docs: lead Vision section with llama-server (always-built path)

Reconfirmed vision-via-llama.cpp end-to-end on 2026-05-19: bundled
Q4_K_M GGUF + mmproj-F16.gguf from unsloth/Qwen3.6-27B-GGUF, Vulkan
backend on Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S), 65/65
layers offloaded. A 1024-px JPEG posted to llama-server's OpenAI-
compat /v1/chat/completions endpoint with an `image_url` data-URL
content block produced an accurate description (478 completion
tokens) of a Tokyo-style alleyway scene — red paper lanterns,
bicycles, Japanese signage, Sapporo sign. Confirms the README
"Vision (mmproj) ✅" claim for the llama.cpp loader.

Two writing fixes from what the retest actually surfaced:

1. The README's CLI example led with `llama-mtmd-cli`, but on this
box that binary wasn't built. Cause was a selective cmake build
(only `llama-cli`/`llama-server`/`llama-bench` targets), not a
missing flag — `LLAMA_BUILD_TOOLS=ON` is already the default.
The note now says to `cmake --build build --target llama-mtmd-cli`
if a selective build skipped it.

2. The `llama-server --mmproj` HTTP path is in stock builds (it's
listed in the loader table) but had no concrete example. Now
shows the server invocation + the wire format (`image_url` data
URL), names where the thinking trace lands (`reasoning_content`)
vs the final answer (`content`), and flags the ≥500 max_tokens
budget needed so the `<think>` block doesn't crowd out the
visible answer.

Vision-prereq fetch (`scripts/fetch_vision.sh F16`) and smoke
(`make smoke` + `make smoke-tools` against the HF-pulled
`hf.co/FoolDev/Thanatos-27B` tag) both passed in the same session.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

CHANGELOG.md

@@ -8,6 +8,25 @@ and documentation**, not the underlying base model.
 ## [Unreleased]
 
 ### Added
+- README "Vision via llama.cpp" subsection now leads with the
+  `llama-server --mmproj` HTTP path (always built into stock llama.cpp,
+  no extra cmake targets needed), reconfirmed working 2026-05-19 with
+  llama.cpp 389ff61 + Vulkan on the Strix Halo reference machine
+  (bundled Q4_K_M GGUF + `mmproj-F16.gguf` from `unsloth/Qwen3.6-27B-GGUF`).
+  Sent a 1024-px JPEG via an OpenAI-style `image_url` data-URL content
+  block; model produced an accurate description (Japanese alleyway
+  with paper lanterns, bicycles, etc.) in 478 completion tokens. The
+  visible answer arrived in `message.content`, the thinking trace in
+  `message.reasoning_content` — the section notes both, plus the
+  ≥500 max_tokens budget needed so the reasoning block doesn't crowd
+  out the final answer. The existing `llama-mtmd-cli` and
+  `llama-cpp-python` examples are still listed; `llama-mtmd-cli` now
+  carries a note that it's a separate cmake target — a plain
+  `cmake --build build` produces it, but a selective build (e.g.
+  `cmake --build build --target llama-cli llama-server llama-bench`)
+  silently skips it, which is what tripped the retest on this box
+  (`LLAMA_BUILD_TOOLS=ON` was already set; the mtmd target just hadn't
+  been requested).
 - Third measured tok/s data point on the Strix Halo reference hardware:
   **Q3_K_S under Vulkan → 12.31 tok/s aggregate** (6182 tokens /
   501.9 s; 12.67 / 12.55 / 12.25 short/medium/long). Now apples-to-apples

README.md

@@ -327,15 +327,33 @@ This repo intentionally does not redistribute either.
 
 ### Vision via llama.cpp
 
+Three flavors, in order of build-time effort:
+
 ```bash
-# CLI:
+# A. HTTP via llama-server (always built — the easiest path).
+#    Reconfirmed working 2026-05-19 against llama.cpp 389ff61 + Vulkan
+#    on a Ryzen AI Max+ 395 / Radeon 8060S iGPU.
+llama-server \
+  -m Qwen3.6-27B-Q4_K_M.gguf \
+  --mmproj mmproj-F16.gguf \
+  --host 127.0.0.1 --port 8765 -c 8192 -ngl 99
+# then POST OpenAI-style chat completions with an image_url content
+# block — e.g. {"type":"image_url","image_url":{"url":"data:image/jpeg;base64,..."}}
+# The thinking trace arrives in message.reasoning_content; the visible
+# answer is in message.content. Budget ≥500 max_tokens so the reasoning
+# block doesn't crowd out the final answer.
+
+# B. CLI via llama-mtmd-cli (one-shot). It's a separate cmake target,
+#    so a selective `cmake --build build --target llama-cli ...` won't
+#    produce it — a plain `cmake --build build` will. If yours didn't,
+#    run `cmake --build build --target llama-mtmd-cli`.
 llama-mtmd-cli \
   -m Qwen3.6-27B-Q4_K_M.gguf \
   --mmproj mmproj-F16.gguf \
   --image photo.jpg \
   -p "Describe this image."
 
-# Python:
+# C. Python via llama-cpp-python:
 python examples/llama_cpp_vision.py \
   --gguf /path/to/Qwen3.6-27B-Q4_K_M.gguf \
   --mmproj /path/to/mmproj-F16.gguf \