Reject perf-comparison iterations with invalid benchmark scripts

[Copilot]

Autoloop evaluation commands only measured their metric and never verified correctness, so iterations that broke benchmark scripts (or, in the sibling pandas-migration program, the type checker / tests) were silently accepted. This updates the perf-comparison evaluator to gate the metric on benchmark-script validity.

Changes

• .autoloop/programs/perf-comparison/program.md — evaluation block now runs validity checks before counting:
  • bun build each benchmarks/tsb/bench_*.ts (catches parse, transpile, and import-resolution errors)
  • python3 -m py_compile each benchmarks/pandas/bench_*.py
  • On any failure, or if bun / python3 is missing, emit {"benchmarked_functions": null, "rejected_reason": "..."} so the autoloop runner rejects the iteration.

for f in benchmarks/tsb/bench_*.ts; do
  [ -e "$f" ] || break
  if ! bun build "$f" --outdir=/tmp/perf-comparison-bench-check >/dev/null 2>&1; then
    echo "{\"benchmarked_functions\": null, \"rejected_reason\": \"invalid TypeScript benchmark: $f\"}"
    exit 0
  fi
done

Running the new evaluator against the current repo immediately surfaces 5 pre-existing broken TS benchmarks (e.g. bench_dataframe_mask.ts imports a non-existent dataFrameMask export) and a broken Python benchmark — exactly the regression class the gate is meant to catch. Those scripts are left as-is (out of scope) and will now correctly block iterations until fixed.

Out of scope

• build-tsb-pandas-typescript-migration — its program definition lives in Issue Build tsb: pandas → TypeScript migration #1, not in a repo file. AGENTS.md forbids in-repo modification and no tool here can edit issue bodies. A maintainer should add the bunx tsc --noEmit + bun test gate to that issue body per the proposal in the originating issue.
• tsb-perf-evolve — already validates via bun test …/series.sortValues.test.ts in evaluate.sh; no change needed.