Fingerprint model outputs and harden composite loader integrity

[CraigMyles]

Summary

• Add per-model fingerprints for patch and slide outputs while keeping grid_hash geometry-only.
• Require receipt, resume, and verify paths to match the current fingerprint and expected dimension; legacy outputs without a fingerprint are recomputed in place.
• Verify SEAL adapters and GigaPath slide weights before model construction, and mark SEAL experimental while its upstream bases remain unpinned.

Details

Fingerprints record the exact weight artifact, revision and SHA-256, preprocessing, pooling and output dimension, resolved execution precision, loader-constructor contract, and composite dependencies. Slide fingerprints also bind to the selected patch-output fingerprint.

Feature-array fingerprints are written only after every row succeeds, so they act as completion markers. Replacing an invalid model preserves the shared grid and unrelated model arrays while invalidating dependent slide embeddings. Multi-device runs must resolve to one common output contract before the store is touched, and standalone slide encoding now validates the complete patch input before using it.

raw2features verify fails closed for unregistered model names rather than trusting a contract read back from the target store. Programmatic custom-model runs continue to supply their current contract through injected embedders or factories.

SEAL now verifies the pinned adapter before deserialisation, validates complete LoRA coverage, and uses a frozen constructor configuration. GigaPath verifies its downloaded checkpoint before passing it to the model factory. SEAL remains experimental for v0.1.1 because its base encoders are not yet locally pinned.

Testing

• ruff check . — passed
• pytest -m "not slow and not network" -q — 442 passed, 9 skipped, 90 deselected
• Focused fingerprint, receipt, loader, slide, CLI, and multi-device suite — 140 passed, 5 skipped
• pytest -q -m network tests/test_registry_revisions_network.py — 40 passed

Refs #6