Fail closed on unknown llama.cpp runtime hashes

[Blackspirits]

Summary

llama.cpp engine archives and the separate Windows CUDA runtime archives are already checked against DownloadHashManager, but the integrity helper currently returns without verification when the key/digest cannot be resolved or when the stream is empty. That turns a future resolver/registry mismatch into an unauthenticated unpack.

This change:

• fails closed when the llama.cpp engine/runtime hash key is missing
• fails closed when the current digest is missing
• rejects empty/tampered payloads instead of treating an empty stream as verified
• rejects streams that cannot be safely read and rewound
• always rewinds the stream after hashing, including failure/cancellation paths
• preserves the existing release tag, URLs, registered digests, model-download behavior, variants and unpack flow
• adds regressions for all 12 current engine/CUDA-runtime digests, the real HTTP engine-download path, null/unknown keys, empty streams and non-seekable streams

Validation

• no URL or digest values were changed
• all 12 current registry hashes were independently compared with the sha256: asset digests published by the official ggml-org/llama.cpp release b10840; all 12 match
• current upstream still uses release tag b10840
• dynamic GGUF model downloads are intentionally out of scope because they do not use this pinned release/hash registry
• full solution restore passed
• full solution build passed
• full solution test suite passed on the first run
• retry step was not used
• validation run: #34704707527

Base is upstream 9be10e12d0f04655f2f134718b00679f89b2b15f.

This PR supersedes internal draft #44.

AI assistance: ChatGPT was used to independently re-audit the llama.cpp integrity path on current upstream, compare the current registry with official release asset digests, strengthen fail-closed verification, and add regression coverage.

[Blackspirits]

Independent adversarial re-check on current upstream base: confirmed all 12 current llama.cpp engine/CUDA-runtime registry hashes against the official ggml-org/llama.cpp b10840 release asset digests; no URL/hash values changed; null/unknown keys, empty streams and non-seekable streams now fail closed before unpack; tampered bytes are rejected and the stream is rewound. Full current-base CI run #34704707527 passed on the first test run with no retry. No blocker identified. Keep draft; no merge/promotion performed.