chore(inference): remove what the fast auto-labeling slices left behind

frontend/ui-core/src/annotator/AnnotationPage.tsx

@@ -372,10 +372,11 @@ export interface AnnotationPageProps {
    * Where somebody goes to set up a model connection, if the app has such a
    * screen (#424, D6).
    *
-   * Optional, and it is expected to be absent for now: the Inference surface
-   * waits on #421's open rail question, and `ui-core` imports no router. Absent,
-   * the suggest tool's panel still says what is missing and simply renders no
-   * control — a host that cannot honour one renders none rather than a dead one.
+   * Optional because `ui-core` imports no router and cannot know whether its
+   * host has such a screen — the app does, and wires this to the Inference
+   * section. Absent, the suggest tool's panel still says what is missing and
+   * simply renders no control: a host that cannot honour one renders none
+   * rather than a dead one.
    */
   readonly onConfigureInference?: () => void;
 }

frontend/ui-core/src/data/inferenceQueries.ts

@@ -93,7 +93,6 @@ export interface SuggestionOut {
   readonly region?: SuggestedRegion | null;
 }
 
-export type ConnectionAction = components["schemas"]["ConnectionAction"];
 export type ConnectionType = components["schemas"]["ConnectionType"];
 export type ConnectionSetupState = components["schemas"]["ConnectionSetupState"];
 export type DownloadSizeOut = components["schemas"]["DownloadSizeOut"];

pyproject.toml

@@ -63,8 +63,12 @@ dependencies = [
 #
 # The canonical spelling is `local-inference`; it supersedes the earlier
 # `autolabel` framing. Version floors rather than pins: this is a user's
-# environment and their CUDA build decides most of it. `uv.lock` does not carry
-# extras, so nothing here is resolved by a `uv sync` in this repository.
+# environment and their CUDA build decides most of it. `uv.lock` does record the
+# extra — once as a member list, once per member under `requires-dist` — so
+# editing these lines without re-running `uv lock` leaves it stale and
+# `uv lock --check` fails on the result. What a plain `uv sync` does not do is
+# *install* it: that takes `--extra local-inference`, which is how two gigabytes
+# of CUDA wheels stay out of a developer's environment by default.
 #
 # `torchvision` is here because SAM 2's image processor imports it, and nothing
 # else in this list brings it: `Sam2ImageProcessor` refuses to construct without

src/visionset/inference/__init__.py

@@ -14,11 +14,10 @@
 **Importing this package imports nothing heavy.** Every reference to torch,
 transformers, accelerate and huggingface_hub is inside a function, and
 torchvision — which nothing here names, and ``transformers`` imports for us — is
-held to the same line — see
-``_extra`` for why that is load-bearing rather than tidy — so a base install
-starts a server, runs a worker and imports this module without the optional
-runtime present. ``tests/architecture/test_optional_runtime.py`` proves it in a
-fresh interpreter.
+held to the same line. See ``_extra`` for why that is load-bearing rather than
+tidy: a base install starts a server, runs a worker and imports this module
+without the optional runtime present, and
+``tests/architecture/test_optional_runtime.py`` proves it in a fresh interpreter.
 
 **There is no plugin registry, and resolution happens in two steps.** #418's
 recorded decision is that adapters are instantiated from user-created model

src/visionset/inference/sam_provider.py

@@ -87,7 +87,7 @@ def points_and_labels(prompt: PointPrompt) -> tuple[list[list[float]], list[int]
     return points, labels
 
 
-def best_of(iou_scores: list[float], masks: list[Any]) -> tuple[int, float]:
+def best_of(iou_scores: list[float]) -> tuple[int, float]:
     """Which of the multi-mask answers to offer, and how sure it is.
 
     A segmenter of this family answers a single click with several masks at
@@ -256,7 +256,7 @@ def _regions(
             outputs.pred_masks, original_sizes=[list(size)], binarize=True
         )[0]
         scores = [float(value) for value in outputs.iou_scores.flatten().tolist()]
-        chosen, confidence = best_of(scores, lifted)
+        chosen, confidence = best_of(scores)
         if confidence < minimum_confidence:
             return ()
         mask = lifted.reshape(-1, *lifted.shape[-2:])[chosen]

tests/architecture/test_optional_runtime.py

@@ -4,10 +4,9 @@
 huggingface_hub in the `local-inference` extra, and `visionset/inference/_extra.py`
 reaches every one of them from inside a function — torchvision by never naming it
 at all, since it is `transformers` that imports it. Neither of those is
-self-enforcing: a single
-module-level `import torch` anywhere under `visionset.inference` would leave the
-metadata unchanged and make roughly two gigabytes of CUDA wheels a condition of
-starting the server.
+self-enforcing: a single module-level `import torch` anywhere under
+`visionset.inference` would leave the metadata unchanged and make roughly two
+gigabytes of CUDA wheels a condition of starting the server.
 
