tests: unblock three stale assertions broken on main (MLX CI + Backend CI)

[danielhanchen]

Summary

MLX CI on Mac M1 and Backend CI have been red on every push to main for days. The underlying code is fine. Three test files have stale anchors / assertions left behind by recently-landed PRs.

This PR fixes all three so PR-time CI on every open PR (including #5743) goes back to green without rebasing past these regressions.

1. MLX CI — tests/studio/run_real_mlx_smoke.py:393

PR #5537 bumped max_steps from 7 to 30 to make the convergence gate seed-robust but left the assertion at == 7. With logging_steps = 1, the callback fires once per step, so 30 entries, not 7. Failure since 2026-05-18:

AssertionError: expected 7 logged steps, got
  [10.5475, 5.4663, 3.1851, 1.6394, ..., 0.0]   (30 entries)

Fix: track config.max_steps so the gate auto-follows future bumps.

2. Backend CI / Repo tests (CPU) — tests/studio/test_composer_rtl_bidi_attribute.py:29

PR #5775 changed studio/frontend/src/components/assistant-ui/thread.tsx:537 from the literal

aria-label="Message input"

to a JSX ternary

aria-label={overlay ? "Image edit instructions" : "Message input"}

but the test still anchors on the old literal form, so src.find(...) returns -1:

AssertionError: anchor 'aria-label="Message input"' not found
assert -1 != -1

Fix: anchor on the inner string literal "Message input".

3. Backend CI / Python 3.10-3.13 — studio/backend/tests/test_desktop_auth.py:487

test_provision_desktop_auth_writes_secret_and_creates_db_without_backend_deps installs a builtins.__import__ guard that blocks ("auth", "fastapi", "structlog", "utils") to prove the CLI does not pull in Studio backend deps. The guard is too broad: it matches the relative from .utils import echo inside typer._click.decorators (typer 0.25+), which calls __import__("utils", ..., level=1):

File ".../typer/_click/decorators.py", line 5, in <module>
File "<string>", line 13, in guarded_import
ModuleNotFoundError: utils

Fix: gate the block on level == 0 so only absolute imports of auth / fastapi / structlog / utils get rejected; relative imports inside third-party packages pass through. This preserves the test's intent (no Studio backend deps) without misclassifying every from .utils import x everywhere on sys.path.

Verification

$ pytest tests/studio/test_composer_rtl_bidi_attribute.py -v
12 passed in 0.03s

$ pytest studio/backend/tests/test_desktop_auth.py::test_provision_desktop_auth_writes_secret_and_creates_db_without_backend_deps -v
1 passed in 0.69s

The MLX one cannot be exercised on Linux but the change is mechanical (7 -> config.max_steps); it will be validated by MLX CI on Mac M1 on this PR.

Test plan

• MLX CI on Mac M1 goes green on this PR
• Backend CI (Repo tests CPU + all Python matrix variants) goes green
• No new regressions in any other workflow

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[gemini-code-assist]

Code Review

This pull request introduces test fixes and improvements across several test files. In test_desktop_auth.py, the guarded_import helper was updated to only block absolute imports (where level == 0), preventing relative imports within third-party packages from being incorrectly blocked. In run_real_mlx_smoke.py, the hardcoded assertion for logged steps was replaced with a dynamic check against config.max_steps. Lastly, in test_composer_rtl_bidi_attribute.py, the search anchor for the main composer test was updated to target the inner string literal "Message input" instead of the full JSX attribute. No review comments were provided for these changes.