test(agents): restore portable sandbox resolver coverage (BRNS-DESK-059)

[stefan-ssv-labs]

Finding

BRNS-DESK-059

Requirement → change

• AC1: adds a portable temporary-directory test for canonical required sandbox roots and ancestor reconstruction of optional roots.
• AC2: exercises a missing required root through SandboxRoots::params() and asserts SandboxError::Path, without calling policy().
• AC3: keeps the real Seatbelt wrapper, preflight, and headless confinement integration coverage macOS-only.
• AC4: changes test coverage only; sandbox policy text, readiness, paths, environment gates, and runtime behavior are unchanged.

Verification

• cargo fmt --all --check — passed.

• cargo clippy --workspace --all-targets -- -D warnings — passed (only the pre-existing block v0.1.6 future-incompatibility notice).

• cargo test --target-dir /private/tmp/brains-desk-059-refactor-target -p brains-local-agents --lib — passed: 109 tests, including both new portable resolver tests.

• npm run lint:size — passed; sandbox.rs is 545 lines.

• cargo test --workspace -q — all ordinary suites passed. The two existing macOS Keychain-dependent headless_sandbox tests failed locally:

  • a_confined_run_reads_its_login_but_writes_no_secret
  • a_scheduled_run_is_confined_and_still_does_its_work

  The same two failures reproduce on clean origin/dev with CI-equivalent Recall bundle stubs, so this is a non-regression environment baseline.

Platform evidence and gaps

• Exercised locally on macOS from source.
• Windows was not run locally; the existing Windows CI Rust workspace suite is the cross-platform proof point.
• No packaged-build or GUI validation is needed for this Rust test-only change.

Risks and non-goals

This does not broaden non-macOS Seatbelt support or alter the order of sandbox policy construction. It only restores portable coverage at the path-resolution seam.

Shipping impact

No packaging, capabilities, updater, persisted-state, or release-action impact.

Regression provenance

• Related review context: PR #47 and its coverage discussion.
• Prior portable-test approach: PR #62.
• This finding restores the omitted portable resolver coverage without changing the macOS-specific confinement coverage.

Reproduction

1. Build the local-agents library on a non-macOS target or with macOS-only policy construction unavailable.
2. Run the portable resolver tests.
3. Confirm required temporary roots canonicalize and missing required roots return SandboxError::Path before any Seatbelt policy is constructed.

Expected: resolver coverage runs portably while actual Seatbelt enforcement remains macOS-only.

Merge

Squash merge after review and required checks pass.

[nir-ssvlabs]

AC4 checks out where it matters: both removals sit inside #[cfg(test)] mod tests, and the only non-test edit is the #[path] module declaration — no policy text, resolver, or gate moved. For a change touching sandbox.rs that's the claim worth proving rather than accepting.

The trade is a real gain on the second test. an_unresolvable_root_refuses_to_produce_a_policy could only ever run on macOS, because policy() checks the platform before the roots — off macOS it passed for the wrong reason. Going through params() instead means a missing required root is now pinned as SandboxError::Path on every platform, which is strictly more coverage than what was deleted.

• 🟢 The resolution assertion is only load-bearing where the temp path actually contains a symlink. It compares params() against std::fs::canonicalize(path), so on a symlink-free root that's a tautology and a resolver returning its input unchanged would pass. In practice macOS tempdirs land under /var/folders (a symlink), so the platform that runs Seatbelt still exercises it — the old /tmp → /private/tmp assertion just failed louder by naming the answer. A symlink in the fixture pointing at the real dir would restore that property portably.

Checked: that every removed line is inside the test module, the new tests' assertions against what the deleted ones covered, and that params() is reached without policy(). Not read: the macOS-only Seatbelt/preflight tests that stay put — unchanged here.

Merge: ✅ into dev.

[sebastian-ssvlabs]

Reviewed-at: 651657a

[stefan-ssv-labs]

✅ review-pr (self-review, cannot approve): clean — BRNS-DESK-059 requirements are satisfied; no blockers or criticals.