test(resume): pin XDG_DATA_HOME in ScopedHome to sandbox opencode

[bdelanghe]

ScopedHome pins $HOME and $TOOLPATH_CONFIG_DIR, which sandboxes every reader that derives its paths from $HOME. The opencode PathResolver does not: it prefers $XDG_DATA_HOME when locating opencode.db.

On a machine that sets $XDG_DATA_HOME — common on Linux, and set on this dev box — the projector goes to the real user database. The test fails with table … already exists, and, worse, mutates live opencode data on the way.

Pinning it alongside the other two keeps every harness inside the sandbox.

Verified

file_input_explicit_opencode_projects_and_records_exec, with $XDG_DATA_HOME exported:

before    test result: FAILED. 9 passed; 1 failed
after     test result: ok.    10 passed; 0 failed

Provenance

This is the commit from #142, which was closed unmerged when the bdelanghe/toolpath fork was retired. The work was preserved beforehand as tag archive/fork/fix-opencode-test-xdg-isolation; this branch is that commit cherry-picked onto current main, unchanged apart from the rebase.

Relationship to #251

Both touch crates/path-cli/tests/support/mod.rs, and both close the same class of hole — an env var that outranks the sandboxed $HOME. They are independent fixes for different variables and different readers, so they are separate PRs; whichever lands second will want a trivial rebase, as they add adjacent fields to ScopedHome.

Together they take cargo test --workspace on a developer machine from 29 failures to 0. Neither is visible in CI, which sets neither variable.

---

1. Independent PR — no bundled or speculative changes
   • Verified: one commit, one file, one concern — $XDG_DATA_HOME escaping the test sandbox. Test-support code only.
2. Changed codepaths verified — targeted unit and full integration tests
   • Verified: cargo test -p path-cli --test resume 10/10 with the fix, 9/10 without it on the same box and the same shell; cargo fmt --all --check clean
3. Root cause identified — every failure traced to source
   • The opencode PathResolver prefers $XDG_DATA_HOME over $HOME; ScopedHome pinned only the latter, so an exported value escaped the sandbox and reached the real opencode.db
4. No duplication — refactoring preferred over copy/paste
   • Follows the save/restore pattern already in ScopedHome for $HOME and $TOOLPATH_CONFIG_DIR rather than adding a parallel mechanism
5. No unrelated changes — housekeeping isolated to its own branch
   • The $CLAUDE_CONFIG_DIR escape is test(cli): the sandbox holds when CLAUDE_CONFIG_DIR is set #251; the clippy sweep is chore(clippy): the lints newer toolchains enforce #252

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

🔍 Preview deployed: https://16c1cf45.toolpath.pages.dev