Measurements: IQ3_XXS at 262K on 93 GB, KV int8 vs fp16 recall, retrieval accuracy vs length (RTX 5070 Ti) #200

Three measurements that fill gaps in docs/DETAILS.md. No action needed; posting them in case they are useful for the tables or for users choosing settings.

Setup: RTX 5070 Ti 16 GB (sm_120), Ryzen 7 7700 (8 cores), 93 GB RAM, Ubuntu 24.04, CUDA 13.2. IQ3_XXS native pack, --kv int8 --kv-resident 32768, --spec 4 + MTP (rt-cjk), --vision (CPU), --vram-reserve-mib 1058, --expert-cache auto, --max-context 262144. The engine was 0.1.27 plus PR #189 and #194; neither touches prompt or decode math, and parity checks against 0.1.27 are in #189.

1. IQ3_XXS at 262K (DETAILS says "not measured" because of 64 GB)

• A 257,466-token prompt read from zero: 96.5 s (2,668 tok/s).
• A 151,004-token prompt: 2,659 tok/s on its 85,468 fresh tokens.
• Output at 151K–257K was 95–118 tok/s. These were list-style answers, which draft well, so treat them as optimistic.
• Prompts over the limit are refused cleanly: prompt (263800 tokens) + max tokens (400) exceeds the context (262144).

2. KV int8 vs fp16 at long context: no recall difference, so int8 is the right default

The test is exact-value retrieval: "What is the weight of Record N?" over N near-identical Record i: weight W kg. lines, 10 positions from 2% to 99.5% of the prompt, greedy.

fp16 cost −18 to −20% prompt speed: at ~60K, 2,993 → 2,463 tok/s; at 257K from zero, 96.5 s → 117.3 s. Output speed was unchanged. The wrong answers are not KV-precision errors.

3. Retrieval accuracy against length (int8, same test)

This is n = 10 per length, so the numbers are coarse, but the trend is monotonic. The errors are spread across positions, not concentrated in the middle. The task is adversarial: thousands of lines differ only in their numbers. Still, it may help users decide whether to enable 262K for agents, or to compact earlier.