Validate native RNGD streaming disconnect and concurrent survivor recovery #278

Problem

Native RNGD context rejection/recovery is covered by #277, but client disconnect during active generation and concurrent survivor requests remain unverified in the four-model matrix.

Evidence

SDK2026.4 public serving_completions.py installs handle_disconnect and invokes async engine abort. A closed client socket alone does not establish that engine work stopped. Native KONI supports the exact selected model; native registry limitations for the other three models remain recorded in PR275.

Scope

Isolated owned native KONI server, cache off/on, inputs128/8192, concurrency1/4/16. Close one streaming request after actual content is received before its1024-token limit. Capture running/waiting and generation counters, concurrent128-token survivors, and same-server normal recovery. Preserve parity failures and inconclusive cancellation evidence rather than calling every closed socket a successful abort.

Done when

Raw event/counter timing, device/model/FXB/runtime identity, survivor/recovery outputs and cleanup are recorded. Distinguish HTTP disconnect, observed idle drain and evidence of avoided generation. No claim of physical cache reclamation from HTTP alone. Related #201, #277 and PR275.

Owner

Codex supplies isolated measurements and portable client. Production engine fixes belong to the engine maintainer. AI-assisted research.