perf: 256 MB HTTP chunks for streaming GGUF indexing

Fewer HTTP round-trips: 18 GB shard = ~72 requests instead of ~1125.
256 MB fits comfortably in RAM alongside the dequantized tensor.

https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7

Author: claude

src/hpc/gguf_indexer.rs

@@ -591,7 +591,7 @@ mod tests {
         };
         eprintln!("  URL resolved, size: {:.2} GB", size as f64 / 1e9);
 
-        let mut reader = HttpRangeReader::with_chunk_size(url, size, 16 * 1024 * 1024); // 16 MB chunks
+        let mut reader = HttpRangeReader::with_chunk_size(url, size, 256 * 1024 * 1024); // 16 MB chunks
 
         let out_path = "/tmp/llama4_scout.bgz7";
         let out = std::fs::File::create(out_path).expect("create output");
@@ -651,7 +651,7 @@ mod tests {
         eprintln!("  URL: {}", url);
 
         // 16 MB chunks for fewer HTTP round-trips
-        let mut reader = HttpRangeReader::with_chunk_size(url, size, 16 * 1024 * 1024);
+        let mut reader = HttpRangeReader::with_chunk_size(url, size, 256 * 1024 * 1024);
 
         let out_path = "/tmp/llama4_scout_shard5.bgz7";
         let out = std::fs::File::create(out_path).expect("create output");

src/hpc/http_reader.rs

@@ -31,8 +31,8 @@ pub struct HttpRangeReader {
 }
 
 impl HttpRangeReader {
-    /// Default chunk: 8 MB (good balance for GGUF tensor reads).
-    const DEFAULT_CHUNK: usize = 8 * 1024 * 1024;
+    /// Default chunk: 256 MB (fewer HTTP round-trips, fits in RAM easily).
+    const DEFAULT_CHUNK: usize = 256 * 1024 * 1024;
 
     /// Create a new HTTP range reader.
     ///