test(installer): lock repo-root resolution in bash launcher scripts

[dislovelhl]

Why this matters

#1332 fixed the off-by-one repo-root bug in start-agent-ui.sh and build-ui-installer.sh (the bash counterparts to the PowerShell fix in #1326), but shipped with only a manual bash -n test plan — nothing in CI stops the $SCRIPT_DIR/.. → installer/ regression from creeping back in if either script is moved or edited again. #1328 is still open precisely because the fix was never pinned.

This adds the regression coverage: a pure-stdlib test that parses both scripts and asserts the two-levels-up traversal resolves to the real repo root, that $WEBUI_DIR/$ELECTRON_DIR point at src/gaia/apps/webui and src/gaia/electron (which exist on a fresh checkout), and that the old one-level traversal would have landed on installer/. No production code changes — the fix is already on main.

Closes #1328.

Test plan

• python -m pytest tests/unit/installer/test_bash_script_paths.py — 3 passed against the current (fixed) scripts.
• Sanity: temporarily reverting either script to $SCRIPT_DIR/.. makes the corresponding test fail (the guard actually guards).
• black --check and isort --check-only clean on the new file.

Out of scope

• The script fix itself (already merged in fix(installer): resolve repo root correctly in bash scripts under installer/ #1332).
• PowerShell counterparts (fix(installer): resolve repo root correctly in scripts under installer/ #1326).

[github-actions]

Review

Approve. This is a tight, well-scoped regression test that pins the $SCRIPT_DIR/../.. two-level traversal in both installer bash launchers, closing the CI gap #1332 left open. I ran all three tests against the live scripts (installer/scripts/start-agent-ui.sh:50, build-ui-installer.sh:15) and confirmed they pass, and verified the guard actually fails on the old one-level traversal — reverting either script to $SCRIPT_DIR/.. makes project_root == REPO_ROOT fail, exactly as the test plan claims.

No production code changes, no security surface, no new dependencies (pure stdlib re + pathlib). No prompt-injection content in the diff.

Issues

🟢 Minor — the first test is documentation, not a guard (test_bash_script_paths.py:45)
test_old_one_level_bash_traversal_would_resolve_to_installer_dir never reads either script — it just asserts installer/scripts/.. == installer/, which is always true by directory structure. It nicely documents the regression, but the real guarding is done by the two path-resolution tests (which read the actual ../.. traversal out of the scripts and assert it lands on the repo root). Worth a one-line docstring note that this case is illustrative, or fold it into the other two — not blocking either way.

🟢 Minor — regex parsing is tight to script formatting (test_bash_script_paths.py:25,35)
The ^VAR="$(cd "$SCRIPT_DIR/..." && pwd)"$ patterns assume exact quoting/spacing, so a benign reformat of the scripts (e.g. a trailing comment or alternate path expression) would trip the assert match, "... does not assign ..." branch rather than the value assertion. That's an acceptable trade-off for a pinning test, and the assert messages are clear about what failed — just be aware future script edits may need the regex touched alongside them.

Strengths

• Pure-stdlib, deterministic, runs anywhere a checkout exists — no Lemonade/network/fixtures needed; correctly placed under tests/unit/installer/ next to the existing installer tests.
• Asserts against the real checkout (WEBUI_DIR.is_dir(), ELECTRON_DIR.is_dir() at src/gaia/apps/webui and src/gaia/electron), so it catches both a traversal regression and a directory move.
• Helper functions (_script_assignment / _literal_assignment / _resolved_from_scripts_dir) keep each test readable and the failure messages actionable.
• PR description is exemplary: leads with the why (the [Bug]: bash installer scripts (start-agent-ui.sh, build-ui-installer.sh) have the same off-by-one path bug as #1325 #1328 gap), states what's out of scope, and the test plan is verifiable.

Verdict

Approve. Both notes are 🟢 minor and optional — nothing blocks merge.