Rerun the launch scatter benchmark and refresh every published number

README.md

@@ -159,9 +159,9 @@ chart = xy.line_chart(
   <img src="spec/assets/launch-benchmark-comparison.svg" alt="Cold-render time for a 10-million-point chart in XY, Matplotlib, and Plotly. Lower is better." width="1200">
 </p>
 
-In the recorded 10-million-point baseline, XY produced a static PNG in 0.023 s
-versus 2.8 s for Matplotlib and 9.6 s for Plotly, and reached first interactive
-render 16–20× sooner.
+In the recorded 10-million-point baseline, XY produced a static PNG in 0.018 s
+versus 2.7 s for Matplotlib and 9.6 s for Plotly, and reached first interactive
+render 16–18× sooner.
 
 That baseline is one proof point, not the finish line. XY is being built to
 outperform every competing charting library across small-data startup,
@@ -172,7 +172,7 @@ the project measures and proves that goal.
 
 The committed launch baseline uses identical seeded data, a 900×420 output,
 and three isolated cold runs. See the
-[launch report](benchmarks/launch_baselines/xy-0.1.0/macos-arm64-m5-pro/report.md),
+[launch report](benchmarks/launch_baselines/xy-main-2026-07-26/macos-arm64-m5-pro/report.md),
 [benchmark runbook](benchmarks/README.md), and
 [competitive benchmark specification](spec/benchmarks/results.md) for the
 environment, methodology, raw results, and full evidence program.

benchmarks/README.md

@@ -35,10 +35,12 @@ interactive CPU fallback through SwiftShader. Each successful cell is the mean
 of three complete cold-process runs; interactive samples also use a fresh
 browser. Terminal 1B failures are attempted once and are not averaged.
 
-Reproduce the 0.1.0 launch environment with its exact dependency versions:
+Reproduce the launch environment with its exact dependency versions. Both
+baseline directories carry the same competitor pins, so either one reproduces
+the comparison environment; use the newest for a fresh run:
 
 ```bash
-BASELINE=benchmarks/launch_baselines/xy-0.1.0/macos-arm64-m5-pro
+BASELINE=benchmarks/launch_baselines/xy-main-2026-07-26/macos-arm64-m5-pro
 uv sync --project "$BASELINE" --frozen --python 3.14.5
 ```
 
@@ -63,9 +65,18 @@ uv run --project "$BASELINE" --frozen python benchmarks/bench_launch_scatter.py
 ```
 
 The immutable 0.1.0 launch baseline, report, and raw results live under
-`benchmarks/launch_baselines/xy-0.1.0/macos-arm64-m5-pro/`. Add a new
-version/environment directory for later launches; never overwrite an earlier
-launch baseline or mix hardware and SwiftShader rows.
+`benchmarks/launch_baselines/xy-0.1.0/macos-arm64-m5-pro/`. The 2026-07-26 rerun
+that the README and public benchmarks page quote lives under
+`benchmarks/launch_baselines/xy-main-2026-07-26/macos-arm64-m5-pro/`; it repeats
+the same contracts against the same pinned competitor versions on the same
+machine. Add a new version/environment directory for later launches; never
+overwrite an earlier launch baseline or mix hardware and SwiftShader rows.
+
+Warm the checkout before a measured run. A first invocation in a fresh worktree
+pays for cold bytecode caches, Matplotlib's font cache, and Kaleido's browser
+download, which inflate every library's first row by an amount that is not part
+of any output contract. Run the suite once and discard it; publish the run after
+it.
 
 ## CI Software GL