Run benchmarks during Pages build so the benchmarks page shows real data

[Copilot]

benchmarks/results.json on main is the empty stub, so benchmarks.html renders "No benchmark data available yet". PR #154 tried to regenerate it inside the perf-comparison autoloop, but that sandbox has no bun, so every iteration evaluates to 0 and the autoloop is stuck.

Move benchmark execution to the Pages workflow (which already has Bun) and revert the autoloop's evaluation to the pre-#154 file-count metric.

• .github/workflows/pages.yml — Setup Python earlier, run bash benchmarks/run_benchmarks.sh, then unconditionally copy the regenerated benchmarks/results.json into the playground artifact. The existence guard is dropped — if benchmarks fail, the build should fail loudly.

• .autoloop/programs/perf-comparison/program.md — Restore the original Evaluation section (counts min(bench_*.ts, bench_*.py) files) so iterations can ratchet again without needing to execute benchmarks. Update the iteration prose to note that the Pages workflow regenerates results.json on merge to main; the autoloop only needs to add benchmark scripts.

Trade-offs

• Pages builds get slower (~10–30 min for 600 pairs), but only on push to main, so PRs aren't blocked. Can be sharded with a matrix later if painful.
• Broken benchmarks surface on merge instead of in-PR. Acceptable since the autoloop still validates that benchmark files are syntactically valid.