[gspark-etri]

Combine the exact reviewed #69/#70/#71 heads: resident SDK documentation, Diff target-plan reuse and exact-profile artifact preparation. Preserve their history and verification records. This completes the first explicit-candidate preparation path; automatic discovery, placement and autoscaling remain future work.

Native proof on one allocated RNGD connects a real init container running Onggi preparation to a fresh standalone prepared HTTP worker. The init rejects a wrong-firmware candidate, verifies the matching graph/config bundle and mounts it read-only. The runtime warms before readiness and serves three batches of four requests: 12 requests /96 reference-equal tokens, consisting of 12 supplied first tokens and 84 native-produced suffix tokens. Upload 1 /graph load 1 /compile 0; all 113 shared projection/LM buffers and compiled identities stay unchanged. Checkpoint 409 /context 400 rejection checks pass.

Full method and bilingual report · sanitized evidence, including both attempts · runtime integration ai-computing/rngd-pd#33.

Validation:

• Independent integration review found no Critical/Important source findings; documentation scope reconciled.
• Combined source 466ef91: Linux format/race/vet/build pass; 154 test/subtest events passed, 10 environment-dependent skips. Binary SHA256 recorded.
• Frozen native execution: Onggi 466ef91 + runtime fafd820fead20ed51dcb7f60ebb5e7dc7d90b2f1, SDK2026.4.0 /Torch2.10.0+cpu, Llama-3.1-8B B4/S256/group2.
• Root independently verified 205 exported small-file hashes, 409 runtime and 125 Onggi canonical Git file hashes, original public reference tokens, each HTTP result, profiles/manifests and buffer identities.
• Final publication head 8f0b4bc: CI41008799474 all 3 jobs passed, including Go race/vet/build and live SeaweedFS/VersityGW profile preparation and warm-cache checks.
• UID-guarded cleanup and independent API checks confirm namespace/all 3 owned Pod UIDs absent. Both final containers exited 0.

Limits: graph/config delivery only; the 16.06 GB checkpoint remains external read-only. The invalid first harness sliced KV to prompt width; preserve its failed requests separately from the corrected full-width KV run. No warmup-only timing, cold-load speedup, production rollout, live gateway/prefill, multiple B/S native switching or ATOM HTTP proof. Large native/model/KV payloads remain remote-only; their remote checks are distinguished from local small-evidence verification. Final publication changes only docs/data; Go source stays identical to the frozen tested source.

No main merge or deployment. Integration branches and original source worktrees remain available for review.