test(benchmark): add chunk reuse accuracy baseline and repair analysis

[7774596]

Summary

• Add benchmarks/bench_accuracy_baseline.py: measures chunk-mode KV reuse vs full prefill accuracy on LongBench QA (hotpotqa, 2wikimqa, musique) through the production HTTP service (/documents + /infer), with identical greedy decoding in both modes.
• Scoring methodology: raw QA F1, extracted-answer F1 (bold span / first sentence, separating factual errors from verbosity drift), and gold-answer substring recall; accuracy means are computed only over samples with a TTFT-verified chunk-KV load so silent full-recompute fallbacks cannot dilute the gap.
• Add docs/optimizations/5_chunk_cross_attention_repair.md: error decomposition (fake attention sinks, missing cross-chunk attention, pad noise, RoPE numerics), hard constraints from the connector interface, ranked repair directions (shared-prefix prefill, composed-prefix partial hits, seam patches, per-chunk correction), and the measured baseline results.
• Documents two measurement pitfalls with remedies: ring-buffer exhaustion across datasets and generation-cap truncation (the initial 64-token cap truncated 35-69% of reuse answers and conflated cap artifacts with KV damage).

Results (Qwen3-8B, RTX 4090, greedy, max_new_tokens=256)

Dataset	N	Full F1	Reuse F1	Ext-F1 Gap	Rec Gap
HotpotQA	19/20	0.7338	0.1714	+0.4486	+0.2632
2WikimQA	20/20	0.2951	0.0665	+0.2060	+0.0500
MuSiQue	16/16	0.2784	0.0138	+0.2073	+0.1250

Test plan

• Ran the full three-dataset comparison end to end against the production service stack (vLLM + DaseR, iouring, chunk mode), restarting DaseR with a clean store per dataset
• ruff check / ruff format --check pass on the benchmark script
• Verified Loaded column detects and excludes silent full-prefill fallbacks