Add stock LPOSS inference runner with explicit maskclip_raw / lposs / lposs_plus modes

[InsightofSPb]

Motivation

• Confirmed root cause: legacy FineTuneWrapper-based paths invoked clip_backbone and MaskCLIP decode-head probabilities directly, bypassing LPOSS.forward (DINO extraction, patch affinity graph, LPOSS label propagation, and LPOSS+ pixel refinement).
• Additional defects: segmentation/evaluation/lposs_eval.py allocated solver tensors on a hardcoded generic CUDA device and contained a duplicated conditional-expression in LPOSS_Infrencer.forward, and the legacy head applied softmax before any propagation so probabilities were treated like logits.
• Goal: restore one reliable stock open-vocabulary inference path that reproduces upstream LPOSS behavior (three explicit modes), preserves historical MaskCLIP-only reproduction paths, and fails closed when graph dependencies are missing.

Description

• Implemented a modular stock inference engine ovs_heritage/stock_lposs.py exposing three explicit modes: maskclip_raw (frozen MaskCLIP features → raw cosine seed scores), lposs (adds frozen DINO extraction + patch-level LPOSS propagation), and lposs_plus (adds LPOSS+ pixel refinement), with injectable propagator and truthful stage metadata in ExecutionMetadata.
• Added a dataset-independent CLI entry point python -m ovs_heritage.infer_ovs (ovs_heritage/infer_ovs.py) supporting single image / directory / manifest input, explicit model config, canonical vocabulary reuse, extra runtime concepts, device selection, optional score saving, optional ledger recording, and deterministic run_manifest.json output; the CLI rejects silent slide-mode approximations and documents behavior and provenance.
• Adapted upstream vendored graph utilities via a thin adapter UpstreamGraphPropagator calling segmentation/evaluation/lposs_eval.py graph builders and solvers, and made the solver device-aware to avoid hardcoded `

---

Codex Task