test(cli): make code-pack + setup-scip tests host-agnostic

[theagenticguy]

Summary

Two @opencodehub/cli tests baked in host-environment assumptions and failed on macOS (they pass on Linux CI, so they were latent). Surfaced while validating an unrelated dep bump on a Mac. Test-only — no production code changes.

The two fixes

1. code-pack — "resolves a relative repo path against process.cwd()"
The test asserted opts.repoPath === resolve(cwd) where cwd came from mkdtemp(tmpdir(), ...). On macOS tmpdir() returns /tmp/..., but /tmp is a symlink to /private/tmp, so after process.chdir(cwd) the production code resolves against process.cwd() → /private/tmp/.... The assertion compared /private/tmp/... (actual) vs /tmp/... (expected) and failed.
→ Read process.cwd() back after chdir and assert against that — exactly what the production code sees.

2. setup-scip — "installs a single tool via injected fetch + allowPlaceholder"
The test hard-coded a { os: "linux", arch: "x64" } platform pin (comment even said "the test runs on AL2023 which is already linux-x64"). On darwin/arm64 the downloader's detectPlatform() found no matching platform entry, installed 0 tools, and result.installed.length was 0 !== 1.
→ Build the pin from detectPlatform() so it matches whatever host runs the test — linux-x64 in CI, darwin-arm64 on a dev Mac.

Why now

Both failures reproduce on a clean main checkout in a macOS sandbox (verified against the clean origin/main worktree). They're invisible in CI (Linux, no /tmp symlink, host is linux-x64), so they only bite local Mac development. This removes the host assumption.

Verification

• pnpm --filter @opencodehub/cli test — 256/256 pass (was 254 pass / 2 fail)
• Full pre-push gate (pnpm -r test recursive + typecheck + verdict) — green
• Banned-strings patterns checked manually against the diff (clean); local hook can't run (needs bash 4+, macOS ships 3.2)

Test plan

• Both tests pass on macOS arm64
• Logic preserves the linux-x64 path (CI host) via detectPlatform()
• No production code touched