feat(evals): compare recorded routing thresholds

[cpakkamisaac-sae]

Summary

• Compare explicit routing thresholds offline against the same recorded decisions and outcomes.
• Report fallback and target-selection counts, quality, cost, and regret for each threshold.
• Preserve the existing single-threshold report; validate and sort comparison thresholds.

Stack

Depends on #883. This branch starts at its current head, so the draft diff includes that PR until it merges. After #883 merges, I will rebase this branch onto main and make this PR ready for review.

Closes #885.

Validation

• Focused replay tests: 18 passed.
• Available Python suite: 158 passed, 2 subtests passed.
• Full Rust workspace tests, Clippy with warnings denied, formatting, Ruff, and mypy passed.
• Docker integration tests were not run because the Docker daemon was unavailable.