docs(benchmark): refresh task-aware measurements

[SarthakWade]

Summary

• refresh the canonical benchmark with five fresh task-aware Linux ARM64 runs
• update the README, benchmark methodology, roadmap, and quality contract with the current medians and honest point-in-time caveat
• validate README claims and every methodology-table median against the generated benchmark document during web lint

Results

• Headless warm: 174 estimated tokens, 3,239 ms wall time, 1,218 ms CPU, 379 MiB peak memory
• Selenium: 410 estimated tokens, 2,828 ms wall time, 1,788 ms CPU, 378 MiB peak memory
• Puppeteer: 499 estimated tokens, 2,400 ms wall time, 1,849 ms CPU, 367 MiB peak memory

Puppeteer remains fastest and lowest-memory. The refreshed claim is limited to the measured agent surface and median CPU result.

Validation

• pnpm test
• pnpm test:runtime
• pnpm test:npm
• pnpm --dir apps/web lint
• pnpm --dir apps/web build
• pnpm test:e2e:linux
• ./apps/headless/build.sh
• shell syntax checks and QA evidence checksum verification
• Prettier 3.6.2 check for all changed files

Closes #50