fix(worker): register a blob for every asset the stub produces

[georgi]

Closes the class behind #1020 (video) and #1021 (audio).

This is one defect class, not four bugs

WorkerContext covers a ref type per entry point, not per type. Any factory it does not override falls through to ProcessingContext, which with no name returns a ref carrying inline data and an empty uri. executor._extract_outputs extracts blobs only for a blob:// uri, and _serialize_asset_ref strips every raw-bytes data field at any depth. The bytes are discarded and the node reports success — 114.6 s of A40 time for an empty mp4, in the case that started this.

So every *_from_* method added to ProcessingContext from now on is a silent data-loss bug in the worker until someone remembers to override it. The coverage test below is what stops that.

Changes

Entry point	Was	Now
image_from_io	ImageRef(data=…, uri="")	blob://
image_from_url	broken via image_from_io	fixed by the above — see finding 1
model3d_from_io	Model3DRef(data=…, uri="")	blob://
audio_from_segment (no name)	AudioRef(uri="memory://<uuid>", data=…)	blob://
video_from_io, audio_from_io	—	as in #1020 / #1021

Branch base. Cut from main as asked, so it also carries the video_from_io and audio_from_io overrides — without them the coverage test cannot be green here. If #1020 and #1021 merge first, rebasing shrinks this diff to image / model3d / segment plus the tests; the conflicts are adjacent method insertions in one file. Their test files are not duplicated here — the coverage test drives those funnels behaviorally.

audio_from_segment keeps the base implementation's WAV encoding and metadata; the override only re-homes the bytes when the base took its memory:// shortcut. That shortcut exists so a later in-process read can get the AudioSegment back without decoding. A worker output crosses a process boundary, so the host can never follow it — the fast path is not available to a worker at all, and the choice is between blob bytes and no bytes.

Finding 1 — image_from_url

Verified, not assumed. Its body is one line:

return await self.image_from_io(await self.download_file(url), name=name, parent_id=parent_id)

It downloads and funnels through image_from_io, so overriding image_from_io covers it. No separate override. tests/worker/test_context_stub_remaining_blob_gaps.py::test_image_from_url_funnels_through_image_from_io pins that with a stubbed download_file — no network.

Finding 2 — asset_to_data

A reader, not a producer. Not this bug, not fixed. It takes an existing AssetRef, and when that ref is a memory:// handle with no data it resolves the handle out of the in-process store and returns a copy with data filled in. It creates no new asset and no new bytes; it materializes bytes that already exist so a caller can read them.

The coverage test excludes it structurally rather than by name: a method with an AssetRef-typed parameter converts a ref, it does not produce output. Nothing is allowlisted.

The coverage test

tests/worker/test_context_stub_blob_coverage.py, two halves, because neither is enough alone:

1. Behavioral. Enumerate every *_from_io producer, drive each with a real BytesIO, and assert it returns a blob:// uri and that the bytes come back out of executor._extract_outputs. A guard test asserts the enumeration actually found a funnel for every type in ASSET_REF_TYPES, so it cannot pass by covering nothing.
2. Structural. Every producer on ProcessingContext — returns a type in ASSET_REF_TYPES, takes no ref parameter — must be overridden by WorkerContext or delegate (transitively, read off the source) to one that is. This is what catches a factory added tomorrow.

The structural half alone would have been worse than nothing: it follows delegation without seeing branches, so it marked audio_from_segment covered while its no-name branch dropped the audio. That is why the behavioral half exists, and the test's docstring says so.

Red / green

./.venv/bin/python -m pytest tests/worker/test_context_stub_blob_coverage.py tests/worker/test_context_stub_remaining_blob_gaps.py -q

Before (stub reverted to main, tests present):

E  AssertionError: audio_from_io returned uri '', so its bytes are dropped
E  AssertionError: image_from_io returned uri '', so its bytes are dropped
E  AssertionError: model3d_from_io returned uri '', so its bytes are dropped
E  AssertionError: video_from_io returned uri '', so its bytes are dropped
E  AssertionError: these produce an asset the worker cannot deliver — override them on WorkerContext
   so the bytes are registered as a blob: ['audio_from_base64', 'audio_from_bytes', 'audio_from_io',
   'audio_from_segment', 'image_from_io', 'image_from_url', 'model3d_from_io', 'video_from_bytes',
   'video_from_frames', 'video_from_io', 'video_from_numpy']
E  AssertionError: expected a blob uri, got ''
E  AssertionError: expected a blob uri, got ''
E  AssertionError: expected a blob uri, got ''
E  AssertionError: expected a blob uri, got 'memory://c57f0789-1b87-4fe6-b08d-6994a77aaca6'
9 failed, 1 passed in 0.09s

Every red is behavioral. The memory:// line is the audio_from_segment hole the structural check could not see.

After:

..........                                                               [100%]
10 passed in 0.06s

Whole worker suite: 225 passed in 9.25s. No network, no GPU, no model downloads — the one download is monkeypatched, and the AudioSegment is 20 ms of generated silence.

Nothing is left failing: every entry point the enumeration finds passes.