feat(pi): eval reproducibility and comparative reporting

[Patel230]

Summary

Implements the remaining feasible item from the Pi adoption follow-up plan (docs/plans/pi-renderer-and-eval-reporting.md):

• Reproducibility hash: evalloop.Result.ReproHash — deterministic SHA-256 over model/provider/prompt/config-version and the transcript.
• Comparative report: evalloop.Compare aggregates runs (token/cost/duration ranges, unique repro hashes) with a stable sort; FormatComparison renders it.
• CLI: hawk eval loop --report prints the comparative summary and includes repro_hash in the JSON output.
• Plan doc records why the Bubble Tea renderer swap and Kitty render-loop integration are NOT adopted: Bubble Tea v2 exposes no public Renderer interface (unexported renderer), so a clean differential-renderer swap is not feasible without forking renderer internals.

Test plan

• Full go test ./... green (twice + via pre-push hook)
• New tests: hash determinism + input-sensitivity, Compare aggregation/sort, FormatComparison
• internal/feature/evalloop race-tested
• make vet, make lint, hawk verify green
• Boundary guards, submodule drift, govulncheck passed in pre-push hooks
• markdownlint-cli2 clean on the new plan doc