docs: point at @gmod/range-cache-filehandle for HTTP reads

A whole-reference query on ce#1000 issues 545 filehandle reads over a
141 KB file, and the README said nothing about what that costs against a
bare RemoteFile. docs/api.md already assumed a byte-range-caching
filehandle without naming one that exists.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: cmdcolin

README.md

@@ -137,6 +137,28 @@ reason to reach for either is a host that runs several worker contexts, since
 one pool serves a context rather than a machine.
 [docs/workers.md](docs/workers.md) has the measurements.
 
+## Reading over HTTP
+
+An indexed query reads a file as many small byte ranges — a whole-reference
+query on a 141 KB test file issues 545 of them. A bare `RemoteFile` turns each
+one into its own range request, so put a byte-range cache underneath:
+
+```js
+import { RemoteFileWithRangeCache } from '@gmod/range-cache-filehandle'
+
+const cram = new IndexedCramFile({
+  cramFilehandle: new RemoteFileWithRangeCache(url),
+  index: new CraiIndex({
+    filehandle: new RemoteFileWithRangeCache(`${url}.crai`),
+  }),
+})
+```
+
+It serves reads from a 256 KiB chunk grid and coalesces the chunks a read is
+missing into one request per contiguous run, so those 545 reads become a handful
+of requests. It also threads the `AbortSignal` below, which is what makes the
+next section worth anything over a network.
+
 ## Cancelling a query
 
 Pass an `AbortSignal` and the query stops decoding and drops the fetch it has in

docs/api.md

@@ -105,9 +105,11 @@ controller.abort() // `records` rejects with an AbortError
 ```
 
 This is worth more than "index reads are short" suggests. A byte-range-caching
-filehandle coalesces adjacent reads into one request, so a small viewport over
-deep data becomes a single multi-megabyte fetch — exactly the thing you want to
-drop when the user pans away.
+filehandle such as
+[@gmod/range-cache-filehandle](https://github.com/GMOD/range-cache-filehandle)
+coalesces adjacent reads into one request, so a small viewport over deep data
+becomes a single multi-megabyte fetch — exactly the thing you want to drop when
+the user pans away.
 
 **Aborting your query never fails anyone else's.** Concurrent queries share two
 things in a `CramFile`: the parsed `.crai`, and each decoded slice in the record