fix(docs): retry snapshot push; portable Pages URL; drop non-positive Y

Three code-review fixes on PR 5:

1. Snapshot push race. The bench-upload Snapshot step pushed devel
   without rebasing, so a concurrent push during a long bench run
   would reject non-fast-forward and fail the entire job. Add a
   fetch + rebase + retry loop (3 attempts) and push HEAD:devel.
   The commit is already [skip ci], so the retry cannot trigger an
   infinite bench loop on a successful push.

2. Hardcoded GitHub Pages verify URL. The docs.yml verification
   step curled https://elijahr.github.io/lockfreequeues/... which
   is wrong on forks or after a repo rename. Derive the URL from
   github.repository_owner and github.event.repository.name so the
   verifier follows the repo wherever it lives.

3. Log-scale accepts non-positive. The chart defaults to log Y
   (distr: 3) but merge_bmf.py only enforces finiteness, so a
   malformed BMF could deliver a 0 or negative throughput value.
   Treat <= 0 as missing data in toUplotData. Throughput in ops/ms
   cannot legitimately be non-positive (a 0 indicates a degenerate
   run that bench_throughput's elapsedNs guard already rejects), so
   surfacing as missing is the correct semantics.

Author: elijahr

.github/workflows/bench.yml

@@ -709,4 +709,25 @@ jobs:
             exit 0
           fi
           git commit -m "chore(bench): refresh snapshot [skip ci]"
-          git push origin devel
+          # Long bench runs leave a wide window for `devel` to advance
+          # before we push (a docs-only follow-up commit, an unrelated
+          # merge, etc.); a plain `git push` would then be rejected
+          # non-fast-forward and fail the entire bench-upload job.
+          # Retry: fetch+rebase+push up to 3 times. Skip-CI is on the
+          # commit so the retry loop cannot trigger an infinite bench
+          # workflow regardless of how many fast-forwards happen.
+          for attempt in 1 2 3; do
+            if git push origin HEAD:devel; then
+              exit 0
+            fi
+            echo "::notice::push attempt $attempt rejected; rebasing onto origin/devel and retrying."
+            git fetch origin devel
+            # Rebase preserves our snapshot commit on top of any new
+            # commits that landed since checkout. Conflict on the
+            # snapshot files themselves is impossible (no other workflow
+            # writes to docs/assets/bench-results/), so a clean rebase
+            # is the expected outcome.
+            git rebase origin/devel
+          done
+          echo "::error::Snapshot push failed after 3 retries; aborting." >&2
+          exit 1

.github/workflows/docs.yml

@@ -67,13 +67,23 @@ jobs:
       # `sleep 30` accommodates GitHub Pages CDN propagation; raise to
       # 60 if this step ever flakes.
       - name: Verify mike asset path
+        # Derive the Pages URL from the repository context so a fork or
+        # repo rename does not falsely fail this verification step. The
+        # default GitHub Pages URL convention is
+        # `https://<owner>.github.io/<repo>/`. If the project ever moves
+        # to a custom domain the URL becomes wrong again, but at that
+        # point we'd update the workflow to read site_url from
+        # mkdocs.yml; for now the convention is stable.
         if: |
           github.ref == 'refs/heads/main' ||
           github.ref == 'refs/heads/master' ||
           github.ref == 'refs/heads/devel'
+        env:
+          OWNER: ${{ github.repository_owner }}
+          REPO: ${{ github.event.repository.name }}
         run: |
           set -euo pipefail
-          URL="https://elijahr.github.io/lockfreequeues/dev/assets/bench-results/latest.json"
+          URL="https://${OWNER}.github.io/${REPO}/dev/assets/bench-results/latest.json"
           sleep 30
           HTTP_CODE=$(curl -sS -o /tmp/latest.json -w "%{http_code}" "$URL")
           if [ "$HTTP_CODE" != "200" ]; then

docs/assets/bench-charts.js

@@ -134,16 +134,24 @@
 
   /* Convert {xLabels, libraries} into the column-major arrays uPlot
    * expects: data[0] is the X axis; data[i+1] is the Y series for
-   * libraries[i]. Missing (library, shape) cells are NaN, which uPlot
-   * skips in line drawing.
+   * libraries[i]. Missing (library, shape) cells are null, which uPlot
+   * skips in line drawing. Non-positive values are also treated as
+   * null: throughput in ops/ms cannot legitimately be <= 0 (a 0 means
+   * a degenerate run our throughput.nim guards already reject), and
+   * the chart's default Y axis is log-scale (`distr: 3`) which is
+   * undefined for non-positive values. merge_bmf.py only enforces
+   * finiteness, so a malformed BMF could in principle deliver a 0 or
+   * negative value; rather than letting it break the log axis,
+   * surface it as missing data.
    */
   function toUplotData(xLabels, libraries) {
     const x = xLabels.map((_, i) => i + 1);
     const series = libraries.map((lib) => {
       const byShape = new Map(lib.points.map((p) => [p.shape, p]));
       return xLabels.map((label) => {
         const p = byShape.get(label);
-        return p ? p.value : null;
+        if (!p) return null;
+        return p.value > 0 ? p.value : null;
       });
     });
     return [x, ...series];