The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).

The first CI run on this PR (https://github.com/jundot/omlx/actions/runs/26738412492)
surfaced three failures introduced by the rebase against upstream's
cb0904d (predictive prefill throttle) and our own change to
``Scheduler.__init__`` (constructs a MemoryMonitor in estimator-only
mode so preflight checks work without prior set_model_info):

1. ``_set_model_info_for_monitor`` accepted MagicMock-style proxies
   as ``head_dim`` / ``num_kv_heads`` / ``num_layers`` because the
   ``if num_layers and num_kv_heads and head_dim:`` truthiness gate
   doesn't reject auto-attribute mocks. The estimator path
   (``estimate_chunk_transient_bytes`` → ``if hd > 128``) then
   blew up the upstream throttle tests with
   ``TypeError: '>' not supported between instances of 'MagicMock'
   and 'int'``. Tighten to ``isinstance(..., int) and ... > 0`` so
   the estimator stays disabled until set_model_info is called with
   genuine model dims.

2. ``test_schedule_waiting_preflight_rejection_releases`` patched
   ``_preflight_memory_check`` to return a plain ``str`` (matching
   the pre-PR contract). Our PR types the return as
   ``_PreflightRejection`` and the rejection-path code in
   ``_schedule_waiting`` reads ``.message`` on it. Update the mock to
   return an actual ``_PreflightRejection`` so the test exercises the
   real shape.

3. ``test_picks_text_config_over_top_level_vision_dims`` (added by
   our merged jundot#1448) asserted ``sched.memory_monitor is None`` at
   init. Our PR now constructs the monitor in estimator-only mode at
   init time; replace the assertion with the actual setup it needs
   (overwrite ``memory_monitor`` with a MagicMock to inspect
   ``set_model_info`` calls).