grill-perf: reported-prefix-hit ineligibility is correct but needs a hint about cache alignment

[plotarmordev]

On a hybrid-KV vLLM serve (GLM-5.3-Flash, block-aligned prefix cache with a 3584-token hit grid) a reported-prefix-hit workload with ~2K prompts reports cached_prompt_tokens: 0 on every measured turn, so every wave is provider_reported_prefix_cache_miss and the run is completed-with-ineligible-measurements. The verdict is right, but the doc should say that reported hits depend on server block/hit alignment and recommend observe or prompts ≥ the server's alignment for warm-decode cells. Evidence: 12/12 waves ineligible until switching to observe; same server hit 7168 cached tokens on an 8.7K prompt.

[plotarmordev]

Resolved in PR #18. Cache guidance now explains engine-specific block/alignment requirements and why warmup alone may not produce a reported hit. It recommends inspecting reported counts and choosing an appropriate prompt/protocol without promising length guarantees reuse. Switching to observe is explicitly a different workload/claim; the correct reported-hit eligibility rule is unchanged.