Harden stock LPOSS inference: PyTorch-1.12 compatibility, pinned upstream, memory preflights, provenance and parity harness

[InsightofSPb]

Motivation

Restore a robust, auditable stock LPOSS inference path while preserving the original LPOSS algorithm and its open-vocabulary behavior. This PR hardens the implementation introduced in the parent PR #58 and prepares it for reproducible local CUDA parity validation.

What changed

• Added PyTorch 1.12-compatible checkpoint loading and deterministic model-state fingerprints.
• Pinned upstream LPOSS and DINO dependencies to immutable revisions.
• Fixed the official LPOSS adapter constructor to match the pinned upstream implementation.
• Added conservative graph-memory preflight checks to reject unsafe runs before large tensors are allocated.
• Added synchronized CUDA timing and truthful memory telemetry.
• Strengthened provenance records with source-manifest hashes, original dataset records, installed dependency versions, model fingerprints, configuration information, and source-control state.
• Added an auditable parity harness that compares official and local intermediate stages, tensor shapes, channel order, semantic IDs, and numerical values.
• Added authoritative validation of canonical dataset-v2 manifests through ovs_heritage.coco_converter.validate_manifest.
• Canonical validation now fails closed: invalid splits, missing masks, invalid mask domains, grid mismatches, and other dataset-contract violations stop inference instead of silently producing canonical results.
• Generic JSONL manifests remain supported but are explicitly marked as noncanonical.
• Expanded CPU contract tests and CI coverage for the LPOSS inference, adapter, provenance, dataset validation, and parity tooling.

Validation

GitHub Actions workflow ovs-heritage-p0 completed successfully:

• 131 passed
• Python compilation checks passed
• Ruff checks passed

Remaining acceptance gate

Real CUDA numerical parity against the pinned official LPOSS implementation has intentionally not been executed in CI. Therefore, this PR does not claim verified numerical equivalence yet.

After this PR is merged into the branch of PR #58, the combined implementation must be tested on the local CUDA environment. PR #58 should be merged into master only after the generated parity manifest reports successful real-GPU parity and all required tensor shapes, class/channel mappings, and numerical tolerances pass.