 It is a condition of *starting* rather than of running a model because
 `visionset.jobs` imports its handler modules at package import to populate the

tests/inference/test_provider.py

@@ -1,9 +1,12 @@
-"""The composition root's refusals, and the adapter's tensor-to-domain half.
+"""What the runtime's presence decides, and the adapter's tensor-to-domain half.
 
-Two subjects, both reachable without the optional runtime and without a GPU:
+Two subjects, both reachable without a GPU:
 
-- `provider_for`, which answers "can this connection predict here?" and must
-  refuse with a sentence rather than a `None` or a stack trace;
+- the half of `provider_for` that turns on whether the optional runtime is
+  installed here — the refusal that names the install command when it is not,
+  and structural conformance to the port when it is. The refusals that turn on
+  the *connection* instead are `test_providers.py`'s, beside the resolution
+  they belong to;
 - `regions_from`, the conversion the adapter does after a forward — the one part
   of running a model that can be wrong in a way no weights are needed to see.
 
@@ -35,8 +38,6 @@
     TextPrompt,
 )
 from visionset.kernel.errors import (
-    InferenceConnectionNotRunnable,
-    InferenceConnectionNotSetUp,
     LocalInferenceUnavailable,
     UnsupportedPrompt,
 )
@@ -59,58 +60,9 @@ def local(
     )
 
 
-def remote() -> InferenceConnection:
-    return InferenceConnection(
-        name="remote",
-        connection_type=ConnectionType.HTTP,
-        model_id="some/model",
-        model_revision="abc123",
-        endpoint_url="https://example.invalid/predict",
-        setup_state=ConnectionSetupState.READY,
-    )
-
-
 # --- resolving a connection to something that can answer ----------------------
 
 
-def test_a_local_connection_without_weights_is_refused_by_state(tmp_path: Path) -> None:
-    """And the message names the action that fixes it.
-
-    "Not set up" alone tells an operator what they already knew; the remedy is
-    the point, and it is a real one — the identical call succeeds after the
-    download.
-    """
-    with pytest.raises(InferenceConnectionNotSetUp) as raised:
-        provider_for(local(), workspace_root=tmp_path)
-    assert "download_weights" in str(raised.value)
-
-
-def test_an_http_connection_is_refused_because_this_build_has_no_adapter(
-    tmp_path: Path,
-) -> None:
-    """A different refusal from the one above, deliberately.
-
-    No state change fixes this and no wait helps — the adapter that would speak
-    to an endpoint is a later slice — so it must not be the error whose whole
-    meaning is "change the state and resubmit".
-    """
-    with pytest.raises(InferenceConnectionNotRunnable) as raised:
-        provider_for(remote(), workspace_root=tmp_path)
-    assert "http" in str(raised.value)
-
-
-def test_the_connections_own_state_is_reported_before_the_machines(tmp_path: Path) -> None:
-    """Order of the two checks, and it matters to whoever reads the answer.
-
-    A not-set-up connection on a machine with no extra has two problems. Telling
-    somebody to run a `pip install` when what they needed was a download sends
-    them to the wrong place; the state is the one they can act on from where they
-    are standing.
-    """
-    with pytest.raises(InferenceConnectionNotSetUp):
-        provider_for(local(), workspace_root=tmp_path)
-
-
 @pytest.mark.skipif(EXTRA_INSTALLED, reason="the local runtime is installed here")
 def test_a_ready_connection_without_the_runtime_names_the_install_command(
     tmp_path: Path,

tests/inference/test_sam_provider.py

@@ -94,11 +94,11 @@ def test_a_prompt_with_no_negatives_still_labels_every_point() -> None:
 
 def test_the_highest_scoring_mask_is_the_one_offered() -> None:
     """A click is ambiguous about scale; the model's own IoU head is what resolves it."""
-    assert best_of([0.2, 0.91, 0.5], [[], [], []]) == (1, 0.91)
+    assert best_of([0.2, 0.91, 0.5]) == (1, 0.91)
 
 
 def test_a_score_outside_the_domains_bounds_is_clamped_rather_than_refused() -> None:
-    assert best_of([1.0000001], [[]])[1] == 1.0
+    assert best_of([1.0000001])[1] == 1.0
 
 
 # --- what it refuses ----------------------------------------------------------