Use SparkCache CUDA restore terminology

[FujitsuPolycom]

Resulting behavior

SparkRing-generated profiles emit these SparkCache connector keys:

• spark_cache_cuda_restore
• spark_cache_cuda_placement_library
• spark_cache_cuda_placement_library_sha256
• spark_cache_cuda_placement_arena_bytes
• spark_cache_cuda_restore_io_workers

SparkCache composition recipes emit cuda_restore. Operator prose names the implementation as SparkCache CUDA restore or SparkCache CUDA placement. The public-base composed-image builder, receipt, verifier, image label, and resolver option use the CUDA placement name as well.

SparkRing's runtime-profile parser and GLM profile resolvers accept the compatibility connector keys and --native-library-sha256. They normalize accepted input to the canonical output contract. A canonical key and compatibility alias with unequal values are rejected before a launch plan is produced. Equal duplicate values are accepted and emitted once under the canonical name.

Compatibility

This draft is stacked on #130 at exact base commit f032df696ad9dd9b0370aad9c1a85fd110b30223.

No runtime, model, B12X, SparkCache, NCCL, or image source pin changed. Cache-identity fields and connector values are unchanged. Cache namespace impact: none.

No image was built, pushed, or published. No serving host or service was contacted or changed.

Validation

• Focused terminology, profile, recipe, launcher, resolver, and composed-image tests: 70 passed, 1 unrelated base test deselected.
• Ruff on changed Python files: passed.
• Seven generated profile JSON documents and four composition recipes: parsed successfully.
• Python bytecode compilation: passed.
• Composed-image Bash syntax: passed.
• Generated-profile namespace audit: all seven identity maps unchanged; connector values differ only by canonical names and placeholder spelling.
• Pin audit: no pins.json changed.
• Maintained CPU suite: 1,935 passed, 9 skipped, 1 failure. The sole failure is inherited from the stack base: scripts/test_glm53_flash_profile.py::test_public_glm53_benchmark_is_sanitized_and_bounded expects C4 and C8 were capacity-limited, which is absent from the base README.md.

[FujitsuPolycom]

The resulting runtime, operator contract, and evidence are consolidated in retained draft stack #146 → #147 → #150. Independent page-base research remains in #149. Closing this superseded draft and deleting only its remote head branch.