Route reconstructed pages to C++ CUDA placement

[FujitsuPolycom]

Resulting behavior

Flat opaque-page snapshots use SparkCache's direct-file C++/CUDA restore function. Authenticated page-delta graphs reconstruct their complete encoded page snapshot and pass those bytes to the separate in-memory C++/CUDA page-placement function.

The connector no longer sends an �ncoded_pages argument to the direct-file function, whose contract requires manifest lookup and cache-root inputs. Page-delta verification failures still reject the restore and recompute the request.

Compatibility

SparkCache wire values, digest salts, 256-token geometry, stored object schemas, and cache identity are unchanged. DeepSeek-V4 and GLM-5.2 source receipts are rebound because connector source and regression coverage changed.

Validation

• Live TP4 reproduction: selected 11,520-token sparkcache-page-delta-manifest/v1 root rejected with �xecute_native_hybrid_restore() got an unexpected keyword argument 'encoded_pages' and safely recomputed
• GPU-free dispatch regression failed before the fix and passes after it
• 739 SparkCache tests passed, 7 skipped
• 108 deployment tests passed, 1 skipped
• Ruff and diff checks passed

Status: implemented; live page-delta C++/CUDA restore qualification remains pending the rebuilt-image replay.

[FujitsuPolycom]

The resulting behavior and evidence are consolidated in retained draft stack #39 → #42 → #44. Independent heat and SSD research remains in #36. Closing this superseded draft and deleting only its remote head branch.