bench/results holds a 500/s run of anonymous code pages through the edge cache and one bypassing it, both with zero refusals and p95 under 100 ms #9622

From #9514's open box. The speed work landed in #9548; what's missing is the measurement on production-shaped data.

Done looks like:

• [ ] bench/load.ts runs at 500/s for at least 60 s against a copy seeded to production-shaped data (bench/seed.ts, sized like production's repositories and file counts, with the counts written into the result).
• [ ] One run goes through the edge cache and one bypasses it. Both are kept as JSON in bench/results/, named for this issue, and each reports zero 429s and zero 5xx and p95 under 100 ms.
• [ ] docs/features/ (the load doc) states both numbers, the machine they came from, and the seed size. A run that only hits best case on a small copy has to be reported that way, not as done.

Files: bench/load.ts, bench/seed.ts, bench/results/, docs/features/. Risks: this waits on #9566 (a cold forge refuses strangers for its first ~20 s). Measure from a warm start only once #9566 lands, and say in the result whether the run started warm or cold.