docs(readme): stop claiming faster than vLLM, the curve is a tie

[localai-bot]

The README's headline claims contradicted its own Performance section.

The table below the fold has always been honest — "treat those as ties", "only c1, at 4.5%, is clearly outside the noise". But everything above it said we beat vLLM outright: the badge read "token-exact + faster at every concurrency", the section was titled "Performance: faster than vLLM", and the hero GIF caption said "vllm.cpp finishes first".

That is exactly the claim Hacker News pushed back on when the LocalAI post went up, and the pushback was right — 0.7% to 1.7% against a 0.5% run-to-run noise band is not a speed win. The LocalAI blog post now says "Those are ties" (mudler/LocalAI#11324, merged). This makes the README agree with it.

Changes

Where	Before	After
badge	token-exact + faster at every concurrency	token-exact + same throughput
section heading	Performance: faster than vLLM	Performance
GIF alt text	ahead at every concurrency	identical output at every concurrency
GIF caption	vllm.cpp finishes first	identical output, finishing a hair ahead, margin inside the noise band at five of six
Fast. bullet	we beat vLLM at all six concurrencies	we match vLLM's throughput, per-point breakdown kept
after the table	A tie is still a good result for a 66 MiB binary against a mature CUDA stack	dropped

That last one grades our own result in the sentence that reports it. Pre-chewing the numbers is what made a skeptical reader look harder at them on HN, not less hard.

Not changed

• No measurement. Every figure is identical (verified by diffing numeric tokens); the only "new" number is 9.1 GiB, already stated 15 lines earlier.
• The real wins stay stated as they were: 1.18x llama.cpp prefill on the same GGUF, ahead of MLX-LM on prefill, and 24.88 GiB vs 28.18 GiB peak host memory, which is a clean result and still called one.
• The Same tokens as vLLM. Same throughput. 140x less to install. tagline was already consistent, so it stands.

Both anchor references to the renamed heading are updated (the badge link and docs/BENCHMARKS.md:144); grep confirms no dangling #performance-faster-than-vllm remains, and the new shields.io URL returns 200.