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

[JArmandoAnaya]

The auto-labeling feature is stable, so this removes what the slices that built it left behind. cf. #464, which carries the full inventory and the items deliberately not acted on.

No behaviour changes. Every item is a removal with zero references, a test deletion with a named surviving test proving the same failure mode, or a comment correction where the code is the authority.

Redundant tests

Three tests in tests/inference/test_provider.py (#434) were re-created in tests/inference/test_providers.py (#458) when the family resolver got its own file, and the originals were never removed — the first pair share a name character for character.

Deleted	Surviving test	Mutation that reddens the survivor
test_provider.py::test_an_http_connection_is_refused_because_this_build_has_no_adapter	test_providers.py::test_an_http_connection_is_refused_because_this_build_has_no_adapter	A
test_provider.py::test_a_local_connection_without_weights_is_refused_by_state	test_providers.py::test_a_connection_without_weights_is_told_which_action_fixes_it	B
test_provider.py::test_the_connections_own_state_is_reported_before_the_machines	test_providers.py::test_a_missing_runtime_is_reported_after_the_connections_own_state	B

In each pair the survivor is the stronger test: it drives provider_for through a real InferenceConnectionService rather than a hand-built model, and matches a longer substring of the message.

The third is worth naming separately, because it did not test what its name claimed. It asserted that a not-set-up connection raises InferenceConnectionNotSetUp — which is the second row's assertion — while stubbing nothing, so there was no second refusal for the order to be against. Its survivor stubs require to raise and then asserts both directions.

test_provider.py keeps everything unique to it: the unstubbed missing-extra refusal, the isinstance-against-Protocol check, both model_ref cases, and the whole regions_from half — including #434's named mutation test, test_duplicate_detections_are_suppressed_before_they_leave_the_adapter, which is untouched. Its module docstring now says which half of provider_for lives where.

Deleting the three left remote() and two error imports with no users; those went with them.

Mutation verification

Committed before the first mutation. Each anchor asserted present exactly once before applying and its replacement asserted after; each reverted by git apply -R on its recorded patch, with git status --porcelain empty afterwards.

Mutation A — provider_for's HTTP arm returns _local(...) instead of raising InferenceConnectionNotRunnable.

FAILED tests/inference/test_providers.py::test_an_http_connection_is_refused_because_this_build_has_no_adapter

Mutation B — the setup_state is not READY guard deleted from _local.

FAILED tests/inference/test_providers.py::test_a_connection_without_weights_is_told_which_action_fixes_it
FAILED tests/inference/test_providers.py::test_a_missing_runtime_is_reported_after_the_connections_own_state
FAILED tests/inference/test_providers.py::test_a_refused_connection_leaves_nothing_behind_for_the_next_request

Both named survivors turn red, and B catches a third test the deletions never covered.

Dead code

• inference/sam_provider.py — best_of(iou_scores, masks) never read masks; the choice comes entirely from the IoU head's scores. Flagged by ruff --select ARG. Three call sites, all in this repository. Not part of the ModelProvider port and not in visionset.inference.__all__.
• frontend/ui-core/src/data/inferenceQueries.ts — ConnectionAction had no importers. Its three neighbours on the adjacent lines are each re-exported from ui-core/src/index.ts; this one was not, so it was not package API either. Anything needing it has it through Connection["allowed_actions"].

Comments that no longer described the code

• pyproject.toml — "uv.lock does not carry extras". It carries them twice: local-inference under [package.optional-dependencies], and one requires-dist row per member with an extra == marker, plus provides-extras. The half of the sentence that was true is kept and sharpened: a plain uv sync does not install the extra — that takes --extra local-inference — but editing those lines without re-locking leaves the lockfile stale and uv lock --check fails on it.
• AnnotationPage.tsx — onConfigureInference was documented as "expected to be absent for now: the Inference surface waits on Inference section — user-configured model connections #421's open rail question". That surface shipped in feat(app): Inference section, rail entry, download-size surface (#421 slice 2) #454, and frontend/app/src/routes.tsx:331 passes the prop. The prop stays optional for the real reason — ui-core imports no router and cannot know whether its host has such a screen.
• inference/__init__.py and tests/architecture/test_optional_runtime.py — two ragged docstring wraps left by the edit that added torchvision to the enumerations. ruff does not reflow docstrings, which is why they passed every gate.

