value is unchanged; the moved tests (zip, tgz, gem, module-zip and the Sink::Validate vs Sink::Write parity tests) pass unchanged.
• cargo test -p socket-patch-core and cargo test -p socket-patch-cli pass; cargo clippy --all-targets is clean.

Dependencies

• Not blocked. It doesn't touch download/build_registry_client, so it doesn't conflict with Bound registry downloads by ApiTimeouts instead of a 60 s total deadline (#872) #876.
• It helps Move the v5 JVM vendor orchestration and Maven acquisition out of maven_repo.rs into vendor/jvm #972 (the JVM fetch.rs move) and E21/Tracking: dispatch vendored backends through one per-ecosystem table instead of string matches in core and the CLI #959 by removing non-vendor importers of vendor/.

Backlog review — 2026-10-08

Consolidated into #959, #833. The retained tracker(s) preserve this issue’s implementation scope and acceptance criteria. Closing this separate scheduling item as not planned, not as completed.

The archive-extractor file move is supporting layering work. Link this proposal from the backend/formats trackers and close the separate scheduling item.