[gspark-etri]

Onggi now documents its implemented P2 scope and the path from checkpoint ranges to resident NPU updates in rngd-pd. Bilingual READMEs/integration records link the runtime APIs and measured evidence.

• Separate implemented storage/delta/cache/peer/CSI/GC from planned CRDs, compile orchestration, compatibility planning and placement-driven prewarming.
• Preserve all SDK2026.2 measurements:24updates/768tokens matched fresh loads; sparse updates improve while full-tensor speedup is unclaimed.
• Link rngd-pd PR31's SDK2026.4 native Runnable loader: fresh-process reconstruction and native8B B4/S256/group2 sparse/LM token parity plus interrupted-copy rollback. Group8 currently fails native compilation. Full cold loading remains roughly19–21s; cold-load and cross-SDK speedups are unclaimed.
• Distinguish historical read-only preflight from later allocated-card validation. Loaded-driver srcversion subsequently matched the installed running-kernel module. Shared SDK/driver installations remained unchanged.

Validation: source/evidence hashes, all update tokens and local links checked; no Go/runtime changes. Final head fcf2fb8 CI passed:gofmt,vet,race tests,LinuxCLI build and S3 conformance/byte-identical round trips on SeaweedFS and VersityGW.

Onggi supplies files/ranges; runtime graph compatibility and device mutation remain rngd-pd responsibilities. This does not provide live FXB mutation. Runtime PR31 is separate and unmerged.

Closes #68.

AI-assisted implementation and verification with Codex.