Fast startup: mmap backward scan instead of reading full log

Uses JsonlBackwardIter (SIMD memrchr3) to scan the conversation log
newest-first without reading/parsing the whole file. Stops as soon
as the conversation budget is full. Only the kept nodes get
retokenized and pushed into context.

18MB log → only tokenize the ~50 nodes that fit in the budget.

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

src/agent/mod.rs

@@ -568,18 +568,17 @@ impl Agent {
     }
 
     pub async fn restore_from_log(&self) -> bool {
-        let all_nodes = {
+        let tail = {
             let ctx = self.context.lock().await;
             match &ctx.conversation_log {
-                Some(log) => match log.read_nodes(64 * 1024 * 1024) {
-                    Ok(nodes) if !nodes.is_empty() => nodes,
-                    _ => return false,
+                Some(log) => match log.read_tail() {
+                    Ok(t) => t,
+                    Err(_) => return false,
                 },
                 None => return false,
             }
         };
 
-        // Walk backwards from the tail, retokenize, stop at budget
         let budget = context::context_budget_tokens();
         let fixed = {
             let ctx = self.context.lock().await;
@@ -588,9 +587,10 @@ impl Agent {
         };
         let conv_budget = budget.saturating_sub(fixed);
 
+        // Walk backwards (newest first), retokenize, stop at budget
         let mut kept = Vec::new();
         let mut total = 0;
-        for node in all_nodes.into_iter().rev() {
+        for node in tail.iter() {
             let node = node.retokenize();
             let tok = node.tokens();
             if total + tok > conv_budget && !kept.is_empty() { break; }