Found, not fixed

• The suggest route's detail field is accepted, published, and ignored #463, filed — POST /inference/suggest accepts a detail field, validates it, publishes it in openapi.json, and ignores it. suggest() takes the parameter and never reads it; the value that reaches polygon_from is the adapter's construction-time default, and providers._local() never passes one. So detail: 0.9 and detail: 0.01 produce the same polygon. Not a cleanup item in either direction: deleting the field removes a published one, and making it work means carrying it per-request, because providers are pooled and shared across calls. No test covers the field reaching anything, which is why it shipped.
• Six items recorded HIGH on Auto-labeling cleanup: remove what the fast slices left behind #464 and deliberately untouched — three unimported names in visionset.inference.__all__ (removing a name from __all__ is a public-API change whatever the reference count says); masks.spans(), which could be derived from masks.runs() but is not an identical helper and would edit the Point-prompted suggestions trace the topmost-leftmost blob, not the one under the click #461 fix's fresh code; the one-letter gap between test_provider.py and test_providers.py, since a rename is not a removal; and every protected test — the mutation-verification tests named in feat(inference): InferenceConnection entity, persistence, wire, REST, CLI (#421 slice 1) #430/feat(inference): ModelProvider port, local adapter, weight download (#418 slice 2) #434/feat(inference): SAM 2 point-prompt adapter and suggest route (#424 slice 3a) #448/feat(annotator): interactive suggest tool with ephemeral preview (#424 slice 3b) #451/feat(app): Inference section, rail entry, download-size surface (#421 slice 2) #454/fix(inference): sam2_video is a point-prompt model, and an unknown type refuses #458, fix(inference): sam2_video is a point-prompt model, and an unknown type refuses #458's refusal tests, the Point-prompted suggestions trace the topmost-leftmost blob, not the one under the click #461 blob-selection tests, and the contract and boundary suites.

What the sweep did not find

No TODO/FIXME/HACK markers on any feature surface, no commented-out code, no .tmp/.bak/spike-residue files tracked or untracked, no unreachable branches. ruff --select ERA,ARG,PIE,RUF,F401,F841,B007 over every backend feature module reported exactly the two unused parameters — one of which is #463 — and nothing else. src/visionset/mcp/ has no inference surface at all.

Gate

Every stage run in this worktree, exit code recorded. Split by directory because of the ~10-minute command ceiling on this machine; the split is ls tests/ at run time, and tests/scripts is node --test, covered by the frontend stage.

Stage	Command	Exit
pytest A	pytest tests/{architecture,cli,examples,fixtures,formats,inference,jobs,mcp,packaging,test_versioning.py}	0 — 988 collected
pytest B	pytest tests/kernel	0 — 1381 collected
pytest C	pytest tests/server	0 — 628 collected
lint	ruff check .	0 — all checks passed
format	ruff format --check .	0 — 341 files already formatted
types	mypy src/visionset	0 — no issues in 150 source files
imports	lint-imports	0 — 4 contracts kept, 0 broken
lock	uv lock --check	0 — the comment corrected above, verified
frontend	scripts/check.sh frontend	0 — annotator 904 passed, ui-core 806 passed, app build, node --test script gates
generated	scripts/check.sh generated	0 — openapi, generated client and MCP tool reference: no drift
browser	scripts/check.sh browser	0 — e2e 233 passed, cycle 1 passed (ports 18411/20459/22507)

Counts rather than pass lines for the pytest stages: addopts = "-q" in pyproject.toml plus the explicit -q makes -qq, which suppresses the summary line. The exit codes are the record.

Closes #459

Closes #464