add memory importance scoring via prompt logprobs

score_memories() drops each memory from the context one at a time,
runs prompt_logprobs against the full conversation, and builds a
divergence matrix: memories × responses.

Row sums = memory importance (for graph weight updates)
Column sums = response memory-dependence (training candidates)

Uses vLLM's prompt_logprobs to check "would the model have said
this without this memory?" — one forward pass per memory, all
responses scored at once. ~3s per memory on B200.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

src/agent/api/mod.rs

@@ -166,6 +166,9 @@ impl ApiClient {
         Ok((build_response_message(content, tool_calls), usage))
     }
 
+    pub fn base_url(&self) -> &str { &self.base_url }
+    pub fn api_key(&self) -> &str { &self.api_key }
+
     /// Return a label for the active backend, used in startup info.
     pub fn backend_label(&self) -> &str {
         if self.base_url.contains("openrouter") {