Build the agent-eval UI run routes #433

Part of #428.

This issue tracks building /runs and /runs/:id so we can browse saved benchmark and experiment runs and interpret their results over time.

Scope

• Define and wire the result input for both agent-eval ui and agent-eval ui build, including how multiple result bundles are discovered. Do not hard-code this repository's results directory layout.
• Provide a shared run/trial index and stable, collision-safe route IDs. Distinguish benchmark and experiment runs, multiple runs of the same resource, and runs created on the same date.
• Read supported historical result schemas through the compatibility layer from Define way for result artifacts to be durable across schema updates #316, normalizing them for the UI rather than introducing page-specific schema handling.
• List runs with their source benchmark/experiment and available creation metadata, with deterministic ordering and a bounded browsing strategy for growing history.
• Show a run's saved configuration, result summary, model/treatment/scenario breakdown, and links to individual trials and available source definitions.
• Keep historical runs viewable when current definitions are renamed, changed, or removed. Clearly distinguish saved metadata from current configuration.
• Handle no results, unknown IDs, unsupported schema versions, malformed bundles, and missing trial files explicitly. One unsupported bundle should not silently erase history or hide unrelated supported runs.
• Keep expensive trial/artifact details out of the initial collection payload and load them when needed.
• Use Primer components and design tokens, support light/dark mode, and provide accessible navigation.
• Verify local preview and static export, including direct links, deployment base paths, mixed supported schema versions, repeated same-day runs, and results without current source definitions.

Dependencies

Blocked by #316 for supported historical result ingestion. Layout and current-schema prototyping can proceed while that foundation is developed. This issue owns the shared result discovery/identity contract consumed by the homepage and other resource pages; the trials issue depends on that contract, not on completing every run-page UI detail.

Starting points

• packages/agent-eval/src/cli/commands/ui.ts:11-15,39-53 currently accepts configuration directories but has no result input.
• website/src/runs.ts:8-17,132 uses a repository-specific experiment-results directory and a date as the run ID. This needs generalization for a global run route.
• website/src/result-files.ts:60-77 currently skips invalid/incompatible bundles. Define way for result artifacts to be durable across schema updates #316 needs to define the supported compatibility boundary and diagnostics.