fix: stabilize core compute budgets

[fabric0de]

Summary

• Stabilize the required 100,000-row to 1,000,000-row core compute performance gate.
• Apply workload-specific scaling ceilings based on repeated GitHub-hosted Linux and local macOS measurements.
• Keep the stricter existing limits for row-model construction and filtering.
• Add explicit boundary tests so meaningful sorting regressions continue to fail.
• Document the updated performance policy and rationale.

Performance Budget Changes

• Keep 2.0x per-row scaling ceilings for:
  • Initial row-model construction
  • Global filtering
  • Column filtering
• Set numeric sorting ceilings to 2.5x.
• Set repeated sort-direction flip ceilings to 3.0x.
• Continue rejecting quadratic scaling regressions.

Rationale

The previous uniform 2.0x ceiling repeatedly failed for sorting workloads without any Core implementation changes:

• GitHub Actions reproduced normalized sorting ratios between 2.245x and 2.300x.
• Local macOS measurement reproduced a 2.112x median and 2.458x p95 for direction flips.
• Unaffected row-model and filtering workloads remained within their existing 2.0x limits.

The revised limits account for the stable allocation and garbage-collection cost observed at one million rows without relaxing unrelated performance contracts.

Verification

• Sorting budget boundaries are covered by tests.
• Ratios immediately below each limit pass.
• Ratios immediately above each limit fail.
• Quadratic scaling regressions remain rejected.
• pnpm benchmark:core-filter:massive:run passes.
• pnpm test passes.
• pnpm check passes.
• pnpm docs:build passes.
• Performance policy documentation is updated.
• CHANGELOG.md is updated.
• No generated benchmark results, credentials, or private data are included.