Fix -g touching the cwd project's state (#436, #445) - #446

LLM Description written by Claude Code:claude-opus-5-5

Fixes #436
Fixes #445

Summary

A -g / --global-prefix run is meant for globally installed packages, which have no project lockfile. But when you ran one from inside a project, it still read and changed that project's hosted and vendored state. Now global scope leaves the project alone:

• get -g --mode hosted|vendored and scan -g --mode vendored (or --vendor) are usage errors (exit 2), the same as scan -g --mode hosted already was. Before, they rewrote the project's lockfile and left the global copy unpatched (get -g --mode hosted|vendored and scan -g --mode vendored rewrite the current project's yarn.lock instead of refusing, leaving the global copy unpatched #436).
• rollback -g and remove <purl> -g no longer run their hosted and vendored legs against --cwd. Before, they restored the project's hosted pins to the registry and unwired and deleted its .socket/vendor/ artifacts, which on vlt also deleted the installed package. They also no longer retire the project's pre-v5 hosted ledger (rollback -g and remove <purl> -g also unwind the current project's hosted pins and vendored wiring; on vlt they delete node_modules/left-pad too #445).
• apply -g and scan -g --mode agent now patch the global copy of a purl that the project vendors. Before, they skipped it with vendored_ownership_retained (rollback -g and remove <purl> -g also unwind the current project's hosted pins and vendored wiring; on vlt they delete node_modules/left-pad too #445, reverse direction).

Root cause

Nothing enforced "global scope excludes project state", so each command decided for itself whether the cwd project's hosted pins and vendor ledger counted under -g:

• get switched to agent mode under -g only when --mode was absent.
• scan refused hosted but not vendored.
• rollback and remove discovered and unwound the cwd hosted pins and vendor ledger with no global check.
• The agent paths (apply_patches_inner and scan's partition_agent_selection) used the cwd ledger's ownership set for global copies too. Scan also folded the cwd ledger's records into update detection, so it counted a global copy as already patched.

Fix

One rule in commands/mod.rs: project_state_in_scope(common) is false under global scope. global_mode_conflict(common, mode) builds the shared usage error.

• scan (resolve_mode_flags) and get refuse hosted and vendored mode through global_mode_conflict. Scan's existing hosted message is unchanged; vendored says ...no project lockfile to wire vendored artifacts into.
• rollback and remove, under global scope, discover no hosted pins and no vendor ledger, so the hosted and vendored legs have nothing in scope. retire_legacy_redirect_ledger is a no-op, and so is rollback's ledgerless-wiring check.
• rollback still reads the ledger, but only to keep the manifest records of purls the project vendors. Dropping them would hand a later vendor reconcile a revert with no backing record. An unreadable ledger still skips cleanup and GC fail-closed, but under -g it no longer fails the run, because no vendored leg runs.
• apply and scan: the cwd ledger no longer owns or records global copies.

CLI_CONTRACT.md is updated in the mode-resolution, get, rollback state-discovery and exit-code sections.

Test evidence

New suite: crates/socket-patch-cli/tests/global_scope_project_state.rs (offline, plus a wiremock API for scan).

Red on main: I reverted src/ to main and ran cargo test -p socket-patch-cli --all-features --test global_scope_project_state: 8 failed, the control passed. With the fix: 26/26 passed (including the shared harness self-tests).

Local checks: cargo clippy --workspace --all-features -- -D warnings is clean. In this container a single cargo test --workspace --all-features run hits the disk limit while linking about 200 test binaries, so I ran every cli and core test target in batches instead. All of them pass except about a dozen write-failure tests (*_write_failure_*, *unremovable*, relax_loop_must_not_traverse_symlinked_root). Those tests rely on read-only permissions, which don't apply to root, and the agent container runs as root. Several are in socket-patch-core, which this PR doesn't touch. CI on 92c71ad: 475 checks passed, 6 skipped by workflow conditions, 0 failed.

cargo fmt --all -- --check already fails on main: #277 landed about 120 files that aren't rustfmt-clean, and CI runs no fmt check. This PR's own hunks are rustfmt-stable (I checked them against cargo fmt output). I left the unrelated files alone so the PR stays reviewable.

The wrappers (npm/, pypi/, gem/) only dispatch to the binary, so they need no change.

Notes and follow-ups

• vlt is fixed through the shared hosted and vendored legs, not by vlt-specific code. The issue's npm/vlt projects take the same code path the pypi hosted and npm vendored tests cover.
• remove <purl> -g still drops the manifest record you named, even when the project vendors that purl through a legacy manifest-tracked (non-detached) vendor entry. v5 vendored entries are manifest-free, so this doesn't affect them.
• Not changed here: vex -g and list -g still read cwd lockfile discovery. Neither one writes to the project.

🤖 Generated with Claude Code

https://claude.ai/code/session_013tdBJBzhf47finAWKtDQVF

Generated by Claude Code