fix(tests): align two stale tests in #1259 with current implementation

[richgoodson]

Summary

Two of the failing tests called out in #1259 are stale against the
current implementation, not regressions in the code under test. Both
were already failing on main for me on a fresh editable install; this
PR brings them back in line with the code they're meant to cover.

Pairs with PR #1268 (test_admin_profiles_api) and #1286
(test_settings) to fully clear the #1259 list.

Changes

tests/test_mlx_lm_mtp_patch.py::TestBatchGeneratorDispatch::test_is_mtp_eligible_requires_mtp_forward_and_solo_batch

The test was written against the pre-#1247 contract where head presence
implied MTP active. Commit
23ca7dc decoupled head
attachment from inference-time MTP for the VLM load path and added an
is_mtp_active() gate inside _is_mtp_eligible. The True-expected
assertion now fails because the process-wide flag is off.

Fix: toggle mlx_lm_mtp.set_mtp_active(True) around the True-expected
assertions, restore the prior value in finally, and add a new
"head attached but flag off" case to lock in the post-23ca7dc
semantics (this is the exact case the VLM runtime patches rely on).

tests/test_vlm_torch_free_image_processor.py::test_patch_wraps_target_processors

The test stubbed the wrong module path and class name for dots_ocr:

	test stubbed	code imports
module	mlx_vlm.models.dots_ocr.processing	mlx_vlm.models.dots_ocr.processing_dots_ocr
class	DotsOcrProcessor	DotsVLProcessor

The fake_import branch never matched, the dots branch silently fell
through to the except (ImportError, AttributeError) clause inside
_patch_torch_free_image_processor, and the wrap-marker assertion
failed. The mismatch has been there since the test was added in
a1987ed.

Fix: align the stubs with the actual (module_path, cls_name) tuples
in omlx/engine/vlm.py:181-184.

Testing

Before:

$ pytest tests/test_mlx_lm_mtp_patch.py tests/test_vlm_torch_free_image_processor.py
... 2 failed, 52 passed

After:

$ pytest tests/test_mlx_lm_mtp_patch.py tests/test_vlm_torch_free_image_processor.py
... 54 passed in 0.21s

Apple Silicon M5, Python 3.13.12, fresh editable install on
51907f0.

Scope

Test-only changes. No code under omlx/ modified.

Refs #1259.

🤖 Generated with Claude Code

[jundot]

Verified both fixes against current code. _is_mtp_eligible gates on is_mtp_active() at omlx/patches/mlx_lm_mtp/batch_generator.py:222-224 (added in 23ca7dc), and _patch_torch_free_image_processor does import processing_dots_ocr.DotsVLProcessor at omlx/engine/vlm.py:183. The new "head attached but flag off" case is a nice lock-in for the post-23ca7dc semantics. Merging.