Validate native RNGD context boundaries and recovery after rejected requests #277

Problem

The four-model matrix includes native RNGD KONI fixed-length serving and cache reuse, but does not establish native context-limit rejection and subsequent healthy inference.

Evidence

Native context9216 KONI artifact and SDK2026.4 evidence are recorded in Draft PR275. Valid input8192/output128 is covered; limit-adjacent input and rejected-request recovery are not.

Scope

One owned native RNGD server per cache mode (off/on). Exercise accepted context-edge requests and malformed/oversized requests through the public completion API. Preserve actual HTTP status/body, explicit input IDs, normal-request reference tokens, health and post-error recovery. Do not infer an engine failure from a client timeout or overwrite failed results. No vendor engine patch.

Done when

Measured statuses and token counts at input+output capacity boundaries are documented with server/FXB identity. Every rejected request is followed by a normal completion compared with its same-server reference. Cache modes are reported independently; unsupported or divergent outcomes remain visible. Related #201 and PR275.

Owner

Codex supplies isolated client and hardware evidence. Engine changes, if needed, go to the maintainer. AI-assisted research.