feat(preprocessing): Pillow resize and augmentation drivers

[JArmandoAnaya]

Position 2 of the pre-processing epic (#785): the pixel side of a recipe, as a package outside the kernel with its own entry-point group.

What changed

• New src/visionset/preprocessing/ package. registry.py scans the visionset.preprocessing entry-point group on every call, keeps what satisfies the PreprocessingDriver port and keys it by step kind; drivers(), pick(installed, kind), driver_for(installed, step) and driver(kind) mirror the format registry's shape. A kind no installed driver applies is refused with PreprocessingDriverNotFound, whose message lists the installed kinds.
• pillow/__init__.py ships the two built-in drivers. PillowResizeDriver applies the EXIF turn first — ingest's PillowImageProcessor bakes the orientation in before it measures, so the manifest's dimensions already describe the oriented picture — then stretches with Image.resize (LANCZOS when the pixel count shrinks, BICUBIC when it grows) or letterboxes by pasting the resized content at the kernel's letterbox_fit offsets onto a pad_value canvas. PillowAugmentDriver mirrors, brightens then contrasts, or turns by quarter turns, reading every draw from the kernel's hflip_applied, brightness_contrast_factors and rot90_quarter_turns; variant 0 comes back untouched apart from orientation and re-encoding.
• PreprocessingDriverNotFound in kernel/errors.py, mapped to 500 PREPROCESSING_DRIVER_NOT_FOUND with an exposed message, with its rows in docs/content/api.md and tests/server/test_errors.py. No route raises it yet; the composition points arrive in the next position.
• pyproject.toml: the visionset.preprocessing entry-point group with pillow-resize and pillow-augment, and import-linter contract 1's forbidden list gains visionset.preprocessing. The contract is tightened, never relaxed; the fresh-process purity probe now also asserts neither visionset.formats nor visionset.preprocessing is loaded by a kernel import.

Re-encode rules

Bytes come back in the source's own format: a JPEG as a JPEG at quality 95 with its chroma subsampling kept (read off the source with get_sampling before the transform, since the resized image is no longer a JPEG Pillow can "keep" from); a PNG as a lossless PNG; anything else Pillow can open as a PNG. No metadata travels — the EXIF that named the orientation is dropped because the orientation is now in the pixels, and no ICC profile or text chunk is copied. media_type(bytes) in the same module answers image/jpeg or image/png for the preview route to return. Palette, bilevel, 16-bit and CMYK sources are converted to RGB (or RGBA when transparent) before any transform, because Pillow silently resamples a palette image nearest-neighbour.

Determinism scope

The same seed over the same bytes gives identical bytes on one machine with one Pillow and one set of codecs; the tests assert exactly that and nothing wider. Byte stability across environments is not promised, as the epic's decisions say.

Found, not fixed

• letterbox_fit has no floor of one pixel on the content size, so an extreme aspect ratio — a 1×200 source onto a 32×32 canvas — rounds one side to 0. The driver refuses that case with UnsupportedMedia before Pillow raises a bare ValueError, and the geometry transform still computes a zero-width strip for the same case. The single fix is a floor inside letterbox_fit, which this position does not touch; the kernel domain is the previous position's surface.
• The stretch resampling rule is decided on total pixel count. A stretch that shrinks one axis and grows the other is one Image.resize call, so it takes one filter; the dispatch's rule names downscaling and upscaling without saying which axis decides.

Stack

Position 2 of the pre-processing stack (#785); based on feat/preprocessing-p1-domain (PR #791).

Checks

Targeted, per commit: uv run pytest tests/preprocessing tests/architecture tests/kernel tests/server/test_errors.py tests/formats/test_registry.py (2119 passed, 6 skipped), uv run lint-imports (4 kept, 0 broken), uv run mypy src/visionset (no issues in 183 files), uv run ruff check . and uv run ruff format.

Full local gate on the final tree, verbatim:

4610 passed, 33 skipped in 38.22s
...
All checks passed.

check.sh: PASSED  ran=python,frontend,generated,browser  skipped=docs

Docs group, verbatim:

check.sh: PASSED  ran=docs  skipped=python,frontend,generated,browser