Add bounded run failure manifest endpoint

[Aryan95614]

Summary

• add a metadata-service endpoint that returns the current failed tasks for a run
• select only the latest failed attempt for each task and include the _exception artifact datastore reference when available
• bound responses with validated limits and opaque cursor pagination
• cover recovery, missing exception artifacts, invalid inputs, and multi-page results with PostgreSQL integration tests

Why

Agent failure investigation currently requires listing tasks, attempts, metadata, and artifacts separately. This endpoint moves the cross-table work into PostgreSQL and returns a bounded manifest that callers can inspect without loading an entire large run.

This is metadata-service-only and independent of the existing pagination and tag-filtering work.

Validation

• repository pre-commit suite: passed
• PostgreSQL-backed suite: 247 passed, 13 skipped, with one unrelated clock-sensitive UI duration test deselected
• the complete suite was also run twice; all new tests passed, while that existing UI duration assertion was the sole failure