Document the medium benchmark input

I've removed the comment header so the file is the exact output of
`helm template`.

Author: featherbread

benches/README.md

@@ -40,7 +40,7 @@ cargo bench json
 The argument to `cargo bench` is a substring match against the full benchmark
 names of the form `{size}_{format}/{source}`.
 
-- **size**: `small` (see below)
+- **size**: `small` or `medium` (see below)
 - **format**: A full format name as given to xt's `-f` or `-t` (e.g. `json`)
 - **source**: `buffer` (non-streaming) or `reader` (streaming)
 
@@ -50,22 +50,58 @@ benchmark run, including charts and comparisons with any previous run.
 
 ## Test Inputs
 
+Each benchmark loads test data into an in-memory buffer by translating a
+"default" version of the input with xt. This approach limits the size of the xt
+repository and ensures that disk I/O performance doesn't influence the results.
+However, it allows changes to xt's output formatting (e.g. whitespace, quoting)
+to influence the results. I expect such changes to be rare, at least compared
+to other changes whose impact is worth benchmarking.
+
+### Small
+
 The small input, `k8s-job.yaml`, is a simple Kubernetes `Job` that runs the
 Docker `hello-world` image. Translation time is usually a few microseconds for
 even the slowest input formats, so each benchmark runs in just a few seconds.
 This provides relatively fast feedback as you work.
 
+### Medium
+
+The medium input, `k8s-kyverno.yaml`, is a full set of Kubernetes manifests for
+deploying [Kyverno][kyverno] v1.16.2, generated from version 3.6.2 of the
+official chart using Helm v4.1.0 on `darwin/arm64`:
+
+```sh
+helm template kyverno kyverno/kyverno \
+  --version 3.6.2 \
+  --set admissionController.replicas=1 \
+  --set backgroundController.replicas=1 \
+  --set reportsController.replicas=1 \
+  --set cleanupController.replicas=1 \
+  --set webhooksCleanup.image.pullPolicy=IfNotPresent
+```
+
+To ensure TOML compatibility:
+
+1. The above `--set` options were chosen to eliminate all `null` values.
+2. The benchmark harness processes the raw Helm output by turning the stream of
+   YAML documents into a single object, with a single `manifests` field
+   containing an array of the documents. It does this by creating a small
+   MessagePack "header" to set up the object structure and type-length marker
+   for an array, then translating the YAML documents with xt. It then
+   translates the complete object to the final format for benchmarking.
+
+The strategy for generating the medium input is intended to be reproducible and
+auditable. The size of the input was chosen to balance space requirements for
+an xt repository checkout with the desire to avoid non-human-readable encodings.
+
+### Large (removed)
+
 The benchmarks previously included a 20 - 30 MB large input based on a sample of
 GitHub events, which was included in the xt repository (and remains in its
 history) as a Zstandard compressed archive of MessagePack data. Based on the
 reveal of the xz-utils backdoor that was obfuscated in part as compressed test
-data, **I have chosen to temporarily eliminate the large benchmarks** until they
-are reimplemented to rely exclusively on human-readable inputs, ideally without
-bloating the size of xt repository checkouts.
+data, **I have chosen to eliminate the large benchmarks** until they are
+reimplemented to rely exclusively on human-readable inputs, ideally without
+bloating the size of xt repository checkouts too much.
 
-Each benchmark loads test data into an in-memory buffer by translating a
-"default" version of the input with xt. This approach reduces the size of the xt
-repository and ensures that disk I/O performance does not influence the
-benchmark results. However, it allows changes to xt's output formatting
-(whitespace, quoting, etc.) to influence the results. I expect such changes to
-be rare, at least compared to other changes whose impact is worth benchmarking.
+[kyverno]: https://kyverno.io/

benches/k8s-kyverno.yaml

@@ -1,5 +1,3 @@
-# helm template kyverno kyverno/kyverno --set admissionController.replicas=1 --set backgroundController.replicas=1 --set reportsController.replicas=1 --set cleanupController.replicas=1 --set webhooksCleanup.image.pullPolicy=IfNotPresent
-# Chart version 3.6.2; Kyverno v1.16.2
 ---
 # Source: kyverno/templates/admission-controller/serviceaccount.yaml
 apiVersion: v1