llm: route every LLM caller through a shared backend client

Replaces the per-plugin Ollama HTTP calls with internal/llm, which picks a
backend from the environment (vLLM or Ollama) behind one Chat interface, plus
internal/plugin/llm_client.go as the plugin-facing wrapper.

Startup now logs llm_backend/llm_endpoint/llm_model instead of the two
OLLAMA_* vars, which no longer describe where inference actually goes.

These files were already running in prod from the vLLM migration but had never
been committed; this is that live state, byte-for-byte.

internal/plugin/llm_client.go

@@ -0,0 +1,45 @@
+package plugin
+
+import (
+	"context"
+	"sync"
+
+	"gogobee/internal/llm"
+)
+
+// The inference backend is process-wide: one endpoint, one model, selected by
+// env at startup. Plugins share a single client rather than each rebuilding one
+// per invocation, and read config through llmConfigured/llmGenerate so that
+// swapping Ollama for vLLM is a config change rather than a code change.
+var (
+	llmOnce   sync.Once
+	llmShared llm.Client
+	llmCfg    llm.Config
+)
+
+func llmInit() {
+	llmOnce.Do(func() {
+		llmCfg = llm.ConfigFromEnv()
+		llmShared = llm.New(llmCfg)
+	})
+}
+
+// llmConfigured reports whether an endpoint and model are set. Plugins call
+// this to stay dormant instead of erroring on every invocation — the same role
+// the old `if ollamaHost == "" || ollamaModel == ""` guards played.
+func llmConfigured() bool {
+	llmInit()
+	return llmCfg.Configured()
+}
+
+// llmClient returns the shared backend client.
+func llmClient() llm.Client {
+	llmInit()
+	return llmShared
+}
+
+// llmGenerate is the one-line path for the common case: a raw prompt in,
+// visible completion out, reasoning blocks already stripped.
+func llmGenerate(ctx context.Context, req llm.Request) (string, error) {
+	return llmClient().Generate(ctx, req)
+}