ci: run benchmark comparison on pull requests only, and shorten it (#3140)

* ci: run benchmark comparison on pull requests only, and shorten it

The benchmark job checks out github.base_ref as the previous code, but
that is only set on pull_request events. On pushes to master both
checkouts were the same commit, so the job spent about 20 minutes
comparing master with itself (the Sep 30 run compared af07b69 with
af07b69). Run it only for pull requests.

Benchmarks now run with -benchtime=100ms instead of the default 1s. Echo
has 51 benchmarks run 8 times on each side, so the 1s default kept the
job at about 20 minutes regardless of the machine. At 100ms one side
takes about a minute.

Both workflows also get a concurrency group, so a new push to a pull
request cancels its superseded runs. Runs on master always finish.

Claude-Session: https://claude.ai/code/session_01QKDYQr53zNKkR7nif2CAAq

* ci: describe the benchmark comparison as advisory

Claude-Session: https://claude.ai/code/session_01QKDYQr53zNKkR7nif2CAAq

Author: vishr

.github/workflows/checks.yml

@@ -12,6 +12,11 @@ on:
 permissions:
   contents: read # to fetch code (actions/checkout)
 
+# Cancel superseded runs of the same pull request; runs on master always finish.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
 env:
   LATEST_GO_VERSION: "1.27"
 

.github/workflows/echo.yml

@@ -12,6 +12,11 @@ on:
 permissions:
   contents: read # to fetch code (actions/checkout)
 
+# Cancel superseded runs of the same pull request; runs on master always finish.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
 env:
   LATEST_GO_VERSION: "1.27"
 
@@ -56,6 +61,9 @@ jobs:
           fail_ci_if_error: false
 
   benchmark:
+    # Compares the pull request against its base branch. github.base_ref is only set for
+    # pull_request events, so on other events both checkouts would be the same commit.
+    if: github.event_name == 'pull_request'
     needs: test
     name: Benchmark comparison
     runs-on: ubuntu-latest
@@ -79,15 +87,17 @@ jobs:
       - name: Install Dependencies
         run: go install golang.org/x/perf/cmd/benchstat@latest
 
+      # -benchtime=100ms instead of the default 1s keeps the job to a few minutes. The
+      # comparison is advisory: it reports differences but never fails the build.
       - name: Run Benchmark (Previous)
         run: |
           cd previous
-          go test -run="-" -bench=".*" -count=8 ./... > benchmark.txt
+          go test -run="-" -bench=".*" -benchtime=100ms -count=8 ./... > benchmark.txt
 
       - name: Run Benchmark (New)
         run: |
           cd new
-          go test -run="-" -bench=".*" -count=8 ./... > benchmark.txt
+          go test -run="-" -bench=".*" -benchtime=100ms -count=8 ./... > benchmark.txt
 
       - name: Run Benchstat
         run: |