Add eval watcher scoreboard artifacts

[mimeding]

Summary

• Adds watcher-oriented eval scoreboard artifacts for stored local/frontier report bundles.
• Adds a script and Makefile target for generating a watcher report, refreshing the latest scoreboard, and failing closed on invalid bundles.
• Gates on the latest release-candidate regressions and run failures while preserving historical comparison totals.
• Consumes unified evidence registry snapshots so watcher scoreboards share the same report source of truth as PR eval evidence.

Rebase update

• Rebased onto current main after Add safe custom JSON Agent Channel runner #1652, Add Agent Channel inbox audit workbench #1784, and Add Russian localization #1782 landed.
• Kept the newer eval harness README content while preserving the PR report and watcher examples.
• Added manifest-backed role recovery for stored report bundles so frontier reports rebuilt from reports/<model>/<suite>.json keep their correct lane.
• Documented artifact-sharing privacy expectations for watcher outputs.

Validation

• swift test --package-path Packages/OsaurusEvals --filter EvalReviewReportTests — 6 tests passed.
• swift test --package-path Packages/OsaurusEvals --filter EvalScoreboardTests — 6 tests passed.
• git diff --check
• bash -n scripts/evals/eval-watcher-report.sh
• swift run --package-path Packages/OsaurusEvals osaurus-evals scoreboard --help

Risk

• This adds report generation/scoreboard artifacts only; it does not change agent-loop behavior, prompts, model routing, or runtime defaults.

[mimeding]

Moving this back to draft because the branch is now dirty against current main even though the required checks were green. I will refresh the branch before asking for review again so the ready queue only contains merge-clean PRs.