evals(community): OsaurusAI/DeepSeek-V4-Flash-0731-JANG on Apple M5 Max

[jax-0n-git]

Crowdsourced model-compatibility run (see COMMUNITY_EVALS.md).

One contribution file; no other changes.

Environment: chip=Apple M5 Max · RAM 128GB · osVersion=26.6.2 · commit=cc19e77 · judge=self-judge · catalogHash=2e766b2b52f4cf37 · contributor=jax-0n-git

[jjang-ai]

Cross-family audit: DSV4 was behaviorally strong here (AgentLoop 115/139, loop-free 94.96%, 17.02 tok/s), but the 110,491.79 MB peak physical footprint is a hard RAM failure, not a blocked/missing measurement. Current strict batch proof separately shows safe serialization [1,1] with engineCapacity; that does not waive this footprint defect. PR #2546 adds exit plus KV/SSM/L2/MTP aggregate telemetry so a current rerun can separate loop exits, cache reuse, and RAM. Required follow-up: current pin, safe-serialized scorer, memory cleanup across multi-turn/Stop/restart, and visible app proof. Draft vMLX #338 remains rejected on speed (0.08/0.75/0.14 tok/s) despite coherent output.