Ollama native adapter

internal/ai/llm/ollama implements llm.Provider over plain net/http:
POST {base}/api/chat with "stream": false, the request schema in
"format", and options.temperature / options.num_predict only when the
request sets them; GET {base}/api/tags for ListModels and Verify. No
Authorization header is sent and redirects are not followed.

Usage comes from prompt_eval_count/eval_count. A schema reply that is
not a JSON object is ErrBadOutput. HTTP 404 on /api/chat is an
ErrProvider naming the configured model, without quoting the body.
Refused connections and timeouts match ErrTimeout and say
"Ollama not reachable at {base}"; a cancelled context stays
context.Canceled. The base URL is required, absolute http or https
(LAN hosts allowed), and may not carry userinfo, a query or a fragment.

Contract tests replay sanitized fixtures adapted from Ollama's public
API docs against httptest servers: success, schema output, model
missing and unreachable host.

Maestro task: 51289131-5abf-48c3-b553-7e3809c09076

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/ai/llm/ollama/testdata/chat_text.json

@@ -0,0 +1,16 @@
+{
+  "model": "llama3.2",
+  "created_at": "2026-10-04T12:13:43.416799Z",
+  "message": {
+    "role": "assistant",
+    "content": "Try a crock pot chili: it is a Quick dinner."
+  },
+  "done": true,
+  "done_reason": "stop",
+  "total_duration": 5191566416,
+  "load_duration": 2154458,
+  "prompt_eval_count": 26,
+  "prompt_eval_duration": 383809000,
+  "eval_count": 298,
+  "eval_duration": 4799921000
+}