Summary

• Replace the benchmark curl/httpbin text and JSON calls with local file:// fixtures.
• Avoids comparing two independent external network responses in the benchmark job.
• Keeps the benchmark exercising the rtk curl path while removing the flaky curl text negative-token failure seen across open PRs.

Verification

• bash -n scripts/benchmark.sh
• git diff --check
• cargo +1.93.0 build --release
• Manual curl -s file://... vs target/release/rtk curl file://... check for both JSON and text fixtures

Note: a full local CI=1 scripts/benchmark.sh run still fails in this worktree because local warm-cache commands like cargo build produce tiny raw output and become existing unrelated negative cases. The curl rows specifically now report equal token counts instead of a negative result.