## Problem

Five checks fail on main, so the `Repo tests (CPU, studio)`, `Repo tests
(CPU, rest)` and `workflow-trigger lint` shards are red on every open PR
(for example #14472, #14482, #14488, #14490 and #14494). All five
reproduce on main `4d7ebf5cea`.

| Test | Introduced by | Fault |
|---|---|---|
|
`tests/security/test_lint_risky_loaders.py::test_the_tree_matches_the_baseline`
| #14437 | Baseline. Two new dynamic-import sites: `import_module` in
`studio/backend/core/rag/embeddings.py`, and a third
`import_from_string` in `unsloth/models/sentence_transformer.py`
(`_require_word_tokenizer`). Both are safe, see below. |
|
`tests/studio/test_tauri_branding_contract.py::test_desktop_surfaces_do_not_restore_studio_branding`
| #14473 | Product. "Studio" became "Unsloth Studio" in the Markdown
import errors and the Tauri webview reject message. The desktop contract
allows "Unsloth" there, never "Unsloth Studio". "Unsloth" meets both
#14473's rule (no bare "Studio") and this one. |
|
`tests/studio/test_usage_examples_model_source_contract.py::test_auto_switch_section_sits_above_the_usage_examples`
| #14455 | Test. The redesign moved the usage examples above the monitor
link on purpose, and its own `api-settings-tab.test.ts` pins that. |
|
`tests/studio/test_usage_examples_model_source_contract.py::test_monitor_can_unload_the_loaded_model`
| #14455 | Test. Unloading moved from a standalone button on the page
into the model picker's eject (`ApiModelLoadControls`). The `/unload`
path itself is unchanged. |
|
`tests/studio/test_settings_compact_overflow_contract.py::test_api_monitor_entries_and_expanded_text_can_shrink`
| #14455 | Test. Entry rows are now rounded cards (no `border-b`), and
the prompt/reply box's `rounded-lg` became conditional, so the class
literals the test matched are gone. The overflow guards themselves are
still in the source. |

## Fix

- **Risky loaders:** two baseline entries with reasons.
- `embeddings.py` imports only the three `sentence_transformers.models`
paths written out in a literal tuple, to wrap their `import_from_string`
with the trust gate.
- `_require_word_tokenizer` imports only after its caller has refused
any `tokenizer_class` outside `sentence_transformers.*`, and only dotted
identifiers.
- **Branding:** "Unsloth Studio" to "Unsloth" in those seven Markdown
import errors and the Rust reject message.
`conversation-markdown-import.test.ts` now matches `/Unsloth Markdown/`.
- **#14455 contracts:** rewritten to guard what the redesign keeps:
- the auto-switch section still follows the monitor link, and the usage
examples come before both;
- the page passes `unloadActiveModel` to the load controls as
`onUnloadActive`, and eject is withheld while `busy` (`modelLoading ||
unloading`, #11223's rule);
  - the `/unload` call still reads the internal id from status;
  - entry rows keep `w-full min-w-0`;
  - the prompt/reply box keeps `max-h-72 overflow-auto break-words`;
  - the model id keeps `min-w-0 break-all`.

## Verification

- The four Python files: 78 passed (5 failed on main).
- `studio/frontend/tests/conversation-markdown-import.test.ts`: 20
passed.
- Sabotage, each run against the rewritten tests, each failing exactly
one test:
- removing `min-w-0` from the entry row, or `break-words` from the text
box: the overflow test fails;
- dropping `unloading` from `busy`, or passing `onEject` while busy: the
unload test fails;
- putting the auto-switch section above the monitor link: the order test
fails.

…utoopt parity checks (#14522)

## Problem

#14473 renamed user-facing "Studio" to "Unsloth" in the backend
messages, but three tests still expect the old wording. They fail on
main and on open PRs (for example #14494 and #14488):

| Test | Expects | Product now says |
|---|---|---|
|
`test_sandbox_linux.py::test_an_interpreter_symlinked_out_of_the_workdir_fails_the_call`
| `Python that runs Studio` | `Python that runs Unsloth`
(`sandbox_linux.py`) |
|
`test_diffusion_attention.py::test_a_hidden_xformers_is_not_reinstalled_in_the_same_process`
| `restart Studio` | `restart Unsloth` (`diffusion_attention.py`) |
| `test_diffusion_autoopt_parity.py` (two tests) | `the Studio fast step
that replays decode steps from tensors`, 16 times in
`fixtures/diffusion_autoopt_parity.json` | `the Unsloth fast step ...`
(`diffusion_capture_safe.py`) |

The rename was intended, so the tests are what is out of date.

## Fix

Test only. Both assertions and the 16 recorded reasons in the parity
fixture now use "Unsloth". Those 16 strings were the only "Studio" in
the fixture. Nothing else in it changes.

## Verification

- `test_sandbox_linux.py` and `test_diffusion_attention.py`: both target
tests fail on main and pass with this change.
- The one remaining local failure,
`test_active_attention_backend_reads_tuple_return`, is an `ImportError`
from this host's `huggingface_hub` and fails the same way on main.
- `test_diffusion_autoopt_parity.py` (CPU torch 2.10, diffusers
0.41.0.dev0):
  - with main's fixture: 2 failed, 1 passed;
  - with this change: 3 passed.