feat(kernel): ingest pipeline — hashing, blob dedup, asset origin, batch materialization (#20)

[JArmandoAnaya]

Closes #20.

IngestService is the one door that turns a registered source into rows: it hashes every item, stores the bytes once, records what the decoder made of them, and puts the result in a draft batch. It closes the last write in the kernel that had no service behind it — _add_assets is gone from examples/sdk_end_to_end.py, which now writes its frames to incoming/, registers the directory as a source and ingests it like a real caller would.

Acceptance criteria

• Duplicate image across two sources ⇒ one blob — and one asset, asserted on the blob directory and on the rows.
• Asset origin metadata correct for both directory and video paths — source_id on every asset, plus frame_index/frame_timestamp and a clip.mp4#frame=7 uri for a decomposed clip.
• Pipeline covered end-to-end with generated fixtures — 51 new tests, nothing reads a checked-in binary.

Design

• One ingest, not two. Registration split register_images/register_video because its arguments genuinely differed; here they do not — the source already carries the kind, the path and the extraction rate, so the branch is on SourceKind.
• Identity is content; origin is provenance. The same bytes arriving again never rewrite the origin, the rule Source.registered_at already follows.
• Four transactions, with the decode in none of them, and blobs written before any row. An out-of-process decoder inside a write transaction is how a single-writer SQLite store starts reporting "database is locked"; an unreferenced blob is harmless, a row naming absent bytes is not.
• Failure splits by remedy. Unsupported and corrupt items are reported per file (IngestFailure, with the name and the reason kept apart and an IngestFailureKind a report can group on) and the run carries on. A missing ffmpeg fails the job and is re-raised — which is exactly why MediaToolUnavailable sits outside MediaError.
• Frames are not re-probed: the port already guarantees FRAME_FORMAT at the probed dimensions, and re-decoding would route our own encoder's output into an operator's report.
• Scope held to kernel: ingest pipeline — SHA-256 hashing, blob-store dedup, technical metadata extraction, materialization into a Batch #20: no transition table, no persisted counters, no persisted report (kernel: IngestJob — state machine pending→running→(completed|failed), queryable progress, idempotency via content addressing #19), no thumbnail_hash (kernel: thumbnails cached in the blob store (needed by the M5 gallery) #21).

Migration 8 — FORMAT_VERSION is now 8

Adds asset.format / source_id / frame_index / frame_timestamp, links ingest_job to its batch, and puts indexes under the two uniqueness rules that had been running as service-level pre-checks with nothing beneath them:

• uq_asset_project_content_hash — "the same bytes are the same asset", claimed by Asset's docstring since M1.
• uq_source_project_kind_path_fps — the index kernel: SourceService — source registration with provenance metadata (origin, date, original framerate, parameters) #18 named as owed once asset.source_id had a target. Its fourth term is coalesce(json_extract(video, '$.extraction_fps'), 0), because SQLite treats NULLs in a unique index as distinct and an image directory would otherwise never collide with itself.

Two findings the fresh-versus-migrated test caught while writing it, both now documented:

• SQLAlchemy cannot reflect an expression-based index, so checkfirst reports it absent and re-issues the CREATE. That one index uses CREATE UNIQUE INDEX IF NOT EXISTS, still compiled from the shared Index object.
• A column carrying a foreign key cannot arrive by ALTER — SQLite spells an added key inline while create_all spells it as a table constraint, so the two paths emit different DDL. ingest_job is rebuilt (no children, provably empty, counted) so batch_id keeps a real key; asset.source_id is added without one, because asset has four cascading children and rows that were legitimately already there.

Duplicate pre-checks refuse a workspace either index could not accept, naming both counts, rather than letting an IntegrityError escape initialize().

Also

• IngestCompleted has its first emitter, flipping the tripwire tests/kernel/test_events.py has held since kernel: in-process EventBus + domain events (BatchApproved, ReleasePublished, IngestCompleted) #13 — narrowed there rather than deleted, so the annotation cycle still cannot quietly claim to be an ingest.
• BatchService._require_draft promoted to public require_draft(uow, batch_id), so a frozen target batch is refused before anything is decoded.
• New docs/ingest.md; sources.md, persistence.md, events.md, examples.md and media.md updated where they described this as future work.

Checks

uv run ruff check .            All checks passed!
uv run ruff format --check .   113 files already formatted
uv run mypy src/visionset/kernel   Success: no issues found in 50 source files
uv run mypy src/visionset          Success: no issues found in 60 source files
uv run lint-imports            Contracts: 2 kept, 0 broken.
uv run pytest                  828 passed
uv run python examples/sdk_end_to_end.py   green
uv run python scripts/export_openapi.py    no diff

VERSION stays 0.0.1.dev0; no new dependency.